<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Anuj Sadani</title>
    <description>The latest articles on DEV Community by Anuj Sadani (@anuj-sadani).</description>
    <link>https://dev.to/anuj-sadani</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4202888%2F6844b643-7509-49b0-8798-eade0d8a416b.png</url>
      <title>DEV Community: Anuj Sadani</title>
      <link>https://dev.to/anuj-sadani</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/anuj-sadani"/>
    <language>en</language>
    <item>
      <title>The Meter Is Not the Money</title>
      <dc:creator>Anuj Sadani</dc:creator>
      <pubDate>Sun, 01 Nov 2026 23:30:00 +0000</pubDate>
      <link>https://dev.to/anuj-sadani/the-meter-is-not-the-money-51kg</link>
      <guid>https://dev.to/anuj-sadani/the-meter-is-not-the-money-51kg</guid>
      <description>&lt;p&gt;A payments company bought a router for a price it never disclosed, and its chief executive justified it by calling tokens the central currency of companies building with AI. The company being bought does not agree in the only place that counts: its own documentation says the credit balance is denominated in US dollars. A token is a meter reading, not a unit of account, and every AI credit that ships is a conversion layer over dollars.&lt;/p&gt;

&lt;p&gt;This is an adapted edition of the original article. It preserves the section argument and evidence; use the source link for the published version and any interactive material.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://tech.anujsadani.in/the-meter-is-not-the-money/" rel="noopener noreferrer"&gt;Read the original article&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;01&lt;/p&gt;

&lt;h2&gt;
  
  
  The meter is not the money
&lt;/h2&gt;

&lt;p&gt;A chief executive says tokens are a currency. His own company's products are built as though they are not.&lt;/p&gt;

&lt;p&gt;On 19 August 2026, Stripe announced it had agreed to acquire OpenRouter.&lt;sup&gt;c-001&lt;/sup&gt; The press release described the target in a single line — a platform that routes and optimizes token usage across more than four hundred models from more than eighty providers&lt;sup&gt;c-011&lt;/sup&gt; — and then handed the microphone to Patrick Collison, who explained what the company thought it was buying.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Tokens are the central currency for companies building with AI, and it’s clear that the real-world economic potential will depend on making good use of scarce compute resources.&lt;sup&gt;c-009&lt;/sup&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;It is a good sentence, and I suspect it is the sentence that sold the deal internally. It is also wrong in a specific way that the rest of this book is about — and the strange thing is that Stripe’s own products appear to know it, because not one of them is built as though it were true.&lt;/p&gt;

&lt;p&gt;Consider the thing Stripe just paid billions for. OpenRouter runs a credit system: you load money in, you spend it down on inference. In what unit is that balance held? The documentation answers without ceremony. OpenRouter uses a credit system in which the base currency is US dollars.&lt;sup&gt;c-015&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;Not tokens. Dollars.&lt;/p&gt;

&lt;p&gt;The competition does the same. Vercel’s AI Gateway sells credits too, and bases its rates on the provider’s list price, with charges deducted from a dollar balance.&lt;sup&gt;c-047&lt;/sup&gt; And Stripe’s own AI billing product, which shipped before this acquisition, is the clearest tell of all: it lets a business define an “AI Credit” and then configure how usage of each model &lt;em&gt;converts&lt;/em&gt; to credits.&lt;sup&gt;c-018&lt;/sup&gt; Converts. From one thing into a different thing, at a rate somebody has to set. That verb is carrying the entire argument.&lt;/p&gt;

&lt;p&gt;Every serious system in this market meters in tokens and settles in dollars. Nobody has built one that does otherwise, and it is not for want of ambition. &lt;strong&gt;The token is the meter. The dollar is still the money.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;What a currency has to do&lt;/p&gt;

&lt;p&gt;The reason is less about finance than about measurement, and it starts with a property so basic that it usually goes unstated: a unit of account has to mean the same thing to both parties at the moment they agree to a trade.&lt;/p&gt;

&lt;p&gt;A dollar is a dollar when you commit to spend it. A kilowatt-hour is a kilowatt-hour, which is why your electricity bill works — you do not know in advance how many you will burn, but you know what one &lt;em&gt;is&lt;/em&gt;, and you know its price before you use it. That is the whole trick of a meter. The quantity floats; the unit does not.&lt;/p&gt;

&lt;p&gt;A token fails this, and it fails it in four different ways at once, none of them exotic and all of them documented by the vendors themselves in the pages they expect their own customers to read.&lt;/p&gt;

&lt;p&gt;The first is that a token’s price depends on who serves it. For open-weight models the weights and the hardware come apart: a lab publishes a model, and then dozens of companies run it on their own machines, at their own prices, with their own throughput. The output is identical. The invoice is not. OpenRouter’s routing documentation puts the spread in its own worked example — Provider A at one dollar per million tokens, Provider B at two, Provider C at three.&lt;sup&gt;c-025&lt;/sup&gt; A factor of three, for the same tokens out of the same model.&lt;/p&gt;

&lt;p&gt;The second is that you do not get to choose which of them serves you, and neither, quite, does the router. You might reasonably assume it picks the cheapest and moves on. It does not. The documented default is to load balance requests across providers, prioritizing price&lt;sup&gt;c-024&lt;/sup&gt; — balance, note, not minimize — and the balancing is probabilistic, with candidates weighted by the inverse square of their price. In the documentation’s own example that makes the cheapest provider nine times likelier to be selected than the dearest.&lt;sup&gt;c-026&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;Nine times more likely still sends roughly one request in ten to a costlier supplier. The die is loaded; it remains a die. Which means that at the instant you press send, the price of that request is not yet a number — it is a distribution, and it will not collapse into a number until after the work is done.&lt;/p&gt;

&lt;p&gt;This is the load-bearing fact of the book and it is worth slowing down for, because it is stronger than the thing people usually say. The complaint is not that AI pricing is hard to predict, or that it varies by workload. It is that the price &lt;em&gt;does not exist yet&lt;/em&gt; at the moment of purchase. It gets settled afterwards, by a draw you neither control nor observe.&lt;/p&gt;

&lt;p&gt;There is a good reason for the design, and it is worth stating plainly rather than treating the nondeterminism as a flaw somebody should fix. Routing every request to the single cheapest provider would pile load onto one supplier, wreck tail latency, and hand that supplier the power to reprice at will. Spreading traffic across a price-weighted field is how a router keeps its own supply side honest. The nondeterminism is deliberate, and it is load-bearing for the market’s health. It is simply fatal to using the token as a unit of account.&lt;/p&gt;

&lt;p&gt;The third failure is that not all tokens are the same kind of token. Prompt caching lets a provider hold part of your context so that repeated calls do not reprocess it, which makes it the single largest lever on real-world inference cost — and which splits one token into several priced categories. On recent OpenAI models, cache &lt;em&gt;writes&lt;/em&gt; are charged at 1.25 times the price of the original input pricing.&lt;sup&gt;c-027&lt;/sup&gt; Cache &lt;em&gt;reads&lt;/em&gt; are charged at a quarter or a half of it, depending on the model.&lt;sup&gt;c-028&lt;/sup&gt; So a token going into the cache costs a quarter more than an ordinary input token, and the same token coming back out costs a half or a quarter of one. Same word, same model, same request, three prices, decided by which side of a cache boundary it happened to fall on.&lt;/p&gt;

&lt;p&gt;If those categories were at least consistent between vendors, a conversion table would rescue the situation. They are not, and that is the fourth failure. The same documentation records that some providers — Anthropic among them — carry a &lt;em&gt;negative&lt;/em&gt; discount on cache writes and a positive one on cache reads.&lt;sup&gt;c-029&lt;/sup&gt; A negative discount is a surcharge. One vendor’s cache write is a rebate and another’s is a penalty, with no shared sign convention to appeal to. And because caching and routing are entangled — the router pins a session to a provider only when that provider’s cache-read price beats its ordinary prompt price&lt;sup&gt;c-030&lt;/sup&gt; — the caching behavior changes the routing, which changes the price, which changes whether the caching was worth doing.&lt;/p&gt;

&lt;p&gt;Where this bites&lt;/p&gt;

&lt;p&gt;Put the four together and you arrive somewhere uncomfortable: you cannot quote a fixed price for a request before it is routed. Not because the tooling is young. Because there is no fixed price there to quote.&lt;/p&gt;

&lt;p&gt;Anyone who has tried to reconcile an AI bill has already met this, usually without having a name for it. The most honest acknowledgment I have found sits in the documentation of LiteLLM, the open-source gateway that a great many teams use to track their spend. It carries a section headed, without euphemism, “Cost does not match your provider bill?” — and beneath it a debugging workflow instructing you to align time ranges, compare token categories including cache, and then decide whether the gap is ingestion, formula, or model-map pricing.&lt;sup&gt;c-031&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;Read that last clause again. Three named ways your own meter can disagree with your supplier’s, catalogued by a mature tool that does this for a living. Metering AI spend is a genuinely hard measurement problem rather than a solved one waiting on a better implementation, and the software that does it best is candid about that.&lt;/p&gt;

&lt;p&gt;The cost of getting this wrong is rarely a catastrophe. It is a slow tax, and it comes due in four places. Budgets drift, because a forecast built on average cost per token moves as the routing mix moves — and the drift is invisible, since the token count was right all along. Margins erode quietly: if you resell inference at a fixed price per unit while your own cost is a distribution, you have written somebody an option and handed it over for free. Reconciliation becomes permanent, a monthly ritual of comparing two numbers that will never agree and deciding how much disagreement is tolerable. And optimization goes after the wrong quarry, because cutting token count is not the same as cutting cost once cache reads are a quarter price and cache writes are a surcharge. Teams tune the meter when they meant to tune the bill.&lt;/p&gt;

&lt;p&gt;None of that is exotic. It is just what happens when you denominate something in a unit that will not hold still.&lt;/p&gt;

&lt;p&gt;What got built instead&lt;/p&gt;

&lt;p&gt;So the interesting question is not why nobody has minted a token currency. It is what the people closest to the problem built once they worked out that they could not.&lt;/p&gt;

&lt;p&gt;They built a conversion layer. Stripe’s billing product does not sell tokens; it lets a business define an abstract AI Credit, specify how each model’s usage converts into that unit, and then sell prepaid packs of it.&lt;sup&gt;c-018&lt;/sup&gt; The metering underneath is fine-grained enough to carry every distinction this section has been describing, segmented by model and by token type including cached tokens.&lt;sup&gt;c-019&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;That is the right architecture, and it concedes the whole argument. The credit is a synthetic unit rather than a token, with an exchange rate against real tokens that is configured per model and can be changed by whoever is selling it. The complexity does not vanish. It moves behind an interface, where one business absorbs it so that its customers never have to look at it. At the time of writing, that product was not generally available; access sat behind a waiting list.&lt;sup&gt;c-020&lt;/sup&gt; The idea has run further ahead than the shipping.&lt;/p&gt;

&lt;p&gt;If you take one thing into your own systems from this section, make it the discipline of keeping the two units apart. Store budgets in dollars and derive tokens, never the reverse — any internal system holding a budget as a token count has a bug that will stay hidden until the routing mix shifts under it. Stamp an access date on every price you rely on, because vendors reprice without changelogs and a price page is evidence about a day rather than a standing fact. Capture the cached-versus-uncached split in your own telemetry from the beginning, since retrofitting it after a surprising invoice means reprocessing logs that may never have kept the field. And decide in advance how much drift between your meter and your supplier’s you are willing to accept, then write the number down — teams that skip this end up settling it during a month-end close, where the answer gets decided by whoever is most tired.&lt;/p&gt;

&lt;p&gt;02&lt;/p&gt;

&lt;h2&gt;
  
  
  What a payments company saw
&lt;/h2&gt;

&lt;p&gt;Four reported prices, none of them disclosed, and one fee structure that explains the interest better than any of them.&lt;/p&gt;

&lt;p&gt;Nobody knows what Stripe paid for OpenRouter.&lt;/p&gt;

&lt;p&gt;That is worth stating flatly at the outset, because the number gets repeated in conversation as though it were settled, and it is not. Stripe disclosed no terms.&lt;sup&gt;c-003&lt;/sup&gt; What exists instead is four figures from four outlets, each citing a source it does not name. Bloomberg reported more than seven billion dollars.&lt;sup&gt;c-004&lt;/sup&gt; The &lt;em&gt;New York Times&lt;/em&gt;, via CNBC, put it at about \$7.5 billion, with \$1.5 billion of that allocated to OpenRouter’s founders.&lt;sup&gt;c-005&lt;/sup&gt; Semafor said eight billion.&lt;sup&gt;c-006&lt;/sup&gt; And earlier, while the talks were still running, The Information had the figure at ten billion.&lt;sup&gt;c-007&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;The temptation is to average these and move on. Resist it. They are not four measurements of one quantity, and averaging them would produce a number that nobody reported and no document supports. They are snapshots of a negotiation taken at different moments by different people with different access. If you need a number for an argument, use a range and say whose it is. If your argument only works at eight billion, it is a weaker argument than you think, because eight billion is the loosest of the four.&lt;/p&gt;

&lt;p&gt;What is not in dispute is the shape of the jump. Less than three months before the acquisition, OpenRouter had raised \$113 million at a valuation of about \$1.3 billion.&lt;sup&gt;c-008&lt;/sup&gt; Whatever the closing figure, the company was worth something on the order of five to six times more in August than it had been in May, and nothing about the product had changed in the interval.&lt;/p&gt;

&lt;p&gt;One more thing was true at the time of writing and may not be by the time you read this: the acquisition had not closed. It remained subject to customary closing conditions, with the parties expecting to close within weeks.&lt;sup&gt;c-002&lt;/sup&gt; Everything in this book describes an agreed deal rather than a completed one.&lt;/p&gt;

&lt;p&gt;So the question is what the buyer saw.&lt;/p&gt;

&lt;p&gt;The shape of the take&lt;/p&gt;

&lt;p&gt;The usual answer is distribution: OpenRouter sits in front of a great deal of traffic, and traffic is valuable. That is true but incomplete, and it misses the detail that should make a payments company sit up.&lt;/p&gt;

&lt;p&gt;Look at how OpenRouter actually earns. It does not mark up inference. Its documentation is explicit that there is no markup on inference pricing, and that it charges a fee when you purchase credits instead.&lt;sup&gt;c-012&lt;/sup&gt; What you pay for tokens is what the underlying provider charges for tokens. The business is instead a percentage taken at the moment money enters the system: the published platform fee on the pay-as-you-go plan is 5.5%.&lt;sup&gt;c-013&lt;/sup&gt; For customers bringing their own provider keys the arrangement shifts but does not change character — free up to \$25,000 of list-price inference per month, then a 5% fee above that.&lt;sup&gt;c-014&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;Sit with the structure rather than the numbers. A software company charges for the software. A reseller buys wholesale and sells retail, and lives on the spread. OpenRouter does neither. It passes the goods through at cost and takes a percentage of the money moved.&lt;/p&gt;

&lt;p&gt;That is a take rate. It is the economic signature of a payment processor, and it is a genuinely unusual thing to find in an infrastructure company. The fee does not scale with how much compute you consume, or how many models you call, or how many seats you have. It scales with &lt;strong&gt;how much money you put in.&lt;/strong&gt; Whether you spend that balance on one enormous batch job or ten million small requests, whether you burn it in a day or leave it sitting for a quarter, the 5.5% was collected at the door.&lt;/p&gt;

&lt;p&gt;A company whose entire business is understanding money at the door would recognize that instantly.&lt;/p&gt;

&lt;p&gt;The economics that come with it are the kind payments people find familiar. OpenRouter was reported to be generating roughly \$140 million in annualized revenue by mid-2026,&lt;sup&gt;c-016&lt;/sup&gt; with the cost of serving the routing product at 28.5% of revenue and gross margin around 70%.&lt;sup&gt;c-040&lt;/sup&gt; That margin is not the product of clever pricing on inference, because there is no pricing on inference. It is what falls out of taking a percentage of transaction value while paying only for the infrastructure that moves the transaction.&lt;/p&gt;

&lt;p&gt;The part that was already there&lt;/p&gt;

&lt;p&gt;The strongest evidence that Stripe understood this is that Stripe had already started building it.&lt;/p&gt;

&lt;p&gt;The press release says so, in a clause that reads like housekeeping and is not: since the previous year, Stripe had been working to help companies optimize their token costs and route tokens efficiently, having launched products including Token Billing.&lt;sup&gt;c-017&lt;/sup&gt; Before any of this, Stripe was already metering AI tokens as a billing primitive. It had shipped the meter. What it lacked was the thing that knows, request by request, which of eighty providers served the work and what that provider charged — and that knowledge is not something you can build from the outside, because it lives in the routing decision itself.&lt;/p&gt;

&lt;p&gt;Read the acquisition that way and it stops looking like a payments company wandering into AI infrastructure. It looks like a metering business acquiring the instrument its meter was missing.&lt;/p&gt;

&lt;p&gt;There is a further clue in how Stripe describes the difficulty. The release frames the problem not as cost but as combinatorial complexity — which model to use for which tasks, at which speed and at what price — and notes that the challenge is exacerbated by the pace at which models are released and repriced.&lt;sup&gt;c-075&lt;/sup&gt; That is a description of an unstable unit of account written by someone who has clearly stared at the problem. And the stated ambition is to help businesses maximize profitability by routing their requests intelligently and spending their tokens efficiently.&lt;sup&gt;c-010&lt;/sup&gt; Revenue and cost. Money in and money out. The two sides of a ledger, which is the one thing Stripe has always sold.&lt;/p&gt;

&lt;p&gt;Where the argument gets uncomfortable&lt;/p&gt;

&lt;p&gt;All of which supports the reading that this was a payments acquisition, and I think that reading is right. But it does not support the further claim that gets attached to it — that the token is therefore becoming a currency, and that owning the router is owning the mint.&lt;/p&gt;

&lt;p&gt;The same company that charges a payment processor’s take rate settles it in dollars. OpenRouter’s credit balance has US dollars as its base currency.&lt;sup&gt;c-015&lt;/sup&gt; The 5.5% is levied on dollars entering, not on tokens leaving. At no point in the flow does a token function as money. It functions as the thing being measured so that dollars can be moved correctly, which is exactly what a meter is for.&lt;/p&gt;

&lt;p&gt;What this section can and cannot carry&lt;/p&gt;

&lt;p&gt;Everything here about the &lt;em&gt;rationale&lt;/em&gt; comes from parties with an obvious interest in the rationale sounding coherent. A press release is a first-rate source for the fact that a deal happened and a poor one for why. Both companies had every reason to present a legible strategic story in August 2026, and neither had any reason to say that they paid a large multiple because the asset was scarce and someone else might have taken it.&lt;/p&gt;

&lt;p&gt;The take-rate structure is different, and it is why this section leans on it. That is not a story anyone told me. It is on a price page, it was there before the deal, and it would be there whether or not the acquisition made sense.&lt;/p&gt;

&lt;p&gt;What follows from it is a question rather than a conclusion. If the position is genuinely a payments position, the 5.5% should be defensible — rents at a chokepoint usually are. Section 4 tests that, and finds the picture considerably messier than the number suggests. Before that, though, it is worth understanding what the router is actually doing when it decides where your request goes.&lt;/p&gt;

&lt;p&gt;03&lt;/p&gt;

&lt;h2&gt;
  
  
  The router as price maker
&lt;/h2&gt;

&lt;p&gt;A component that selects suppliers by price, at volume, is not observing a market. It is setting one.&lt;/p&gt;

&lt;p&gt;A router looks like plumbing. Requests arrive, requests leave, and the interesting work is assumed to be happening at the ends. That framing survives right up until you notice that the router is choosing, on every single request, which supplier gets paid — and that it makes the choice on price.&lt;/p&gt;

&lt;p&gt;Once a component selects suppliers by price and controls enough volume, it stops being a passive observer of a market. Supplier revenue is now a function of the router’s weighting. That is the argument of this section, and the way to see it is to look closely at what the routing decision actually involves, because there are far more priced dimensions in it than most people building on top of it realize.&lt;/p&gt;

&lt;p&gt;More dials than anyone expects&lt;/p&gt;

&lt;p&gt;Start with the one from Section 1: provider selection weighted by the inverse square of price, so that the cheapest candidate is nine times likelier than the dearest in the documentation’s own three-provider example.&lt;sup&gt;c-026&lt;/sup&gt; That alone makes the cost of a request a distribution rather than a number.&lt;/p&gt;

&lt;p&gt;But price is only the default sort. A caller can filter on measured performance first and &lt;em&gt;then&lt;/em&gt; sort on price, and the filtering is empirical rather than advertised — OpenRouter will select hosts that have sustained at least 40 tokens per second for 90% of requests over the last five minutes, and route to the cheapest provider within that qualified group.&lt;sup&gt;c-064&lt;/sup&gt; That is a rolling five-minute window over a ninetieth percentile. The set of providers eligible to serve you changes continuously, based on how they have been performing against other people’s traffic in the last few minutes. Two identical requests, sent ninety seconds apart, can face different candidate sets and therefore different price distributions, without anything in your code changing and without any provider having announced a price change.&lt;/p&gt;

&lt;p&gt;Then consider what else carries a multiplier once you look beyond the router at the vendors underneath.&lt;/p&gt;

&lt;p&gt;Latency tolerance is priced. Amazon Bedrock offers selected models for batch inference at a 50% lower price than on-demand inference.&lt;sup&gt;c-063&lt;/sup&gt; Half price, for agreeing to wait. If your workload can tolerate a delay and you have not told anyone, you are paying twice what you need to.&lt;/p&gt;

&lt;p&gt;Geography is priced. Anthropic applies a 1.1x multiplier across all token pricing categories when a caller specifies US-only inference.&lt;sup&gt;c-060&lt;/sup&gt; Not a surcharge on a line item — a ten percent uplift across input, output, cache writes and cache reads alike. A compliance requirement, expressed as a coefficient.&lt;/p&gt;

&lt;p&gt;The cache is the economy&lt;/p&gt;

&lt;p&gt;Section 1 gave the OpenAI figures: cache writes at 1.25x input on recent models, cache reads at a quarter or a half. Those figures belong to one vendor rather than to the industry, and the others differ.&lt;/p&gt;

&lt;p&gt;Anthropic’s published table runs on three multipliers rather than two, because it sells cache duration as a product. A five-minute cache write costs 1.25x base input price. A one-hour cache write costs 2x. A cache read costs 0.1x.&lt;sup&gt;c-058&lt;/sup&gt; Bedrock, meanwhile, prices cache reads at 75% below its on-demand input price&lt;sup&gt;c-062&lt;/sup&gt; — a 0.25x multiplier, which is neither Anthropic’s 0.1x nor the 0.5x that some OpenAI models carry.&lt;/p&gt;

&lt;p&gt;Three vendors, three different discounts for the identical operation of reading a cached token. There is no convergence here and no sign of any, which means a tool that wants to predict cost across vendors cannot hold a single number for “cache read.” It has to hold a table, and keep it current, against vendors who reprice without changelogs.&lt;/p&gt;

&lt;p&gt;Anthropic, to its credit, publishes the break-even directly: caching pays off after one cache read for the five-minute duration, or after two cache reads for the one-hour duration.&lt;sup&gt;c-059&lt;/sup&gt; That is an unusually honest piece of documentation and it is worth dwelling on, because it makes the decision structure visible. Whether to cache is not a performance question with a cost side effect. It is a bet on reuse. Write to the five-minute cache and you need one hit to break even. Buy the hour and you need two. Get the reuse pattern wrong and caching costs you money — which is the opposite of how almost everyone describes caching in every other part of computing.&lt;/p&gt;

&lt;p&gt;Now couple that back to routing, because the two are not independent. A cache lives on a specific provider’s hardware. To benefit from it, subsequent requests have to reach the same provider, which is why the router pins sessions — and pins them conditionally, activating sticky routing only when the provider’s cache-read price beats its ordinary prompt price.&lt;sup&gt;c-030&lt;/sup&gt; The router is running the break-even calculation on your behalf and changing its routing policy based on the answer.&lt;/p&gt;

&lt;p&gt;The entanglement produces failure modes that look like nothing at all in a diff. Changing a model slug mid-conversation — appending a latency-oriented variant suffix, say — invalidates the sticky key.&lt;sup&gt;c-065&lt;/sup&gt; A one-word change to a configuration string, made to improve latency, silently discards a warm cache and re-routes the conversation. The latency probably does improve. The bill moves in a direction nobody predicted, because the person who made the change was reasoning about speed and the system was also reasoning about cache residency.&lt;/p&gt;

&lt;p&gt;What this makes the router&lt;/p&gt;

&lt;p&gt;Assemble the dials: price-weighted probabilistic selection, rolling throughput filters, batch tiers, geographic multipliers, three-way cache multipliers that differ per vendor, and sticky sessions that engage conditionally on a price comparison. Every one of those is a lever on what a request costs, and the router holds most of them.&lt;/p&gt;

&lt;p&gt;A component holding that many pricing levers over that much volume is not a neutral pipe. When a router weights by inverse square of price, it is not merely finding the cheap provider — it is &lt;em&gt;manufacturing&lt;/em&gt; demand elasticity. A provider that drops its price does not simply win the requests where it is cheapest; it wins a larger share of a probabilistic draw across the whole eligible set, immediately, with no sales cycle and no contract renegotiation. And a provider that raises its price does not lose its customers, because it never had customers in the retail sense. It loses weight in a distribution.&lt;/p&gt;

&lt;p&gt;That is a real market mechanism, and it is a better one than the alternative in most respects. It is also, from the perspective of anyone trying to answer “what will this cost,” an active adversary. The very properties that make the routing layer good at disciplining suppliers are the properties that make a fixed price impossible to quote.&lt;/p&gt;

&lt;p&gt;Mechanism established, magnitude not&lt;/p&gt;

&lt;p&gt;Everything above describes documented mechanism, not measured outcome. I have not seen data on how much the price distribution actually moves in practice for a typical workload, and I am not aware of any published study that measures it. It is entirely possible that for most applications the variance is small enough to ignore, and that the practical answer is to use a trailing average and stop worrying.&lt;/p&gt;

&lt;p&gt;The mechanism is certain; the magnitude is not, and nobody should take a claim about magnitude from this book, because this book does not have one.&lt;/p&gt;

&lt;p&gt;What the mechanism does establish is where the leverage sits. If a supplier’s revenue depends on a weighting function held by an intermediary, the intermediary is the most commercially valuable component in the chain — regardless of whether it charges anything for the privilege. Which raises the question the next section has to answer: given that position, what does the market actually charge for standing in it?&lt;/p&gt;

&lt;p&gt;The answer turns out to be four incompatible things at once.&lt;/p&gt;

&lt;p&gt;04&lt;/p&gt;

&lt;h2&gt;
  
  
  Who charges for standing in the middle
&lt;/h2&gt;

&lt;p&gt;Five firms, one function, and prices that range from 5.5% of everything you spend to nothing at all.&lt;/p&gt;

&lt;p&gt;If the routing layer is as valuable a position as the last two sections suggest, the market should agree on what it is worth. Markets usually do converge on a price for a well-understood function, even a contested one — card acquiring settled into a band, cloud storage settled into a band, and you can predict what a new entrant will charge to within a few basis points.&lt;/p&gt;

&lt;p&gt;The AI gateway market has not converged. It has not even nearly converged. The same function — sit between an application and a set of model providers, route the request, meter the spend — is monetized four incompatible ways by five serious companies, and the prices are not close.&lt;sup&gt;c-069&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;OpenRouter charges a share of the money you load: 5.5% on the pay-as-you-go plan.&lt;sup&gt;c-013&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;Vercel charges nothing on the tokens. Its documentation states plainly that the AI Gateway charges no markup and no platform fee on tokens, with the customer paying the provider’s list price and Vercel deducting from a credit balance.&lt;sup&gt;c-021&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;Portkey charges a subscription. Its production plan is \$49 a month with a quota of 100,000 recorded logs&lt;sup&gt;c-067&lt;/sup&gt; — a software license with usage-based logging, entirely decoupled from how much inference flows through it. Below that sits a permanently free tier, explicitly not suitable for production workloads.&lt;sup&gt;c-068&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;Cloudflare charges nothing at all. Its AI Gateway’s core features are offered free on all plans, and the core it names is not a stub: dashboard analytics, caching, and rate limiting.&lt;sup&gt;c-066&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;And beneath all of them, LiteLLM gives the metering away as open-source software you run yourself, tracking spend for keys, users, and teams across more than a hundred models.&lt;sup&gt;c-023&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;Five providers. A percentage of funds, zero, a flat monthly fee, zero again, and free source code.&lt;/p&gt;

&lt;p&gt;What a spread like that means&lt;/p&gt;

&lt;p&gt;A price dispersion this wide is not noise. It is a signal that the participants disagree about what business they are in.&lt;/p&gt;

&lt;p&gt;Take a percentage of money loaded and you have declared yourself a financial intermediary whose revenue scales with customer spend. Charge a flat subscription and you have declared yourself a software vendor whose revenue scales with customer count. Charge nothing and you have declared the gateway a feature of something else you sell — hosting, in Vercel’s case; the broader network, in Cloudflare’s. Publish the source and you have declared it infrastructure that should not be a business at all.&lt;/p&gt;

&lt;p&gt;Those are four different businesses wearing the same coat. And it means the 5.5% is not the market price for routing. It is one firm’s answer to a question the market has not settled, and it is the highest answer on the board by a wide margin — infinitely higher than three of the four alternatives, which charge nothing for the gateway itself.&lt;/p&gt;

&lt;p&gt;I want to be careful about what this does and does not establish. It does not establish that the 5.5% is doomed. Prices persist above zero-priced alternatives all the time, for reasons that are perfectly rational: switching costs, breadth of model coverage, reliability, the fact that the free options make you operate them yourself. OpenRouter’s fee buys access to a catalogue and an operational burden somebody else carries, and plenty of teams will consider that a bargain at 5.5% of spend. The presence of a free competitor does not refute a paid one; if it did, nobody would sell databases.&lt;/p&gt;

&lt;p&gt;What it does establish is that the fee is a &lt;strong&gt;business-model choice rather than a property of the position.&lt;/strong&gt;&lt;sup&gt;c-051&lt;/sup&gt; Occupying the routing layer does not entitle you to a take rate. Two of the largest firms in this list occupy the same layer and have decided the correct rent is nothing, because they monetize elsewhere and would rather have the traffic.&lt;/p&gt;

&lt;p&gt;And it complicates the payments framing from Section 2 in a specific way. A payment processor’s take rate is defensible because moving regulated money at scale is genuinely hard, genuinely licensed, and genuinely not something a competitor can give away as a loss leader. A routing fee has none of those protections. Cloudflare can zero it out on Tuesday and absorb the cost inside a network business, and there is no regulator, no license, and no settlement obligation standing in the way.&lt;/p&gt;

&lt;p&gt;The other pressure&lt;/p&gt;

&lt;p&gt;The competitive squeeze is the visible half. The other half is that the thing being charged for does not work perfectly, and its own practitioners say so.&lt;sup&gt;c-054&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;Section 1 quoted LiteLLM’s documentation on reconciliation. Read in the context of this section that admission carries more weight. If metering were both hard &lt;em&gt;and&lt;/em&gt; reliable, it would be a defensible service. If it were easy and reliable, it would be a commodity but at least a working one. What the documentation describes is a capability that is available for free and does not fully work — which is close to the worst possible configuration for anyone hoping to charge rent on it.&lt;/p&gt;

&lt;p&gt;Sacra’s analysis of OpenRouter names the pressure directly, observing that open-source alternatives such as LiteLLM let enterprises self-host routing infrastructure, reducing willingness to pay for standalone routing.&lt;sup&gt;c-045&lt;/sup&gt; That is a private-market research firm modeling a company it does not have access to, so treat it as a considered opinion rather than a finding. But it is the same conclusion the price list arrives at from the other direction.&lt;/p&gt;

&lt;p&gt;Five data points are not a distribution&lt;/p&gt;

&lt;p&gt;I have five prices, gathered because they were reachable, and no systematic survey of the gateway market. There are more gateways than these — hyperscalers bundling routing into existing contracts, observability vendors extending into it, enterprise platforms building it internally. I do not know the shape of the whole distribution, only that its visible members disagree violently.&lt;/p&gt;

&lt;p&gt;A proper market survey would strengthen or overturn the argument here, and it has not been done.&lt;/p&gt;

&lt;p&gt;What can be said with the evidence in hand is narrower and still useful. The fee at this layer is contested, several well-capitalized firms have set it to zero, the underlying capability is available as free software, and that software concedes it does not always agree with the invoice. Whatever Stripe bought, it was not an uncontested toll booth.&lt;/p&gt;

&lt;p&gt;Which leads directly to the strongest argument against this entire book — the possibility that everyone involved already knows all of this, and that the routing fee going to zero was never the point.&lt;/p&gt;

&lt;p&gt;05&lt;/p&gt;

&lt;h2&gt;
  
  
  The loss leader
&lt;/h2&gt;

&lt;p&gt;The strongest case against this book, presented without a rebuttal appended, because I could not build one.&lt;/p&gt;

&lt;p&gt;Everything so far has built toward a conclusion: that the routing layer matters because it is where compute becomes a billable quantity, and that a payments company bought it for that reason. This section sets out the strongest case that the conclusion is wrong, or at least that it is right for reasons that make it far less interesting than it sounds.&lt;/p&gt;

&lt;p&gt;I am not going to answer it. I could not, with the evidence I gathered, and appending a rebuttal I cannot support would be worse than leaving it standing.&lt;/p&gt;

&lt;p&gt;The case that routing is meant to be worthless&lt;/p&gt;

&lt;p&gt;Start with the version Semafor published two days after the announcement, which is the sharpest reading of the deal I have seen anywhere.&lt;/p&gt;

&lt;p&gt;The observation is that OpenRouter’s business could be destined for commoditization, and that this is actually kind of the point.&lt;sup&gt;c-041&lt;/sup&gt; The reasoning follows from the market structure described in Section 4 rather than contradicting it. If routing collapses into finding the cheapest provider that clears a quality and reliability bar, then routing is a solved problem with a shrinking margin, and everyone in it is heading for zero. But zero is survivable — attractive, even — if you are not trying to earn there: Stripe can afford to let routing become low-margin, because it can monetize everything surrounding it, from billing and tax and fraud through to stablecoin settlement, treasury and possibly financing.&lt;sup&gt;c-042&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;Take that seriously and the thesis of this book inverts. The routing layer is not the valuable position; it is the &lt;em&gt;entry point&lt;/em&gt; to a set of valuable positions, none of which is routing. Stripe does not need a take rate on inference because it has billing, tax, fraud, settlement and treasury waiting on the other side of the relationship, all of which are businesses it already knows how to run at scale. Under this reading, buying OpenRouter is buying distribution into every AI-native company’s financial stack, and the fact that the acquired product happens to route tokens is almost incidental.&lt;/p&gt;

&lt;p&gt;Notice what this does to Section 4’s argument. I presented the four-way price dispersion as a problem for the payments framing — evidence that the fee is contested and might go to zero. Semafor’s reading takes the identical fact and calls it the plan. Under this account the 5.5% is not a rent to be defended but a legacy price that a rational owner will happily compete away, because the customer relationship is worth more than the fee ever was. The same evidence, pointing the opposite direction. I do not have a way to distinguish the two from outside.&lt;/p&gt;

&lt;p&gt;The case that it is an expense-management deal&lt;/p&gt;

&lt;p&gt;The second reading is less dramatic and possibly more likely.&lt;/p&gt;

&lt;p&gt;TechCrunch noted that Stripe’s large acquisitions had historically been about helping companies collect and manage incoming money, and that this one looked different — a move to the other side of the ledger, into expense management, beginning with AI expenses.&lt;sup&gt;c-043&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;That reframes the deal as being about &lt;em&gt;cost control software&lt;/em&gt;, which is a real and large category with nothing much to do with payment rails. And the competitive evidence supports it, because that category was already filling up before Stripe arrived: Databricks developed its own AI gateway, Rippling launched one focused on employee AI spend, and Ramp launched one for AI expense management too.&lt;sup&gt;c-044&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;Rippling and Ramp are not infrastructure companies. They are spend-management companies, and they built AI gateways because AI spend is a line item their customers cannot see clearly. If the natural competitor set for this product is Ramp rather than Cloudflare, then the position Stripe bought is a seat in the expense-management market, and the elaborate argument about units of account is a description of a technical problem rather than an explanation of a purchase.&lt;/p&gt;

&lt;p&gt;The uncomfortable part, for me, is that this reading explains the price better than mine does. Expense management is a proven category with known multiples and an obvious buyer synergy. Owning the settlement layer for a future compute economy is a story that requires several things to happen that have not happened yet.&lt;/p&gt;

&lt;p&gt;The case that the price was simply too high&lt;/p&gt;

&lt;p&gt;And then there is the arithmetic, which is unkind whichever story you prefer.&lt;/p&gt;

&lt;p&gt;At the figure floated during the talks, Stripe would have been paying roughly seventy times OpenRouter’s annualized revenue.&lt;sup&gt;c-038&lt;/sup&gt; Seventy times. The same reporting describes that price as high compared to other AI acquisitions when calculated as a multiple of forward revenue.&lt;sup&gt;c-039&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;The defense is that the unit economics are unusually good for the category, and they are — Section 2 set out the margin structure. But a strong margin on \$140 million of revenue does not obviously justify a price several dozen times that revenue unless you believe the revenue is about to become very much larger, which brings you back to needing the strategic story to be true.&lt;/p&gt;

&lt;p&gt;There is also a piece of evidence about the buyer’s state of mind that is hard to categorize and too good to omit. A leaked letter from Stripe’s founders to investors explained the company’s operating posture by saying it had decided that 1 January marked the beginning of the singularity and it had been operating on that basis.&lt;sup&gt;c-076&lt;/sup&gt; The reporting is careful to note the phrase is tongue-in-cheek and that Patrick Collison had used it in that spirit publicly.&lt;sup&gt;c-077&lt;/sup&gt; I am not going to build an argument on a joke in a letter. But a firm operating on an explicitly stated assumption that the technological curve has gone vertical is a firm whose acquisition multiples should be read in that light, and it is at least worth knowing that the assumption was written down.&lt;/p&gt;

&lt;p&gt;What would have to be true for this book to be right&lt;/p&gt;

&lt;p&gt;Set the three counter-readings side by side and the honest position is that all of them are consistent with the evidence I have.&lt;sup&gt;c-052&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;For the argument in this book to hold, the metering-and-settlement position has to be worth more than the expense-management seat, and the router has to remain the place where the metering happens rather than being reduced to a commodity component inside somebody else’s billing stack. Neither is established. Both are assertions about a market that is roughly eighteen months old.&lt;/p&gt;

&lt;p&gt;What I can say is that the counter-arguments are about &lt;em&gt;where the value accrues&lt;/em&gt;, not about the mechanism. Nothing in this section disputes that the token is a poor unit of account, that prices are nondeterministic at submission, or that the credit is a conversion layer over dollars. Semafor, TechCrunch and PYMNTS are arguing about who captures the value created by that mess. They are not arguing that the mess is imaginary.&lt;/p&gt;

&lt;p&gt;So the mechanical claims of Sections 1 and 3 survive this section intact. The commercial claim of Section 2 — that this was a payments acquisition in the meaningful sense rather than the metaphorical one — does not survive it cleanly, and I would not want a reader to finish this book believing it had been demonstrated. It is the reading I find most persuasive. It is not the only one the facts allow.&lt;/p&gt;

&lt;p&gt;One small thing cuts the other way, and it deserves a mention because it is the only piece of post-deal evidence anyone has. OpenRouter’s own announcement commits to neutrality in unusually specific language: routing decisions will remain driven by one thing, which is what is best for the user.&lt;sup&gt;c-046&lt;/sup&gt; If routing were merely a funnel into a financial-services relationship, that commitment would be an odd thing to make and an expensive one to keep.&lt;/p&gt;

&lt;p&gt;06&lt;/p&gt;

&lt;h2&gt;
  
  
  What can be built
&lt;/h2&gt;

&lt;p&gt;Prepaid credits, gifted compute, and the regulatory line that explains why you cannot buy a friend a hundred thousand tokens.&lt;/p&gt;

&lt;p&gt;The practical question underneath all of this is whether you can sell someone a quantity of AI compute the way you can sell them a quantity of anything else. A gift card. A prepaid balance. A coffee bought for a stranger on the internet.&lt;/p&gt;

&lt;p&gt;The answer turns out to be yes, with a conversion layer, inside a closed loop — and the three qualifications are doing all the work.&lt;/p&gt;

&lt;p&gt;The unit everybody invents&lt;/p&gt;

&lt;p&gt;By now the pattern should be familiar, but it is worth seeing how far it extends, because the conversion layer is not a Stripe peculiarity. It is what everyone builds when they need a stable number.&lt;/p&gt;

&lt;p&gt;Stripe’s billing product, as established, lets a business define an AI Credit and configure how each model’s usage converts into it, then sell prepaid packs of that unit.&lt;sup&gt;c-018&lt;/sup&gt; The metering underneath tracks model and token type including cached tokens,&lt;sup&gt;c-019&lt;/sup&gt; and the product was not yet generally available at the time of writing.&lt;sup&gt;c-020&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;Anthropic has already shipped the same idea, and almost nobody has noticed. When Claude is sold through the AWS or Azure marketplaces, Anthropic rates the token usage in USD at standard per-model rates, applies any negotiated discount, and converts the result to Claude Consumption Units at one cent each.&lt;sup&gt;c-061&lt;/sup&gt; A CCU is worth a penny. It exists because a cloud marketplace needs a single countable line item and cannot express a charge such as &lt;em&gt;1.25x base input price for a five-minute cache write&lt;/em&gt;, so the vendor rates everything in dollars, converts to a synthetic unit at a fixed rate, and reports a quantity.&lt;/p&gt;

&lt;p&gt;That is the whole thesis of this book, shipped as a production billing integration by a model vendor, for reasons that have nothing to do with any acquisition. Dollars are the unit of account. The synthetic unit is a presentation layer over dollars, invented because the underlying quantity would not sit still. And note the direction of the conversion — usage is rated &lt;em&gt;in USD first&lt;/em&gt;, then converted. The dollars are not derived from the units. The units are derived from the dollars.&lt;/p&gt;

&lt;p&gt;Anyone building a credit product is going to reinvent this, and should. The design question is not whether to have a conversion layer but who absorbs the risk in it. If you sell a credit today and the underlying provider reprices tomorrow, either your margin moves or your customer’s balance quietly buys less than it did. Someone eats that. Fixing the exchange rate is a promise about a price you do not control, and Section 3 is a long argument for why you should be nervous about making it.&lt;/p&gt;

&lt;p&gt;The expiry problem&lt;/p&gt;

&lt;p&gt;There is a second feature of every prepaid AI product currently shipping, and it is one customers dislike: the value goes away.&lt;/p&gt;

&lt;p&gt;OpenRouter’s terms provide for expiry of unused credits after one year of purchase.&lt;sup&gt;c-036&lt;/sup&gt; Anthropic’s gift subscriptions expire 365 days after the purchase date.&lt;sup&gt;c-035&lt;/sup&gt; These are not outliers; they are the standard shape.&lt;/p&gt;

&lt;p&gt;Flagged, not analyzed&lt;/p&gt;

&lt;p&gt;Expiry is where prepaid products stop being a pricing decision and start being a regulated one. Money taken now for a service delivered later is a liability on the seller’s balance sheet, and unredeemed value is revenue only under specific accounting conditions. Consumer protection rules in many jurisdictions constrain when prepaid value may expire and what fees may be charged against it.&lt;/p&gt;

&lt;p&gt;I gathered no sources on the accounting or the consumer-protection treatment. This is a prompt to get advice, not advice.&lt;/p&gt;

&lt;p&gt;Buying someone else compute&lt;/p&gt;

&lt;p&gt;Which brings us to the consumer version, and the reason it is harder than it looks.&lt;/p&gt;

&lt;p&gt;The patronage model is well established and boringly profitable. Buy Me a Coffee charges a 5% transaction fee, creators keep 95% of the earnings, and payouts go directly to a bank account.&lt;sup&gt;c-032&lt;/sup&gt; It is a thin layer over a payment processor, and the processor is Stripe: the platform states that payment processing is handled by Stripe and Wise.&lt;sup&gt;c-033&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;So the obvious product suggests itself. Instead of buying a writer a coffee, buy them the inference to write with. Both legs already exist and both already terminate at the same company — the patronage rail runs on Stripe, and Stripe supplies the token-metering primitive.&lt;sup&gt;c-055&lt;/sup&gt; There is no missing infrastructure.&lt;/p&gt;

&lt;p&gt;And yet nobody sells it, while two adjacent things do ship. Anthropic will sell you a gift, but the gift is a plan: you choose a tier and a duration, and the recipient receives full access to the features of the selected plan.&lt;sup&gt;c-034&lt;/sup&gt; It is a subscription with a bow on it, not a balance. You cannot gift someone eighty thousand tokens and let them spend it where they like.&lt;/p&gt;

&lt;p&gt;The reason is not technical, and I think it is not squeamishness about the product either. It is that a transferable balance spendable across unaffiliated vendors is a different kind of object in law from a subscription to one vendor’s service.&lt;/p&gt;

&lt;p&gt;The line that actually matters&lt;/p&gt;

&lt;p&gt;Here I have to correct a framing I nearly used, because it is the sort of error that survives in a book for years.&lt;/p&gt;

&lt;p&gt;The natural place to reach for is FinCEN’s guidance on stored value, which held that the agency does not currently interpret the definition of stored value to include closed system products such as a mall-wide gift card program.&lt;sup&gt;c-037&lt;/sup&gt; That is a clean statement of the closed-versus-open distinction, and it is tempting to quote and move on.&lt;/p&gt;

&lt;p&gt;It is also from 2003,&lt;sup&gt;c-056&lt;/sup&gt; the same document warns that FinCEN intends to engage in further rulemaking on the definition of stored value,&lt;sup&gt;c-057&lt;/sup&gt; and — decisively — that rulemaking happened. FinCEN’s 2011 final rule amended the Bank Secrecy Act regulations by renaming stored value as prepaid access, defining that term, and deleting the terms issuer and redeemer of stored value.&lt;sup&gt;c-070&lt;/sup&gt; The vocabulary the 2003 ruling operates in no longer exists in the regulations. The rule was framed as addressing regulatory gaps that had resulted from the proliferation of prepaid innovations over the preceding twelve years&lt;sup&gt;c-073&lt;/sup&gt; — precisely the category a cross-vendor compute credit would fall into.&lt;/p&gt;

&lt;p&gt;The 2011 regime imposes suspicious activity reporting and recordkeeping requirements on both providers and sellers of prepaid access, plus registration for providers,&lt;sup&gt;c-071&lt;/sup&gt; while exempting certain lower-risk categories from certain requirements&lt;sup&gt;c-072&lt;/sup&gt; — and it retains closed loop prepaid access as a defined category.&lt;sup&gt;c-074&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;The shape of the answer is therefore visible even from outside: closed loop is treated as lower risk and carries lighter obligations; open loop does not. A gift card for one vendor’s service sits comfortably on one side of that line. A transferable balance spendable across eighty inference providers sits on the other, or at minimum sits close enough to it that you would need a lawyer before you shipped.&lt;sup&gt;c-053&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;Where this book’s competence ends&lt;/p&gt;

&lt;p&gt;I have read a 2011 final rule and the definitions it amended. I have not researched the exemption thresholds in detail, state money-transmitter licensing, the CFPB’s prepaid account rules under Regulation E, the CARD Act’s gift-card provisions, or anything outside the United States. Several of those could be decisive.&lt;/p&gt;

&lt;p&gt;What I can say is narrow: the reason you cannot buy your friend a hundred thousand tokens is not that the payments plumbing is missing. It is that the object you would be selling them is a regulated instrument, and the closed-loop products that ship today are shaped the way they are because that shape avoids the question.&lt;/p&gt;

&lt;p&gt;What would change my mind&lt;/p&gt;

&lt;p&gt;The argument of this book comes down to one claim: that tokens are a meter reading, that every credit is a conversion layer over dollars,&lt;sup&gt;c-049&lt;/sup&gt; and that the interesting position is therefore the exchange desk rather than the mint.&lt;/p&gt;

&lt;p&gt;Here is what would falsify it, and what I would watch.&lt;/p&gt;

&lt;p&gt;If a vendor began quoting a &lt;strong&gt;fixed price per token guaranteed at submission time&lt;/strong&gt; — absorbing the routing variance itself, the way an insurer absorbs risk — the central mechanical claim would weaken considerably.&lt;sup&gt;c-048&lt;/sup&gt; The variance would still exist; it would simply have found someone willing to be paid to hold it, which is a real product and one I would expect to see before I expect to see a token currency.&lt;/p&gt;

&lt;p&gt;If a &lt;strong&gt;transferable, cross-vendor balance shipped and survived&lt;/strong&gt;, the argument in this section is wrong, and the constraint I have attributed to regulation was really just nobody having bothered.&lt;/p&gt;

&lt;p&gt;If the &lt;strong&gt;cache multipliers converged&lt;/strong&gt; across vendors — if 0.1x and 0.25x and 0.5x became one number the way transaction fees converged into a band — then a common unit becomes tractable, and the mess described in Section 3 turns out to have been an artifact of an immature market rather than a structural feature.&lt;/p&gt;

&lt;p&gt;And the single most informative thing, the one that costs nothing to observe: whether OpenRouter’s commitment that routing decisions will remain driven by what is best for the user holds under its new ownership.&lt;sup&gt;c-046&lt;/sup&gt; If in three years the router quietly prefers providers that settle through its parent’s rails, then the exchange desk was the point, and it was worth more than the eight billion dollars nobody will confirm was paid for it.&lt;/p&gt;

&lt;p&gt;If it does not, then Section 5 was right, and I will have written a book about the plumbing of a business that turned out to be about something else. That would be a fine outcome. The plumbing is real either way, and the bills still have to be reconciled by someone on a Tuesday afternoon, holding two numbers that do not match.&lt;/p&gt;




&lt;p&gt;&lt;a href="https://tech.anujsadani.in/the-meter-is-not-the-money/" rel="noopener noreferrer"&gt;Read the original on Anuj Sadani | Tech&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>machinelearning</category>
      <category>devtools</category>
      <category>programming</category>
    </item>
    <item>
      <title>Clone Your Voice Locally With Qwen3-TTS: Recordings, Mic Checks, and Reference Audio</title>
      <dc:creator>Anuj Sadani</dc:creator>
      <pubDate>Sun, 01 Nov 2026 19:30:00 +0000</pubDate>
      <link>https://dev.to/anuj-sadani/clone-your-voice-locally-with-qwen3-tts-recordings-mic-checks-and-reference-audio-4o2n</link>
      <guid>https://dev.to/anuj-sadani/clone-your-voice-locally-with-qwen3-tts-recordings-mic-checks-and-reference-audio-4o2n</guid>
      <description>&lt;p&gt;Qwen3-TTS can produce narration in a cloned voice offline. My first three recording sessions mostly taught me what not to trust: Bluetooth headsets and laptop microphones can cap capture at 8 kHz while the exported file still reports 48 kHz. In this workflow, four seconds of clean reference audio beat ninety seconds of poor capture.&lt;/p&gt;

&lt;p&gt;The guide focuses on recording and measurement as much as on the model, so you can diagnose the input before judging the generated voice.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Follow the practical steps:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/clone-your-voice-qwen3-tts/#why-yours" rel="noopener noreferrer"&gt;Why use your own voice&lt;/a&gt;: what voice cloning adds to a personal narration workflow.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/clone-your-voice-qwen3-tts/#passage" rel="noopener noreferrer"&gt;Choose the passage&lt;/a&gt;: why the reference text affects the sample.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/clone-your-voice-qwen3-tts/#microphone" rel="noopener noreferrer"&gt;Check what the microphone captured&lt;/a&gt;: find hidden sample-rate limits instead of trusting file metadata.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/clone-your-voice-qwen3-tts/#measure" rel="noopener noreferrer"&gt;Measure before you trust the take&lt;/a&gt;: inspect the signal before running inference.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/clone-your-voice-qwen3-tts/#shorter" rel="noopener noreferrer"&gt;Use a shorter reference&lt;/a&gt;: compare a clean short clip with a longer reference.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/clone-your-voice-qwen3-tts/#pace" rel="noopener noreferrer"&gt;Check the speaking pace&lt;/a&gt;: decide whether the generated narration is usable for long listening.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://tech.anujsadani.in/clone-your-voice-qwen3-tts/" rel="noopener noreferrer"&gt;Open the complete Qwen3-TTS walkthrough&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>python</category>
      <category>machinelearning</category>
      <category>tutorial</category>
    </item>
    <item>
      <title>Narrate Your Writing Locally With Kokoro: Voices, Acronyms, and Silence</title>
      <dc:creator>Anuj Sadani</dc:creator>
      <pubDate>Sun, 01 Nov 2026 15:30:00 +0000</pubDate>
      <link>https://dev.to/anuj-sadani/narrate-your-writing-locally-with-kokoro-voices-acronyms-and-silence-3dn3</link>
      <guid>https://dev.to/anuj-sadani/narrate-your-writing-locally-with-kokoro-voices-acronyms-and-silence-3dn3</guid>
      <description>&lt;p&gt;Kokoro is an 82M-parameter text-to-speech model that can make a written article listenable on a laptop. Getting a model to speak is straightforward. Getting the result to sound consistent across a long piece takes more care.&lt;/p&gt;

&lt;p&gt;This guide focuses on the practical details that changed the output: voice styles are vectors you can blend, the espeak frontend needs help with acronyms, and two silence settings matter more than another model parameter.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The learning path:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/narrate-writing-kokoro/#why-local" rel="noopener noreferrer"&gt;A narration you can re-run&lt;/a&gt;: why keep the pipeline local and repeatable?&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/narrate-writing-kokoro/#blending" rel="noopener noreferrer"&gt;Voices are vectors&lt;/a&gt;: combine voice characteristics rather than switching voices abruptly.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/narrate-writing-kokoro/#acronyms" rel="noopener noreferrer"&gt;The words the model actually reads&lt;/a&gt;: handle acronyms as the speech frontend sees them.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/narrate-writing-kokoro/#silence" rel="noopener noreferrer"&gt;Silence is a parameter&lt;/a&gt;: tune pauses and pacing for an article, not a short sample.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/narrate-writing-kokoro/#caching" rel="noopener noreferrer"&gt;Re-renders should cost nothing&lt;/a&gt;: make iteration cheaper with cached segments.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The piece also documents loudness, file size, and the limits of the setup. &lt;a href="https://tech.anujsadani.in/narrate-writing-kokoro/" rel="noopener noreferrer"&gt;Read the full Kokoro narration guide&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>python</category>
      <category>opensource</category>
      <category>tutorial</category>
    </item>
    <item>
      <title>Run Small AI Models on an Old GPU: Fix the Software Path First</title>
      <dc:creator>Anuj Sadani</dc:creator>
      <pubDate>Sun, 01 Nov 2026 11:30:00 +0000</pubDate>
      <link>https://dev.to/anuj-sadani/run-small-ai-models-on-an-old-gpu-fix-the-software-path-first-3m91</link>
      <guid>https://dev.to/anuj-sadani/run-small-ai-models-on-an-old-gpu-fix-the-software-path-first-3m91</guid>
      <description>&lt;p&gt;A six-year-old GTX 1650 with 4 GB of VRAM can still run small models. In my case, the obstacles were mostly software: a CPU-only PyTorch wheel, a missing PATH entry, a dtype unsupported by Turing, and a batch size that pushed data into system RAM.&lt;/p&gt;

&lt;p&gt;This field guide walks through the investigation rather than promising that every old GPU will run every workload. It is especially useful if your device appears capable on paper but your framework reports that it cannot see CUDA.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The main sections:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/revive-old-gpu-small-models/#the-card" rel="noopener noreferrer"&gt;The card nobody recommends&lt;/a&gt;: establish the constraints before choosing a model.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/revive-old-gpu-small-models/#driver" rel="noopener noreferrer"&gt;The driver and CPU wheel&lt;/a&gt;: check what the installed driver and PyTorch build actually support.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/revive-old-gpu-small-models/#path" rel="noopener noreferrer"&gt;Why CUDA looks broken on Windows&lt;/a&gt;: verify the executable and library paths.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/revive-old-gpu-small-models/#dtype" rel="noopener noreferrer"&gt;Why fp16 crashes on Turing&lt;/a&gt;: choose a precision mode the hardware can handle.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/revive-old-gpu-small-models/#batch" rel="noopener noreferrer"&gt;Tune against memory, not the clock&lt;/a&gt;: avoid misleading speed measurements caused by memory spill.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/revive-old-gpu-small-models/#payoff" rel="noopener noreferrer"&gt;What the old card actually does&lt;/a&gt;: see the practical result and limits.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The guide also covers DirectML, watchdog resets, and interrupted runs. &lt;a href="https://tech.anujsadani.in/revive-old-gpu-small-models/" rel="noopener noreferrer"&gt;Read the complete walkthrough&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>python</category>
      <category>machinelearning</category>
      <category>tutorial</category>
    </item>
    <item>
      <title>The Attention Invoice</title>
      <dc:creator>Anuj Sadani</dc:creator>
      <pubDate>Sun, 01 Nov 2026 07:30:00 +0000</pubDate>
      <link>https://dev.to/anuj-sadani/the-attention-invoice-58ao</link>
      <guid>https://dev.to/anuj-sadani/the-attention-invoice-58ao</guid>
      <description>&lt;p&gt;How AI slop transfers its cost to everyone else. LinkedIn's own definition is the right one: not whether a machine helped, but whether anyone exercised enough judgment to make the result worth someone else's attention. Eight chapters on what changed when generation got cheap, why detection is the wrong question, and what to do about it on Monday. Includes a downloadable designed PDF.&lt;/p&gt;

&lt;p&gt;This is an adapted edition of the original article. It preserves the section argument and evidence; use the source link for the published version and any interactive material.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://tech.anujsadani.in/ai-slop/" rel="noopener noreferrer"&gt;Read the original article&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;§ 01 · Definitions&lt;/p&gt;

&lt;h2&gt;
  
  
  What Is AI Slop?
&lt;/h2&gt;

&lt;p&gt;LinkedIn now has a feedback option called “Seems like AI slop.” Open the menu on a post, choose it, and LinkedIn treats that as a signal that the post did not feel valuable. It does not remove the post or make a policy finding about its author.&lt;sup&gt;1&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;The feature is useful because it names a feeling that has become familiar. A person opens a post promising an insight about leadership, AI agents, or the future of work. The prose is smooth. The paragraphs are clean. There is a confident opening, a tidy list, and a final lesson. Then the reader reaches the end and realizes that no one has actually said anything.&lt;/p&gt;

&lt;p&gt;But the label also creates a confusion. People hear “AI slop” and infer “text that an AI wrote.” That is too broad to be useful. An AI can help a person research, outline, translate, edit, test an argument, or make a difficult idea clearer. None of those acts makes the result slop.&lt;/p&gt;

&lt;p&gt;The real question is harder: &lt;strong&gt;did someone exercise enough judgment to make this worth another person’s attention?&lt;/strong&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  The distinction LinkedIn gets right
&lt;/h3&gt;

&lt;p&gt;LinkedIn’s own definition is better than the name of its button. It describes AI slop as low-effort, likely AI-generated content that sounds polished on the surface but lacks a point of view, unique perspective, or substance. It says the distinction is not how content was created, but whether it adds value and reflects a person’s perspective, experience, or expertise.&lt;sup&gt;2&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;That is the right center of gravity. The generated sentence is not the offense. The unexamined sentence is.&lt;/p&gt;

&lt;p&gt;Imagine three people publishing an article about AI agents:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Generate → publish&lt;/strong&gt;: &lt;strong&gt;What happened:&lt;/strong&gt; A model produces 1,500 words and the author pastes them into a post.; &lt;strong&gt;Slop?:&lt;/strong&gt; Usually.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Human idea → AI draft → human edit&lt;/strong&gt;: &lt;strong&gt;What happened:&lt;/strong&gt; The author supplies a thesis, examples, corrections, and a real rewrite.; &lt;strong&gt;Slop?:&lt;/strong&gt; Usually not.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Human research → AI synthesis → human judgment&lt;/strong&gt;: &lt;strong&gt;What happened:&lt;/strong&gt; The author decides which sources matter, checks the claims, and stands behind the conclusion.; &lt;strong&gt;Slop?:&lt;/strong&gt; No.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The table is intentionally about process rather than style. The visible artifact is a weak signal. A heavily directed and carefully edited piece can retain the rhythm of a model. A person can write generic, derivative, attention-seeking prose without touching an AI. Manual typing has never been a reliable measure of thought.&lt;/p&gt;

&lt;p&gt;Two articles can therefore look similar while being fundamentally different. One may be prompt, generate, publish. The other may contain an idea, reporting, context construction, criticism, correction, and verification before the final generation. A reader cannot reliably see that history from punctuation, em dashes, headings, or an AI-detection score.&lt;/p&gt;

&lt;p&gt;Calling both pieces slop because both contain generated tokens collapses authorship into keystrokes. Authorship is better understood as the work of choosing: choosing the question, the evidence, the examples, the omissions, and the words one is prepared to defend.&lt;/p&gt;

&lt;h3&gt;
  
  
  A family resemblance, not a test
&lt;/h3&gt;

&lt;p&gt;There is no settled technical definition of AI slop. A recent scholarly paper argues that it has resisted formal definition and proposes three recurring properties instead: superficial competence, asymmetric effort, and mass producibility.&lt;sup&gt;3&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;Those properties explain why the phrase feels precise even when its boundary is fuzzy.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Superficial competence&lt;/strong&gt; is the polished shell. The grammar works. The image looks plausible. The memo has the expected headings. Underneath, the claims are unsupported, the summary never gets specific, or the apparent conclusion is just the premise repeated in better phrasing.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Asymmetric effort&lt;/strong&gt; is the production problem. It takes seconds to make a convincing-looking artifact and much longer for another person to decide whether it is accurate, useful, or original. The producer gets the convenience; the reader, reviewer, customer, or colleague inherits the verification cost.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Mass producibility&lt;/strong&gt; is the ecosystem problem. One mediocre image or awkward paragraph is not necessarily what people mean by slop. The term becomes more apt when a system can make thousands of near-substitutes for search rankings, ad impressions, engagement, or the appearance of output. Reporting on the rise of the term has repeatedly connected it to low-quality text, images, videos, and entire sites produced cheaply for search traffic.&lt;sup&gt;4&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;None of these is a pass-fail test. A personalized surreal image can be silly and still mean something to the person who made it. A concise generated answer to a narrow question can be exactly what someone needs. The scholarly account makes this point too: so-called slop can meet demand for niche or personalized content and can carry aesthetic or social value.&lt;sup&gt;5&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;That is why “I dislike it” and “it is AI slop” should not be interchangeable. The term is most useful when it names a failure that affects someone beyond the author: an output that asks for attention, trust, or review without having been given enough thought to earn any of them.&lt;/p&gt;

&lt;h3&gt;
  
  
  The missing word is judgment
&lt;/h3&gt;

&lt;p&gt;The usual list of tells is familiar: generic language, repeated ideas, unnecessary length, fake profundity, too many headings, confident claims with no evidence. Those are symptoms. They matter because they suggest an absence of judgment, but none proves it.&lt;/p&gt;

&lt;p&gt;A person can use every fashionable AI writing habit and still make an original, careful argument. Another can avoid every tell and publish a confident falsehood they wrote alone. The important difference is whether someone took responsibility for the content: checked what could be checked, removed what did not hold up, supplied the context a model could not know, and decided that the piece had a reason to exist.&lt;/p&gt;

&lt;p&gt;This makes a better working definition:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;AI slop is generated output whose cost of production has fallen much faster than the effort spent deciding whether it deserves to exist.&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Or, in the shortest form: &lt;strong&gt;AI slop is a failure of judgment, not a method of generation.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;The definition does not let every AI-assisted author off the hook. It raises the bar in the place that matters. If a model makes drafting cheap, then choosing, checking, and editing become more valuable, not less. The person who publishes the result still owns the claim.&lt;/p&gt;

&lt;h3&gt;
  
  
  The software version of the same problem
&lt;/h3&gt;

&lt;p&gt;This is not only a problem for feeds. Generative AI can produce code, documentation, tests, issue comments, design notes, and pull requests much faster than a team can understand them. The same asymmetry appears: generating an artifact is cheap; reviewing it, maintaining it, and recovering its rationale are expensive.&lt;/p&gt;

&lt;p&gt;The danger is not that AI wrote a function. A well-directed model can remove drudgery and leave more time for work that needs judgment. The danger arrives when a team measures generated volume as if it were progress, accepts code that nobody can explain, or creates documentation that no one has verified against the system it describes.&lt;/p&gt;

&lt;p&gt;That is engineering slop: artifacts that create the appearance of movement while increasing uncertainty about why they exist, whether they are correct, and who understands them. The technical debt is not simply AI-generated code. It is the gap between what the repository contains and what its humans can account for.&lt;/p&gt;

&lt;p&gt;The practical response is ordinary, even if the tools are new. Ask before publishing or merging: What is the point? What is new here? What did we verify? Who will own this when it breaks? If those questions do not have good answers, the artifact is not finished merely because it looks finished.&lt;/p&gt;

&lt;p&gt;LinkedIn’s button cannot reconstruct the invisible process behind a post. No detector can do that reliably either. But the phrase can still be useful if it reminds us what readers are reacting to: not the presence of a machine, but the absence of a mind that cared enough to decide.&lt;/p&gt;

&lt;p&gt;§ 02 · The Debate&lt;/p&gt;

&lt;h2&gt;
  
  
  The Argument About the Word
&lt;/h2&gt;

&lt;p&gt;“Slop” is a hostile word. That is part of its usefulness and part of its risk. It carries the picture of something poured into a trough: plentiful, uniform, barely worth choosing. It is not a neutral description of a production method.&lt;/p&gt;

&lt;p&gt;That matters because the word is now used for at least four different complaints. Sometimes it means false information. Sometimes it means ugly or derivative culture. Sometimes it means a volume problem: too much content competing for the same attention. And sometimes it means a moral complaint about a creator who outsourced too much of the work.&lt;/p&gt;

&lt;p&gt;Those are related, but they are not the same.&lt;/p&gt;

&lt;p&gt;A fabricated local-news story is dangerous because it can be false. A hundred perfectly grammatical product descriptions can be useless because they add no information. A strange AI video made for one family may be neither useful nor dangerous; it may simply be a new form of play. If every one of these is called slop, the word stops helping us decide what to do next.&lt;/p&gt;

&lt;p&gt;The academic account of the term is unusually careful here. It does not offer a test that separates slop from non-slop. It offers a family resemblance: superficial competence, asymmetric effort, and mass producibility.&lt;sup&gt;6&lt;/sup&gt; That is a better starting point than an AI detector because it points at the social conditions around the artifact.&lt;/p&gt;

&lt;p&gt;The boundary will remain fuzzy because value is contextual. A generated bedtime story made for one child may be cheap to create and deeply valuable to that child. A heavily researched report may use model-written sentences and still earn attention because the author made the decisive choices. The object does not announce its process.&lt;/p&gt;

&lt;p&gt;The question is therefore not “Was AI used?” It is: &lt;strong&gt;what burden did this artifact place on other people, and what care did its maker take before placing it there?&lt;/strong&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  In practice
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Name the failure precisely.&lt;/strong&gt; Say false, repetitive, derivative, or unreviewed when that is what you mean.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Do not make style a verdict.&lt;/strong&gt; Smooth prose is neither evidence of care nor evidence of its absence.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Reserve “slop” for the larger failure.&lt;/strong&gt; It is most useful when cheap output has shifted verification, attention, or cleanup onto somebody else.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;§ 03 · The Mechanism&lt;/p&gt;

&lt;h2&gt;
  
  
  When Output Became Cheap
&lt;/h2&gt;

&lt;p&gt;For most of the internet’s life, publishing imposed a small amount of friction. Someone had to write the article, make the image, record the voiceover, or assemble the page. The cost was never high enough to guarantee quality. It was high enough to force a choice.&lt;/p&gt;

&lt;p&gt;Generative systems changed the shape of that choice. They did not make judgment cheap. They made the artifacts that normally signal judgment cheap: paragraphs, illustrations, outlines, summaries, scripts, reviews, and code.&lt;/p&gt;

&lt;p&gt;That creates a predictable temptation. When the cost of production falls, the producer can run more experiments. Some experiments are good: a teacher can make an explanation for a student who needs it; a small business can translate a manual; an engineer can draft a test case. Others are attempts to capture attention at a price close to zero.&lt;/p&gt;

&lt;p&gt;Content farms have found the second use especially attractive. NewsGuard’s tracker describes thousands of unreliable AI-generated news sites operating with little or no human oversight. It reports that such sites often publish dozens of articles a day, and that programmatic advertising can reward them without regard to the quality of the site carrying the ad.&lt;sup&gt;7&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;The important fact is not that the articles are machine-written. The business model does not need a reader to learn anything. It needs a page to be indexed, clicked, and shown an advertisement.&lt;/p&gt;

&lt;p&gt;This is why scale changes the moral texture of mediocre content. A weak essay asks for five minutes. Ten thousand weak essays change the cost of finding anything strong. They turn discernment into unpaid labor for everyone downstream: the reader searching, the editor reviewing, the platform ranking, the model training on the resulting web.&lt;/p&gt;

&lt;p&gt;Cheap supply does not create demand for meaning. It can, however, create a great deal of material that resembles meaning at first glance.&lt;/p&gt;

&lt;h3&gt;
  
  
  In practice
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Measure a publishing system by reader benefit, not artifacts shipped.&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ask what each item is for.&lt;/strong&gt; “To rank” and “to help a reader decide” are not interchangeable purposes.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Put human review where distribution is automated.&lt;/strong&gt; Scale without review is the condition in which a small quality problem becomes a public one.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;§ 04 · The Limit&lt;/p&gt;

&lt;h2&gt;
  
  
  The Artifact Cannot Tell You
&lt;/h2&gt;

&lt;p&gt;A reader sees the final text. They do not see the notebook beside it, the rejected drafts, the source tabs, the fact check, or the moment someone removed a sentence because it could not be defended.&lt;/p&gt;

&lt;p&gt;This is the central difficulty in calling something AI slop. We can observe the artifact and infer from it. We cannot observe the process that produced it.&lt;/p&gt;

&lt;p&gt;The familiar tells are weak evidence. Certain transitions, an abundance of headings, em dashes, a polished but generic cadence: these may suggest a model, but they do not tell us whether an author supplied a real claim, constructed the context, or checked the result. A detector faces the same limit. Even if it identified machine-generated prose perfectly, it would identify generation, not judgment.&lt;/p&gt;

&lt;p&gt;The opposite error is equally common. A human can type every word of a vague, derivative post. The labor is real; the contribution may still be zero. We should not confuse authorship with keystrokes.&lt;/p&gt;

&lt;p&gt;Authorship is the chain of decisions that makes an artifact answerable: why this question, why these sources, why this example, why this conclusion, why now. AI can participate in that chain. It cannot remove the need for someone to own it.&lt;/p&gt;

&lt;p&gt;This is why disclosure alone is not a quality test. “Written with AI” can be honest and still leave a reader with no reason to trust the work. “Written by a human” can conceal carelessness. The more useful convention is responsibility: the named author should be able to explain and defend the work.&lt;/p&gt;

&lt;h3&gt;
  
  
  In practice
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Keep an evidence trail for consequential work.&lt;/strong&gt; A short list of sources and decisions is more meaningful than a purity claim about tools.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Use disclosure to clarify process, not to borrow credibility.&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ask the author question.&lt;/strong&gt; Can the person whose name is on this explain why the claim is here and what would change their mind?&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;§ 05 · The Cost&lt;/p&gt;

&lt;h2&gt;
  
  
  The Price Paid by Readers
&lt;/h2&gt;

&lt;p&gt;The complaint about AI slop is often described as aesthetic: too many strange images, too much bland prose, another LinkedIn post that sounds like every other LinkedIn post. That is real, but it is not the deepest cost.&lt;/p&gt;

&lt;p&gt;The deeper cost is attention. Every plausible-looking low-value artifact asks a reader to make a small investigation. Is this true? Is it copied? Is there a real person behind it? Is the advice specific enough to use? Each investigation is cheap in isolation. At scale, it makes the information environment tiring.&lt;/p&gt;

&lt;p&gt;That fatigue changes behavior. Readers skim more aggressively. They trust less. They favor familiar sources because checking new ones is expensive. Small publishers and careful novices then pay a tax created by people who can flood the same channels with cheaper imitations.&lt;/p&gt;

&lt;p&gt;Researchers discussing AI slop describe the asymmetry directly: output can take far less effort to generate than a comparable human-made artifact, while its surface competence can disguise a lack of substance.&lt;sup&gt;8&lt;/sup&gt; The reader must do the expensive part after the producer has skipped it.&lt;/p&gt;

&lt;p&gt;The problem resembles spam, but only up to a point. Spam filters work because many messages share patterns and because the cost of a false negative is often low. Slop may be persuasive, topical, and nearly correct. The most expensive cases are not absurd. They are the calm, credible paragraph that leaves out the one condition that matters.&lt;/p&gt;

&lt;p&gt;This is also why a blanket hostility to synthetic media is a poor answer. The goal is not a return to a world where only people with spare time can publish. The goal is a world where the cost of publishing does not become an invoice sent to everyone who reads.&lt;/p&gt;

&lt;h3&gt;
  
  
  In practice
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Respect attention as a finite resource.&lt;/strong&gt; Do not publish a summary when a link and one sentence would do.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Make claims checkable.&lt;/strong&gt; Link the source, name the uncertainty, and remove decorative certainty.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Reward useful specificity.&lt;/strong&gt; It is harder to mass-produce and easier for a reader to test.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;§ 06 · The Workplace&lt;/p&gt;

&lt;h2&gt;
  
  
  When Slop Enters the Repository
&lt;/h2&gt;

&lt;p&gt;Software teams are beginning to meet the same problem in a form they recognize immediately. An AI assistant can generate a feature branch, a test suite, a migration, a design memo, and a pull-request description before the reviewer has finished reading the ticket.&lt;/p&gt;

&lt;p&gt;The code may run. That is not the end of the question.&lt;/p&gt;

&lt;p&gt;Code has a longer afterlife than a post. It will be reviewed, deployed, debugged, changed by someone new, and used as evidence of how the system works. A patch that its author cannot explain transfers all of that future reasoning to other people.&lt;/p&gt;

&lt;p&gt;A recent paper frames this as a software commons problem: individual gains from AI-generated output can externalize costs onto reviewer capacity, codebase integrity, shared knowledge, and collaborative trust.&lt;sup&gt;9&lt;/sup&gt; Another qualitative study of developer discussion identified review friction and quality degradation as central themes in complaints about AI-assisted software development.&lt;sup&gt;10&lt;/sup&gt; These are early accounts, not settled measurements. They are still useful because they name where the cost appears.&lt;/p&gt;

&lt;p&gt;The engineering version of slop is therefore not “code made with AI.” It is code that creates the appearance of progress while increasing uncertainty about behavior, rationale, and ownership.&lt;/p&gt;

&lt;p&gt;The crucial question is not whether the assistant wrote the first draft. It is whether a human can trace the change from requirement to design to tests and can carry the responsibility when it fails. When that chain exists, AI can reduce drudgery. When it does not, generated volume becomes technical debt at production speed.&lt;/p&gt;

&lt;h3&gt;
  
  
  In practice
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Review for explanation, not only execution.&lt;/strong&gt; A passing test is necessary; an understood change is durable.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Require a rationale for non-obvious code.&lt;/strong&gt; Generated comments are not a rationale unless someone verified them.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Keep changes small enough to own.&lt;/strong&gt; A reviewer should be able to say what changed, why it is safe, and who will maintain it.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;§ 07 · The Test&lt;/p&gt;

&lt;h2&gt;
  
  
  The Responsibility Test
&lt;/h2&gt;

&lt;p&gt;There is no reliable detector for care. That does not leave us helpless. It means the standard should be behavioral rather than forensic.&lt;/p&gt;

&lt;p&gt;A useful responsibility test has five questions.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What is the point?&lt;/strong&gt; A piece should be able to name the question it answers for a particular reader. “Content about AI” is a category, not a point.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What is new here?&lt;/strong&gt; The answer may be a firsthand observation, a synthesis, a worked example, a useful framing, or a decision. If nothing is new, a link is usually better than a post.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What was checked?&lt;/strong&gt; This does not demand impossible certainty. It demands that a publisher distinguish verified facts, interpretation, and speculation.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What was removed?&lt;/strong&gt; Editing is evidence of judgment. The sentences that did not survive are often more important than the sentences that did.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Who owns the result?&lt;/strong&gt; The person who publishes or merges should be able to answer questions about it without blaming the model.&lt;/p&gt;

&lt;p&gt;This test is deliberately indifferent to how many tokens a machine produced. It makes room for a writer who uses AI extensively and for one who uses none. It also makes the actual obligation visible: publishing is a promise to bear some of the reader’s verification burden yourself.&lt;/p&gt;

&lt;p&gt;In organizations, this should be a process rather than a cultural wish. People need enough time to read outputs, enough permission to say “this does not add up,” and incentives that do not reward artifact count over outcomes.&lt;/p&gt;

&lt;p&gt;The strongest defense against slop is not better taste. It is a system in which care is cheaper to exercise than carelessness is to pass downstream.&lt;/p&gt;

&lt;h3&gt;
  
  
  In practice
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Add the five questions to publishing and PR templates.&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Treat uncertainty as useful information.&lt;/strong&gt; “We do not know” is often the most valuable sentence in a report.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Make ownership explicit.&lt;/strong&gt; A named owner is not decoration; it tells others where understanding is supposed to live.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;§ 08 · The Practice&lt;/p&gt;

&lt;h2&gt;
  
  
  What to Do on Monday
&lt;/h2&gt;

&lt;p&gt;The useful response to AI slop is not to ban tools or to become a detective of punctuation. Both answers attack the visible surface and miss the real failure.&lt;/p&gt;

&lt;p&gt;Start with the work that leaves your hands.&lt;/p&gt;

&lt;p&gt;Before publishing, ask whether the piece contains a claim, a reader, and a reason. Remove the paragraph that only restates the title. Link the source that carries the fact. Replace a generic example with the actual case that changed your mind. If the answer is not ready, keep it as a note.&lt;/p&gt;

&lt;p&gt;Before merging, ask whether the author can narrate the change. Ask what was tested, what assumptions remain, and what would make the design wrong. Keep the change small enough for a reviewer to understand. These are not anti-AI rules. They are the old rules of responsible work, made urgent by a new rate of output.&lt;/p&gt;

&lt;p&gt;At the platform level, reward useful material rather than the frequency of posting. LinkedIn’s own guidance makes the same broad distinction: AI-assisted content is welcome when it reflects a real person’s perspective, experience, or expertise; generic and repetitive material is the problem.&lt;sup&gt;11&lt;/sup&gt; That is a reasonable direction, even if no feedback button can see the full process behind a post.&lt;/p&gt;

&lt;p&gt;The word “slop” will remain imprecise. It may even become too broad to survive. The underlying problem will not disappear with the vocabulary: when production gets easier, judgment becomes the scarce resource.&lt;/p&gt;

&lt;p&gt;That is the thing worth defending. Not human typing for its own sake. Not a nostalgia for artificial scarcity. The expectation that when a person asks for our attention, someone has done the work of deciding why we should give it.&lt;/p&gt;

&lt;h3&gt;
  
  
  In practice
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Use AI to increase the quality of decisions, not only the quantity of artifacts.&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Publish less when less is all you have.&lt;/strong&gt; Silence is cheaper than asking strangers to sort through a draft.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Stand behind the work.&lt;/strong&gt; That is what authorship has always meant.&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;&lt;a href="https://tech.anujsadani.in/ai-slop/" rel="noopener noreferrer"&gt;Read the original on Anuj Sadani | Tech&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>machinelearning</category>
      <category>devtools</category>
      <category>programming</category>
    </item>
    <item>
      <title>Tool Selection at Scale: What Top-K RAG Leaves Out</title>
      <dc:creator>Anuj Sadani</dc:creator>
      <pubDate>Sun, 01 Nov 2026 03:30:00 +0000</pubDate>
      <link>https://dev.to/anuj-sadani/tool-selection-at-scale-what-top-k-rag-leaves-out-4aj9</link>
      <guid>https://dev.to/anuj-sadani/tool-selection-at-scale-what-top-k-rag-leaves-out-4aj9</guid>
      <description>&lt;p&gt;Embed, retrieve, let the LLM decide works for about 100 tools. At 1,000 it needs a control plane: hybrid retrieval, hard filters, a learned reranker, progressive disclosure and per-stage evals.&lt;/p&gt;

&lt;p&gt;This is an adapted edition of the original article. It preserves the section argument and evidence; use the source link for the published version and any interactive material.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://tech.anujsadani.in/tool-selection-at-scale/" rel="noopener noreferrer"&gt;Read the original article&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;01 / Baseline&lt;/p&gt;

&lt;h2&gt;
  
  
  What the pipeline gets right
&lt;/h2&gt;

&lt;p&gt;The idea is that a reliable classifier is a pipeline, not a prompt. Embeddings find the neighborhood. Retrieval returns definitions. The LLM compares them. A validation step chooses between a category and an abstention. I would keep every stage.&lt;/p&gt;

&lt;p&gt;The pipeline has a ceiling, and it deserves one extension. If the correct category is not retrieved, the LLM cannot pick it. True. It is equally true of every stage after retrieval, and that is where scale starts to hurt.&lt;/p&gt;

&lt;p&gt;At around 100 items the plain pipeline is enough. A 100-tool catalog at roughly 400 tokens per schema is 40k tokens. That figure is the one I used in my &lt;a href="https://arxiv.org/abs/2604.21816" rel="noopener noreferrer"&gt;Tool Attention paper&lt;/a&gt;, and real servers run higher: the public audits I cited put the full GitHub server at about 55,000 tokens for 93 tools, close to 590 each. Forty thousand tokens is wasteful but it fits, and a single top-10 embedding lookup rarely misses.&lt;/p&gt;

&lt;p&gt;At 1,000 tools the same arithmetic gives 400k tokens. Retrieval stops being an optimization and becomes the only way in. That is the point where embed, retrieve, decide and validate is necessary and no longer sufficient. The rest of this post is about the parts I would add.&lt;/p&gt;

&lt;p&gt;02 / Framing&lt;/p&gt;

&lt;h2&gt;
  
  
  Tools are not categories
&lt;/h2&gt;

&lt;p&gt;It is tempting to treat tool selection as classification with a bigger label set. The analogy holds for the retrieval half and breaks for everything around it.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Wrong answer&lt;/strong&gt;: &lt;strong&gt;Category classification:&lt;/strong&gt; A wrong value in a row. Fixable later.; &lt;strong&gt;Tool selection:&lt;/strong&gt; A side effect: a sent email, a deleted record, a charged card.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ground truth&lt;/strong&gt;: &lt;strong&gt;Category classification:&lt;/strong&gt; A taxonomy that changes slowly.; &lt;strong&gt;Tool selection:&lt;/strong&gt; The catalog, plus per-user permissions, plus live connection state. It changes per session.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Decisions&lt;/strong&gt;: &lt;strong&gt;Category classification:&lt;/strong&gt; One per input.; &lt;strong&gt;Tool selection:&lt;/strong&gt; A sequence. Each observation changes which tools are relevant next.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Near-duplicates&lt;/strong&gt;: &lt;strong&gt;Category classification:&lt;/strong&gt; Adjacent categories.; &lt;strong&gt;Tool selection:&lt;/strong&gt; Overlapping tools from different servers, often three different &lt;code&gt;search&lt;/code&gt; tools.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Abstention&lt;/strong&gt;: &lt;strong&gt;Category classification:&lt;/strong&gt; One exit: insufficient evidence.; &lt;strong&gt;Tool selection:&lt;/strong&gt; Three exits: ask the user, widen the search, or report the missing capability.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Trust&lt;/strong&gt;: &lt;strong&gt;Category classification:&lt;/strong&gt; The input text is untrusted.; &lt;strong&gt;Tool selection:&lt;/strong&gt; The definitions are untrusted too. A tool description is text the model obeys.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Row one is the one that changes the architecture. Safety cannot live in the prompt when a wrong choice writes to a production system. It has to live in code that runs regardless of what the model says.&lt;/p&gt;

&lt;p&gt;03 / Why retrieval at all&lt;/p&gt;

&lt;h2&gt;
  
  
  It is a context-quality problem, not only a knowledge problem
&lt;/h2&gt;

&lt;p&gt;The baseline pipeline frames retrieval as supplying information the model lacks. That is half of it. The other half is that a model with all the information in a crowded context still does worse, even when the right answer is technically in there.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Position effects.&lt;/strong&gt; Liu et al. (&lt;a href="https://aclanthology.org/2024.tacl-1.9/" rel="noopener noreferrer"&gt;Lost in the Middle&lt;/a&gt;, TACL 2024) found that accuracy is lowest when the relevant item sits in the middle of a long context. A tool list is exactly that shape.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Tool overload, measured.&lt;/strong&gt; RAG-MCP (&lt;a href="https://arxiv.org/abs/2505.03275" rel="noopener noreferrer"&gt;Gan and Sun, 2025&lt;/a&gt;) reports tool selection accuracy of 43.13% with retrieval against 13.62% for the baseline that shows the model everything, and prompt tokens cut by more than half.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Vendor numbers agree in direction.&lt;/strong&gt; Anthropic's &lt;a href="https://www.anthropic.com/engineering/advanced-tool-use" rel="noopener noreferrer"&gt;tool search&lt;/a&gt; write-up reports Opus 4 going from 49% to 74% on their MCP evaluation and Opus 4.5 from 79.5% to 88.1%, with about 85% fewer tokens.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;My own number is a simulation.&lt;/strong&gt; On a calibrated 120-tool, six-server catalog, gating cut per-turn tool tokens from 47.3k to 2.4k (95.0%). No live agent ran. Treat it as an arithmetic result, not an accuracy result.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;These are four different benchmarks, so I am not comparing them to each other. They agree on the direction, and that is enough for the design rule: retrieval is a context-control mechanism. Expose the smallest decision space that still contains the answer.&lt;/p&gt;

&lt;p&gt;04 / Measurement&lt;/p&gt;

&lt;h2&gt;
  
  
  The recall chain
&lt;/h2&gt;

&lt;p&gt;Once you add a reranker, the ceiling is no longer one number. Every stage can drop the right tool, and the drops compound. The precise form is a product of conditional recalls:&lt;/p&gt;

&lt;p&gt;End-to-end recall&lt;/p&gt;

&lt;p&gt;P(right&amp;nbsp;tool&amp;nbsp;used)&amp;nbsp;=&amp;nbsp;&amp;nbsp;&amp;nbsp;R_retrieve&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;#&amp;nbsp;gold&amp;nbsp;is&amp;nbsp;in&amp;nbsp;the&amp;nbsp;candidate&amp;nbsp;set\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;*&amp;nbsp;R_filter&amp;nbsp;&amp;nbsp;|&amp;nbsp;retrieved&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;#&amp;nbsp;gold&amp;nbsp;survives&amp;nbsp;policy&amp;nbsp;filters\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;*&amp;nbsp;R_menu&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;|&amp;nbsp;filtered&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;#&amp;nbsp;gold&amp;nbsp;survives&amp;nbsp;the&amp;nbsp;reranker&amp;nbsp;cut\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;*&amp;nbsp;R_select&amp;nbsp;&amp;nbsp;|&amp;nbsp;menu&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;#&amp;nbsp;the&amp;nbsp;LLM&amp;nbsp;picks&amp;nbsp;it&amp;nbsp;from&amp;nbsp;the&amp;nbsp;menu&lt;/p&gt;

&lt;p&gt;The word "conditional" matters. Multiplying raw retrieval recall by raw reranker recall double counts, because the reranker only ever sees what retrieval returned. Here is an illustration with numbers I made up to show the shape, not measurements:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Retrieve 40&lt;/strong&gt;: &lt;strong&gt;Conditional recall:&lt;/strong&gt; 0.95; &lt;strong&gt;Cumulative:&lt;/strong&gt; 0.950&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hard filters&lt;/strong&gt;: &lt;strong&gt;Conditional recall:&lt;/strong&gt; 0.98; &lt;strong&gt;Cumulative:&lt;/strong&gt; 0.931&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Rerank to 6&lt;/strong&gt;: &lt;strong&gt;Conditional recall:&lt;/strong&gt; 0.92; &lt;strong&gt;Cumulative:&lt;/strong&gt; 0.857&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;LLM selects&lt;/strong&gt;: &lt;strong&gt;Conditional recall:&lt;/strong&gt; 0.90; &lt;strong&gt;Cumulative:&lt;/strong&gt; 0.771&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Raising retrieval from 0.95 to 0.99 lifts the end result by about three points. Raising selection from 0.90 to 0.95 lifts it by about four. The weak stage is rarely the one people tune first, because retrieval is the stage everyone knows how to tune. The fix is to log four sets per request and compute the numbers per stage:&lt;/p&gt;

&lt;p&gt;stage_recall.py&lt;/p&gt;

&lt;p&gt;STAGES&amp;nbsp;=&amp;nbsp;["retrieved",&amp;nbsp;"filtered",&amp;nbsp;"menu",&amp;nbsp;"chosen"]&amp;nbsp;&amp;nbsp;&amp;nbsp;#&amp;nbsp;each&amp;nbsp;is&amp;nbsp;a&amp;nbsp;set&amp;nbsp;of&amp;nbsp;tool_ids&amp;nbsp;on&amp;nbsp;the&amp;nbsp;trace\&lt;br&gt;
\&lt;br&gt;
def&amp;nbsp;stage_recall(traces):\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;alive,&amp;nbsp;out&amp;nbsp;=&amp;nbsp;traces,&amp;nbsp;{}\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;for&amp;nbsp;stage&amp;nbsp;in&amp;nbsp;STAGES:\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;kept&amp;nbsp;=&amp;nbsp;[t&amp;nbsp;for&amp;nbsp;t&amp;nbsp;in&amp;nbsp;alive&amp;nbsp;if&amp;nbsp;t.gold&amp;nbsp;in&amp;nbsp;getattr(t,&amp;nbsp;stage)]\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;out[stage]&amp;nbsp;=&amp;nbsp;len(kept)&amp;nbsp;/&amp;nbsp;max(len(alive),&amp;nbsp;1)&amp;nbsp;&amp;nbsp;&amp;nbsp;#&amp;nbsp;conditional&amp;nbsp;on&amp;nbsp;the&amp;nbsp;previous&amp;nbsp;stage\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;alive&amp;nbsp;=&amp;nbsp;kept\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;return&amp;nbsp;out&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;#&amp;nbsp;multiply&amp;nbsp;the&amp;nbsp;values&amp;nbsp;for&amp;nbsp;end-to-end&lt;/p&gt;

&lt;p&gt;Even in the baseline pipeline, evaluating only the LLM is misleading, so you track retrieval and selection separately. At scale two numbers are not enough. You need four, and you need a labeled set where the gold tool is one the user could actually call.&lt;/p&gt;

&lt;p&gt;One caveat: the chain above starts at retrieval, but in the per-subgoal design of section 06 there is a stage before it, splitting the request into subgoals. A bad split caps everything downstream, so give it its own term, &lt;code&gt;R_decompose&lt;/code&gt;: the share of requests where the gold tool is covered by some extracted subgoal.&lt;/p&gt;

&lt;p&gt;05 / Control part 1&lt;/p&gt;

&lt;h2&gt;
  
  
  The catalog is a product
&lt;/h2&gt;

&lt;p&gt;The first thing that breaks at scale is not the model or the index. It is the catalog. In my simulation 29 of 500 tasks failed with tool gating. Fourteen (48%) came from a query matching several semantically similar tools. Seven (24%) came from cryptic legacy descriptions. That is 72% of failures where what the catalog says was the proximate cause. Five more (17%) were multi-hop cases, covered below, and the last three fit none of these buckets. It is a simulated workload, so read the proportions as a hypothesis worth testing on your own traces.&lt;/p&gt;

&lt;p&gt;So keep the registry outside the prompt and give every tool a card written for retrieval, not for the implementer:&lt;/p&gt;

&lt;p&gt;tool card (indexed fields in the top block, the rest loaded on demand)&lt;/p&gt;

&lt;p&gt;{\&lt;br&gt;
&amp;nbsp;&amp;nbsp;"tool_id":&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;"github.search_code",\&lt;br&gt;
&amp;nbsp;&amp;nbsp;"domain":&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;"source_control",\&lt;br&gt;
&amp;nbsp;&amp;nbsp;"summary":&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;"Find&amp;nbsp;where&amp;nbsp;something&amp;nbsp;is&amp;nbsp;implemented&amp;nbsp;across&amp;nbsp;connected&amp;nbsp;repositories",\&lt;br&gt;
&amp;nbsp;&amp;nbsp;"capabilities":&amp;nbsp;["code_search",&amp;nbsp;"repository_read"],\&lt;br&gt;
&amp;nbsp;&amp;nbsp;"entities":&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;["repository",&amp;nbsp;"file",&amp;nbsp;"commit"],\&lt;br&gt;
&amp;nbsp;&amp;nbsp;"input_concepts":&amp;nbsp;["query",&amp;nbsp;"repository",&amp;nbsp;"organization"],\&lt;br&gt;
&amp;nbsp;&amp;nbsp;"prefer_over":&amp;nbsp;&amp;nbsp;["web.search"],\&lt;br&gt;
&amp;nbsp;&amp;nbsp;"not_for":&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;"Reading&amp;nbsp;one&amp;nbsp;known&amp;nbsp;file.&amp;nbsp;Use&amp;nbsp;github.read_file.",\&lt;br&gt;
&amp;nbsp;&amp;nbsp;"examples":&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;["Find&amp;nbsp;where&amp;nbsp;login&amp;nbsp;sessions&amp;nbsp;are&amp;nbsp;validated&amp;nbsp;in&amp;nbsp;the&amp;nbsp;checkout-service&amp;nbsp;repository"],\&lt;br&gt;
\&lt;br&gt;
&amp;nbsp;&amp;nbsp;//&amp;nbsp;not&amp;nbsp;indexed:&amp;nbsp;fetched&amp;nbsp;only&amp;nbsp;after&amp;nbsp;the&amp;nbsp;tool&amp;nbsp;is&amp;nbsp;selected\&lt;br&gt;
&amp;nbsp;&amp;nbsp;"requires":&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;{&amp;nbsp;"connection":&amp;nbsp;"github",&amp;nbsp;"scopes":&amp;nbsp;["repo:read"]&amp;nbsp;},\&lt;br&gt;
&amp;nbsp;&amp;nbsp;"risk":&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;"read_only",\&lt;br&gt;
&amp;nbsp;&amp;nbsp;"cost":&amp;nbsp;"low",&amp;nbsp;"p50_ms":&amp;nbsp;900,\&lt;br&gt;
&amp;nbsp;&amp;nbsp;"owner":&amp;nbsp;"platform-team",&amp;nbsp;"schema_version":&amp;nbsp;7,\&lt;br&gt;
&amp;nbsp;&amp;nbsp;"input_schema":&amp;nbsp;{&amp;nbsp;/*&amp;nbsp;full&amp;nbsp;JSON&amp;nbsp;Schema&amp;nbsp;*/&amp;nbsp;}\&lt;br&gt;
}&lt;/p&gt;

&lt;p&gt;Three fields do more work than they look. &lt;code&gt;summary&lt;/code&gt; is written in the voice of the user's intent. In my paper, regenerating summaries that way cut their average length by 63% and raised retrieval F1 by 8 points, on the same synthetic benchmark. &lt;code&gt;not_for&lt;/code&gt; and &lt;code&gt;prefer_over&lt;/code&gt; are the disambiguation fields. They give the reranker and the model a reason to choose between near-duplicates, and near-duplicates were the largest failure bucket. &lt;code&gt;owner&lt;/code&gt; and &lt;code&gt;schema_version&lt;/code&gt; exist because at 1,000 tools somebody has to be paged when a description rots.&lt;/p&gt;

&lt;p&gt;One structural change belongs here too. A flat top-k over thousands of tools tends to fill the shortlist with ten variants of one family and crowd out the second family you needed. Group tools into domains and route in two hops, domain first and tool second. This is my rule from building these systems, not a measured result. Test it against a flat index on your own catalog before you trust it.&lt;/p&gt;

&lt;p&gt;06 / Control part 2&lt;/p&gt;

&lt;h2&gt;
  
  
  Discovery per subgoal, three retrievers, one pre-filter
&lt;/h2&gt;

&lt;p&gt;The baseline pipeline embeds the input once. A tool request is rarely one intent. "Find where sessions are validated in checkout-service and check whether the fix is deployed" is two subgoals with different tools. Embedding the whole sentence returns a blurred neighborhood that serves neither. I extract structured intent first and run discovery per subgoal:&lt;/p&gt;

&lt;p&gt;intent extraction output&lt;/p&gt;

&lt;p&gt;{\&lt;br&gt;
&amp;nbsp;&amp;nbsp;"subgoals":&amp;nbsp;[\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;{&amp;nbsp;"goal":&amp;nbsp;"locate&amp;nbsp;session&amp;nbsp;validation&amp;nbsp;code",\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;"operations":&amp;nbsp;["repository_search",&amp;nbsp;"code_read"],\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;"entities":&amp;nbsp;["checkout-service&amp;nbsp;repository"],\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;"required_access":&amp;nbsp;["GitHub"],&amp;nbsp;"mutation":&amp;nbsp;false&amp;nbsp;},\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;{&amp;nbsp;"goal":&amp;nbsp;"check&amp;nbsp;deployment&amp;nbsp;status",\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;"operations":&amp;nbsp;["rollout_inspect"],\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;"entities":&amp;nbsp;["checkout-service"],\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;"required_access":&amp;nbsp;["Kubernetes"],&amp;nbsp;"mutation":&amp;nbsp;false&amp;nbsp;}\&lt;br&gt;
&amp;nbsp;&amp;nbsp;]\&lt;br&gt;
}&lt;/p&gt;

&lt;p&gt;Retrieval then runs three ways and fuses the ranks. Vectors catch paraphrase. BM25 catches the exact tokens embeddings blur: tool ids, product names, error codes. Capability matching uses the structured &lt;code&gt;operations&lt;/code&gt; field. Reciprocal rank fusion (Cormack et al., SIGIR 2009, constant 60) merges them without needing comparable scores.&lt;/p&gt;

&lt;p&gt;Two more signals feed the same stage: historical success on similar tasks, and whether the connector is available in this environment. I use history as a ranker feature and availability as a pre-filter. False positives are acceptable here and a miss is not. A missing tool costs the whole task, while an irrelevant one costs a few tokens that the reranker removes.&lt;/p&gt;

&lt;p&gt;The detail people miss is the pre-filter. If you filter for permissions after retrieval, the top 40 can be full of tools this user cannot call, and your recall of &lt;em&gt;usable&lt;/em&gt; tools collapses while the recall you measure looks fine. Push the static predicates into the index query.&lt;/p&gt;

&lt;p&gt;discover.py (sketch)&lt;/p&gt;

&lt;p&gt;def&amp;nbsp;rrf(rankings,&amp;nbsp;k=60):\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;score&amp;nbsp;=&amp;nbsp;defaultdict(float)\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;for&amp;nbsp;ranking&amp;nbsp;in&amp;nbsp;rankings:\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;for&amp;nbsp;rank,&amp;nbsp;tool_id&amp;nbsp;in&amp;nbsp;enumerate(ranking,&amp;nbsp;start=1):\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;score[tool_id]&amp;nbsp;+=&amp;nbsp;1.0&amp;nbsp;/&amp;nbsp;(k&amp;nbsp;+&amp;nbsp;rank)\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;return&amp;nbsp;sorted(score,&amp;nbsp;key=score.get,&amp;nbsp;reverse=True)\&lt;br&gt;
\&lt;br&gt;
def&amp;nbsp;discover(sub,&amp;nbsp;ctx,&amp;nbsp;index,&amp;nbsp;n=40,&amp;nbsp;menu=6):\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;where&amp;nbsp;=&amp;nbsp;{"connection":&amp;nbsp;ctx.connections,&amp;nbsp;"env":&amp;nbsp;ctx.env,\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;"scopes_within":&amp;nbsp;ctx.granted_scopes}&amp;nbsp;&amp;nbsp;&amp;nbsp;#&amp;nbsp;static&amp;nbsp;per&amp;nbsp;session,&amp;nbsp;pushed&amp;nbsp;into&amp;nbsp;the&amp;nbsp;index\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;fused&amp;nbsp;=&amp;nbsp;rrf([\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;index.vector(sub.text,&amp;nbsp;n=n,&amp;nbsp;where=where),\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;index.bm25(sub.text,&amp;nbsp;n=n,&amp;nbsp;where=where),\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;index.by_capability(sub.operations,&amp;nbsp;n=n,&amp;nbsp;where=where),\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;])[:n]\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;allowed&amp;nbsp;=&amp;nbsp;[t&amp;nbsp;for&amp;nbsp;t&amp;nbsp;in&amp;nbsp;fused&amp;nbsp;if&amp;nbsp;request_ok(t,&amp;nbsp;sub,&amp;nbsp;ctx)]&amp;nbsp;&amp;nbsp;&amp;nbsp;#&amp;nbsp;depends&amp;nbsp;only&amp;nbsp;on&amp;nbsp;this&amp;nbsp;subgoal\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;ranked&amp;nbsp;=&amp;nbsp;reranker.score(sub,&amp;nbsp;allowed)&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;#&amp;nbsp;learned,&amp;nbsp;relevance&amp;nbsp;only\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;return&amp;nbsp;prefer_cheap_within(ranked,&amp;nbsp;margin=0.05)[:menu]\&lt;br&gt;
\&lt;br&gt;
def&amp;nbsp;request_ok(tool,&amp;nbsp;sub,&amp;nbsp;ctx):\&lt;br&gt;
&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;return&amp;nbsp;(tool.risk&amp;nbsp;!=&amp;nbsp;"read_only")&amp;nbsp;==&amp;nbsp;sub.mutation&amp;nbsp;&amp;nbsp;&amp;nbsp;#&amp;nbsp;write&amp;nbsp;tools&amp;nbsp;for&amp;nbsp;write&amp;nbsp;subgoals,&amp;nbsp;read&amp;nbsp;tools&amp;nbsp;for&amp;nbsp;read&lt;/p&gt;

&lt;p&gt;07 / Control part 3&lt;/p&gt;

&lt;h2&gt;
  
  
  Filter, then rank, and where I would change the score
&lt;/h2&gt;

&lt;p&gt;Hard filters are deterministic policy, not LLM judgment. They remove a missing connection or scope, incompatible data types, a write tool for a read-only request (and a read-only tool for a write request), an action that policy disallows or that is riskier than the request needs, and any tool unavailable in this environment. They run before any model sees a candidate, and they are cheap to test.&lt;/p&gt;

&lt;p&gt;Each stage optimizes something different, and it helps to say so out loud:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Retrieval&lt;/strong&gt; optimizes recall. Missing the right tool is the expensive error.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Filters&lt;/strong&gt; optimize correctness. They remove what cannot or should not run.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The reranker&lt;/strong&gt; optimizes precision and order. It puts the strongest candidates first.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The LLM&lt;/strong&gt; does a bounded comparison over a small menu.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Validation&lt;/strong&gt; checks that the choice is supported and safe to run.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The reranker then cuts roughly 30 to 50 candidates to a menu of 5 to 8. I keep the cut generous. A reranker error deletes the right tool for good, so this is the stage where I trade a few extra tokens for recall. In my paper I recommended k of 8 to 12 with the threshold doing the precision work. The examples here use 6 because they show one-line menus; widen it if your measured &lt;code&gt;R_menu&lt;/code&gt; is low.&lt;/p&gt;

&lt;p&gt;A common way to specify the reranker is a weighted sum over the candidate tool &lt;code&gt;t&lt;/code&gt;:&lt;/p&gt;

&lt;p&gt;Reranking score&lt;/p&gt;

&lt;p&gt;S(t)&amp;nbsp;=&amp;nbsp;w1*I&amp;nbsp;+&amp;nbsp;w2*C&amp;nbsp;+&amp;nbsp;w3*E&amp;nbsp;+&amp;nbsp;w4*P&amp;nbsp;+&amp;nbsp;w5*H&amp;nbsp;-&amp;nbsp;w6*R&amp;nbsp;-&amp;nbsp;w7*L&amp;nbsp;-&amp;nbsp;w8*K&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;I&lt;/code&gt; intent match, &lt;code&gt;C&lt;/code&gt; capability match, &lt;code&gt;E&lt;/code&gt; entity or data-source match.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;P&lt;/code&gt; permission and availability, &lt;code&gt;H&lt;/code&gt; historical success on similar tasks.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;R&lt;/code&gt; operational risk, &lt;code&gt;L&lt;/code&gt; expected latency, &lt;code&gt;K&lt;/code&gt; execution cost.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The weights should move with the task. For a sensitive action, permission and risk count for more. For an interactive request, relevance and latency count for more. I like the inputs and I like that instinct. I would change how the terms combine:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Permission&lt;/strong&gt;: &lt;strong&gt;Problem inside a weighted sum:&lt;/strong&gt; A missing scope is not a small penalty. It is a no. It also double counts with the hard filter.; &lt;strong&gt;Where I put it instead:&lt;/strong&gt; Filter only.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Risk&lt;/strong&gt;: &lt;strong&gt;Problem inside a weighted sum:&lt;/strong&gt; Subtracting risk lets a slightly more relevant destructive tool beat a safe adequate one, or the reverse, and nobody audits the weights.; &lt;strong&gt;Where I put it instead:&lt;/strong&gt; A risk policy after selection: confirm, dry-run or deny.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cost, latency&lt;/strong&gt;: &lt;strong&gt;Problem inside a weighted sum:&lt;/strong&gt; They should never outvote relevance.; &lt;strong&gt;Where I put it instead:&lt;/strong&gt; A tiebreaker among tools within a small relevance margin.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;History&lt;/strong&gt;: &lt;strong&gt;Problem inside a weighted sum:&lt;/strong&gt; Popular tools collect more history and crowd out better new ones.; &lt;strong&gt;Where I put it instead:&lt;/strong&gt; A learned feature, with deliberate exploration of new tools.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hand-set weights&lt;/strong&gt;: &lt;strong&gt;Problem inside a weighted sum:&lt;/strong&gt; They do not survive a catalog change.; &lt;strong&gt;Where I put it instead:&lt;/strong&gt; Train the relevance ranker on (subgoal, tool used, success) traces.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The advice to weight permission and risk more heavily for sensitive actions points the same way. Past some weight they stop being weights and become gates.&lt;/p&gt;

&lt;p&gt;The rule I use&lt;/p&gt;

&lt;p&gt;Anything that must never happen belongs in a filter or a policy. Anything that is a preference belongs in the ranker. A weighted sum blurs the two, and the blur is where incidents come from.&lt;/p&gt;

&lt;p&gt;08 / Control part 4&lt;/p&gt;

&lt;h2&gt;
  
  
  Progressive disclosure
&lt;/h2&gt;

&lt;p&gt;Retrieval narrows which tools are visible. Progressive disclosure narrows how much the model sees about each one. The token figures below are rough, and they follow from the 400-token schema assumption above.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Discovery&lt;/strong&gt;: &lt;strong&gt;The model sees:&lt;/strong&gt; Name and one-line capability; &lt;strong&gt;~Tokens per tool:&lt;/strong&gt; 30 to 40; &lt;strong&gt;Trigger:&lt;/strong&gt; Every subgoal&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Selection&lt;/strong&gt;: &lt;strong&gt;The model sees:&lt;/strong&gt; Constraints, risk, cost, one example; &lt;strong&gt;~Tokens per tool:&lt;/strong&gt; 100 to 150; &lt;strong&gt;Trigger:&lt;/strong&gt; Only when two menu items are close&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Invocation&lt;/strong&gt;: &lt;strong&gt;The model sees:&lt;/strong&gt; Full input schema; &lt;strong&gt;~Tokens per tool:&lt;/strong&gt; 400+; &lt;strong&gt;Trigger:&lt;/strong&gt; Only for the chosen tool&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Recovery&lt;/strong&gt;: &lt;strong&gt;The model sees:&lt;/strong&gt; The error, alternatives, a wider menu; &lt;strong&gt;~Tokens per tool:&lt;/strong&gt; varies; &lt;strong&gt;Trigger:&lt;/strong&gt; After a failure or a rejected call&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A six-tool menu at discovery level is about 200 tokens. The same six with full schemas is about 2,400. The saving is real, and so is the risk: the model chooses from one-liners, so the one-liners must be good. That takes you straight back to the catalog. This is all the model sees at the discovery level:&lt;/p&gt;

&lt;p&gt;discovery-level menu&lt;/p&gt;

&lt;p&gt;1.&amp;nbsp;github.search_code&amp;nbsp;&amp;nbsp;&amp;nbsp;Search&amp;nbsp;code&amp;nbsp;inside&amp;nbsp;connected&amp;nbsp;repositories\&lt;br&gt;
2.&amp;nbsp;github.read_file&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;Read&amp;nbsp;a&amp;nbsp;specific&amp;nbsp;repository&amp;nbsp;file\&lt;br&gt;
3.&amp;nbsp;jira.search_issues&amp;nbsp;&amp;nbsp;&amp;nbsp;Search&amp;nbsp;issues&amp;nbsp;and&amp;nbsp;tickets\&lt;br&gt;
4.&amp;nbsp;web.search&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;Search&amp;nbsp;public&amp;nbsp;internet&amp;nbsp;sources&lt;/p&gt;

&lt;p&gt;Only after it picks &lt;code&gt;github.search_code&lt;/code&gt; does the system reveal that tool's full input schema. Two operational details. First, keep the summary pool in the stable prefix of the prompt and put per-turn schemas right before the user message. In my simulation this gave an 84% prompt-cache hit rate over a 30-turn session against 22% for naive full-schema injection, because the tool list stopped invalidating the cache. This only works if the schemas travel as message content after the cached prefix. On the Claude API, tool definitions sit at the front of the prefix, so changing the &lt;code&gt;tools&lt;/code&gt; parameter between turns invalidates everything behind it. Second, give the model an escape hatch: a &lt;code&gt;search_tools&lt;/code&gt; meta-tool it can call when the menu is wrong. Anthropic's tool search is the same idea shipped as a platform feature.&lt;/p&gt;

&lt;p&gt;Pair that with a gate: if the model calls a tool that is not in the active set, reject it with a structured error listing what is available. In my simulation the gate fired on 2.3% of turns. The model recovered on the next turn 78% of the time and asked the user a clarifying question the other 22%.&lt;/p&gt;

&lt;p&gt;09 / Control parts 5 and 6&lt;/p&gt;

&lt;h2&gt;
  
  
  Validate, iterate, recover
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Validate before executing.&lt;/strong&gt; The model may pick one tool or a short sequence from the menu, and each call gets deterministic checks on required parameters, types, permissions, read versus write intent, confirmation requirements and target scope. For writes I add a dry-run where the tool supports one. Feed validator errors back to the model verbatim. They are the cheapest correction signal in the system.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Return observations, not payloads.&lt;/strong&gt; The executor should cap and normalize the result, and keep a handle to the full output outside the context. A 200 KB response dumped into the prompt undoes everything the earlier stages saved.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Rediscover on every subgoal.&lt;/strong&gt; A tool's output can reveal the next need: a config reference that points to a deployment system you had no tools for. In my simulation 5 of the 29 failures (17%) were multi-hop cases where the right tool only became relevant after an intermediate result. Re-embedding the query after each observation partly fixed them. The agent should not carry every plausibly useful tool for the whole session.&lt;/p&gt;

&lt;p&gt;one request, three rounds of discovery (illustrative)&lt;/p&gt;

&lt;p&gt;search&amp;nbsp;the&amp;nbsp;repository&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;#&amp;nbsp;round&amp;nbsp;1:&amp;nbsp;github.search_code\&lt;br&gt;
&amp;nbsp;&amp;nbsp;-&amp;gt;&amp;nbsp;found&amp;nbsp;a&amp;nbsp;config&amp;nbsp;reference\&lt;br&gt;
config&amp;nbsp;says&amp;nbsp;it&amp;nbsp;ships&amp;nbsp;as&amp;nbsp;a&amp;nbsp;Kubernetes&amp;nbsp;rollout&amp;nbsp;#&amp;nbsp;new&amp;nbsp;subgoal\&lt;br&gt;
retrieve&amp;nbsp;Kubernetes&amp;nbsp;tools&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;#&amp;nbsp;round&amp;nbsp;2:&amp;nbsp;k8s.get_rollout_status\&lt;br&gt;
read&amp;nbsp;the&amp;nbsp;failing&amp;nbsp;pod's&amp;nbsp;logs&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;#&amp;nbsp;round&amp;nbsp;3:&amp;nbsp;k8s.get_pod_logs&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Name the failure class.&lt;/strong&gt; "The agent used the wrong tool" is not a diagnosis. There are at least seven distinct failures, each with a different detector and a different fix:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;No suitable tool exists&lt;/strong&gt;: &lt;strong&gt;How you see it:&lt;/strong&gt; Gold label is "none". Abstention rate on those cases.; &lt;strong&gt;Where the fix lives:&lt;/strong&gt; Report the missing capability. Roadmap input.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Request mis-split&lt;/strong&gt;: &lt;strong&gt;How you see it:&lt;/strong&gt; Low &lt;code&gt;R_decompose&lt;/code&gt;: the gold tool's need is missing from, or merged into, another subgoal; &lt;strong&gt;Where the fix lives:&lt;/strong&gt; Intent-extraction prompt, examples and output schema.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Exists, not retrieved&lt;/strong&gt;: &lt;strong&gt;How you see it:&lt;/strong&gt; Low &lt;code&gt;R_retrieve&lt;/code&gt;; &lt;strong&gt;Where the fix lives:&lt;/strong&gt; Catalog text, hybrid retrieval, intent decomposition.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Retrieved, cut by reranker&lt;/strong&gt;: &lt;strong&gt;How you see it:&lt;/strong&gt; Low &lt;code&gt;R_menu&lt;/code&gt;; &lt;strong&gt;Where the fix lives:&lt;/strong&gt; Reranker training data, a larger menu, &lt;code&gt;not_for&lt;/code&gt; fields.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Shown, not chosen&lt;/strong&gt;: &lt;strong&gt;How you see it:&lt;/strong&gt; Low &lt;code&gt;R_select&lt;/code&gt;; &lt;strong&gt;Where the fix lives:&lt;/strong&gt; Better one-liners, a selection-level disclosure step.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Chosen, invoked wrongly&lt;/strong&gt;: &lt;strong&gt;How you see it:&lt;/strong&gt; Validator rejection rate per tool; &lt;strong&gt;Where the fix lives:&lt;/strong&gt; Examples in the schema, clearer parameter names.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ran, returned too little&lt;/strong&gt;: &lt;strong&gt;How you see it:&lt;/strong&gt; Observation check fails; the subgoal is still open; &lt;strong&gt;Where the fix lives:&lt;/strong&gt; Next tool in the candidate list, or a new discovery round.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Recovery paths follow from the table. Low discovery confidence: widen retrieval, reformulate the intent, or ask the user. A selected tool fails: return to the candidate list and take the next compatible one. Nothing viable: say which capability or connection is missing. Do not let the model invent an answer to fill the gap.&lt;/p&gt;

&lt;p&gt;10 / Worked example&lt;/p&gt;

&lt;h2&gt;
  
  
  One request, end to end
&lt;/h2&gt;

&lt;p&gt;Everything above is easier to judge on a single request. This is an illustration, not a benchmark: the catalog size, the counts and the tool names are mine, chosen to be realistic. The catalog has 1,200 tools across 30 servers, and the request is:&lt;/p&gt;

&lt;p&gt;&lt;em&gt;"The checkout-service deploy failed last night. Find what changed, check the rollout, and tell the payments channel on Slack."&lt;/em&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Intent extraction&lt;/strong&gt;: &lt;strong&gt;What happens:&lt;/strong&gt; Three subgoals: find what changed (GitHub, read), check the rollout (Kubernetes, read), post a summary (Slack, write).; &lt;strong&gt;The model sees:&lt;/strong&gt; The request only.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Retrieval, per subgoal&lt;/strong&gt;: &lt;strong&gt;What happens:&lt;/strong&gt; 40 candidates each, out of 1,200, after the connection and scope pre-filter (this user has no Argo CD connection, so those tools never enter the query). For the rollout subgoal, BM25 catches the literal word "rollout" in &lt;code&gt;k8s.get_rollout_status&lt;/code&gt;, which the vector search had ranked 63rd, outside its top 40.; &lt;strong&gt;The model sees:&lt;/strong&gt; Nothing yet.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hard filters&lt;/strong&gt;: &lt;strong&gt;What happens:&lt;/strong&gt; Rollout subgoal: &lt;code&gt;k8s.rollout_restart&lt;/code&gt; and nine other write tools removed because the subgoal is read-only. 30 remain.; &lt;strong&gt;The model sees:&lt;/strong&gt; Nothing yet.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Rerank&lt;/strong&gt;: &lt;strong&gt;What happens:&lt;/strong&gt; 30 become a menu of 6. The &lt;code&gt;not_for&lt;/code&gt; field on &lt;code&gt;logs.search&lt;/code&gt; ("for Kubernetes pod logs use &lt;code&gt;k8s.get_pod_logs&lt;/code&gt;") settles a near-duplicate.; &lt;strong&gt;The model sees:&lt;/strong&gt; Nothing yet.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Disclosure and selection&lt;/strong&gt;: &lt;strong&gt;What happens:&lt;/strong&gt; The menu is shown at discovery level. The model picks &lt;code&gt;k8s.get_rollout_status&lt;/code&gt;, and only then is its full schema revealed.; &lt;strong&gt;The model sees:&lt;/strong&gt; Six one-liners (about 200 tokens), then one schema (about 400).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Validation&lt;/strong&gt;: &lt;strong&gt;What happens:&lt;/strong&gt; The required parameter &lt;code&gt;namespace&lt;/code&gt; is missing. The validator returns an error, the model supplies &lt;code&gt;payments&lt;/code&gt;, and the second attempt passes.; &lt;strong&gt;The model sees:&lt;/strong&gt; The error string.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Execution&lt;/strong&gt;: &lt;strong&gt;What happens:&lt;/strong&gt; 200 KB of pod events is capped to a 1.2 KB observation, with a handle to the full output kept outside the context.; &lt;strong&gt;The model sees:&lt;/strong&gt; 1.2 KB.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Write subgoal&lt;/strong&gt;: &lt;strong&gt;What happens:&lt;/strong&gt; Posting to Slack is a write. The risk policy shows the user the draft message and waits for a confirmation before &lt;code&gt;slack.post_message&lt;/code&gt; runs.; &lt;strong&gt;The model sees:&lt;/strong&gt; The draft message.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Count the tokens with the 400-token schema assumption from earlier. Showing the whole catalog would cost about 480k tokens on every turn, more than most context windows hold. This run used three menus of about 200 tokens and three schemas of about 400, roughly 1,800 tokens, a cut of more than 99%. Those are assumptions multiplied out, not a measurement.&lt;/p&gt;

&lt;p&gt;The more useful observation is where the LLM acted. It made two decisions that choose tools: splitting the request, and choosing from a menu of six. It also supplied one missing parameter and drafted the Slack message, but neither of those picks a tool. Every other step was an index, a rule or a learned ranker. That is the point of the control plane. If this run had gone wrong, it would have gone wrong in one identifiable place. Had &lt;code&gt;k8s.get_rollout_status&lt;/code&gt; never been retrieved, that is a retrieval miss. Had it been retrieved and cut, that is the reranker. Had it been chosen and called badly, the validator catches it before anything runs.&lt;/p&gt;

&lt;p&gt;11 / Judgment&lt;/p&gt;

&lt;h2&gt;
  
  
  What to build at which scale
&lt;/h2&gt;

&lt;p&gt;Here is the whole flow in one place, in three stages. It starts where the user does, with the request. The tag on each step says whether it is rules, a learned model, or the LLM. Two of the eight steps are the LLM, and every loop is decided by code.&lt;/p&gt;
Fig. 1. Discovery is a stage separate from execution, and the reasoning model never sees the full catalog. Tags: LLM (green), LEARNED (accent), RULES and INDEX (blue). Red arrows are failure paths, the accent arrow is the loop for a new subgoal.



&lt;p&gt;You do not need all of it. The thresholds below are my judgment from building these systems, not measured boundaries. In my simulation the full-schema baseline crossed the 70% context-utilization line at about 50 tools, which is why the first row is short. That figure depends on the context window and on how much the system prompt and conversation already occupy, so rerun it with your own numbers.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Up to ~50&lt;/strong&gt;: &lt;strong&gt;What breaks first:&lt;/strong&gt; Nothing yet.; &lt;strong&gt;Add:&lt;/strong&gt; Nothing. Send them all.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;50 to 200&lt;/strong&gt;: &lt;strong&gt;What breaks first:&lt;/strong&gt; Cost and mid-list misses.; &lt;strong&gt;Add:&lt;/strong&gt; The baseline pipeline: one embedding, top-k, definitions, validation.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;200 to 1,000&lt;/strong&gt;: &lt;strong&gt;What breaks first:&lt;/strong&gt; Near-duplicates, multi-step tasks, unmeasurable failures.; &lt;strong&gt;Add:&lt;/strong&gt; Hybrid retrieval, hard filters, progressive disclosure, per-stage recall.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;1,000+&lt;/strong&gt;: &lt;strong&gt;What breaks first:&lt;/strong&gt; Shortlists crowded by one family, catalog rot, stage latency.; &lt;strong&gt;Add:&lt;/strong&gt; Domain routing, a trained reranker, per-subgoal discovery, catalog ownership and linting.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Where this breaks in production&lt;/p&gt;

&lt;p&gt;Each stage adds latency, and I have not measured a full stack end to end, so budget it before you commit. Indexes go stale when a server ships a new schema, which is why &lt;code&gt;schema_version&lt;/code&gt; exists. And tool descriptions are attacker-controlled text. In my simulation, gating cut the success rate of 50 poisoned descriptions from 38% to 6%. That was a side effect of never showing irrelevant tools, not a defense, and it needs a real one. The attacks to design for are tool poisoning (instructions hidden in a description), tool shadowing (a third-party tool describing itself to win selection over a trusted one) and a description that changes after it was approved. In the control plane that means hashing each description at approval and re-reviewing on change, treating third-party descriptions as untrusted input in catalog linting, and keeping write tools behind the risk policy whatever any description says. Log per-stage latency (p50 and p95) next to per-stage recall.&lt;/p&gt;

&lt;p&gt;12 / Closer&lt;/p&gt;

&lt;h2&gt;
  
  
  The short version
&lt;/h2&gt;

&lt;p&gt;For classification, the rule is to stop asking the LLM to remember the taxonomy and give it the taxonomy when it needs it. I would keep that and add its counterpart for tools.&lt;/p&gt;

&lt;p&gt;Do not ask the LLM to be the only thing standing between a request and a side effect. Give it a small, filtered, well-described menu. Run the rules around it in code. Measure every stage, so that when the agent is wrong you know which part to fix.&lt;/p&gt;

&lt;p&gt;My design rule fits in one line. The LLM reasons over the actions that are viable right now and does not carry the platform's whole capability surface in every prompt. Retrieve broadly, filter deterministically, rerank intelligently, and reveal full tool details only when the model needs them.&lt;/p&gt;

&lt;p&gt;At 100 tools, the model does the work. At 1,000, the control plane does most of it, and the model makes the last call.&lt;/p&gt;




&lt;p&gt;&lt;a href="https://tech.anujsadani.in/tool-selection-at-scale/" rel="noopener noreferrer"&gt;Read the original on Anuj Sadani | Tech&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>data</category>
      <category>architecture</category>
      <category>programming</category>
    </item>
    <item>
      <title>Decision Models: From a Typed Prediction to an Accountable Action</title>
      <dc:creator>Anuj Sadani</dc:creator>
      <pubDate>Sat, 31 Oct 2026 23:30:00 +0000</pubDate>
      <link>https://dev.to/anuj-sadani/decision-models-from-a-typed-prediction-to-an-accountable-action-20me</link>
      <guid>https://dev.to/anuj-sadani/decision-models-from-a-typed-prediction-to-an-accountable-action-20me</guid>
      <description>&lt;p&gt;A model can return probabilities and a recommendation. A product still needs to decide what action is allowed, what evidence is sufficient, and who remains accountable. This 13-chapter guide introduces Jev and other decision models through that boundary between prediction and action.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;A chapter map:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/decision-models-guide/#chapter-1.-the-ticket-that-isnt-one-question" rel="noopener noreferrer"&gt;Chapters 1–3: tickets and the decision-model idea&lt;/a&gt;: work through a ten-ticket example, then define what a decision model is and is not.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/decision-models-guide/#chapter-4.-reading-the-numbers" rel="noopener noreferrer"&gt;Chapters 4–6: read the numbers and inspect a measured case&lt;/a&gt;: interpret an output, follow one real call, and see what the reported experiments establish.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/decision-models-guide/#chapter-7.-the-cascade-that-made-things-worse" rel="noopener noreferrer"&gt;Chapters 7–9: alternatives and limits&lt;/a&gt;: compare the cascade that regressed and examine what changes around the model.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/decision-models-guide/#chapter-10.-between-prediction-and-action" rel="noopener noreferrer"&gt;Chapters 10–13: policy, accountability, and what to build&lt;/a&gt;: place policy between prediction and action, then consider extensions and a first implementation.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/decision-models-guide/#appendix-a.-what-was-and-was-not-validated" rel="noopener noreferrer"&gt;Appendix A: what was and was not validated&lt;/a&gt;: separate the reproduced work from reported measurements and unvalidated claims.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The reading edition includes chapter mind maps and synthetic narration. The guide takes about 114 minutes to listen to. Its offline workflow uses synthetic tickets, not measured model performance. &lt;a href="https://tech.anujsadani.in/decision-models-guide/" rel="noopener noreferrer"&gt;Read or listen to the complete guide&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>machinelearning</category>
      <category>architecture</category>
      <category>tutorial</category>
    </item>
    <item>
      <title>When Software Hires Software: The Executive Playbook for the Agent-to-Agent Economy</title>
      <dc:creator>Anuj Sadani</dc:creator>
      <pubDate>Sat, 31 Oct 2026 19:30:00 +0000</pubDate>
      <link>https://dev.to/anuj-sadani/when-software-hires-software-the-executive-playbook-for-the-agent-to-agent-economy-5176</link>
      <guid>https://dev.to/anuj-sadani/when-software-hires-software-the-executive-playbook-for-the-agent-to-agent-economy-5176</guid>
      <description>&lt;p&gt;An executive playbook on the agent-to-agent economy: autonomous procurement, algorithmic corporate wallets, machine-negotiated mortgage lending, and healthcare claims that argue with themselves. Five chapters -- one cross-cutting thesis, then four domains -- locate the Mandate Boundary in each: where software may decide alone, where a human still has to sign, and what to build by 2027 versus only position for by 2030. Free to read here and to listen to (102 minutes, synthetic voice); the typeset PDF is on Ko-fi.&lt;/p&gt;

&lt;p&gt;This is an adapted edition of the original article. It preserves the section argument and evidence; use the source link for the published version and any interactive material.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://tech.anujsadani.in/when-software-hires-software/" rel="noopener noreferrer"&gt;Read the original article&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  When Software Hires Software
&lt;/h2&gt;

&lt;p&gt;Every roadmap answers one question, whether it says so or not: who acts. For thirty years the answer was a person, and the software just carried the paperwork. That is the assumption breaking right now, and it is breaking on the buying side first.&lt;/p&gt;

&lt;p&gt;Look at what actually shipped between April 2025 and April 2026, not what got keynoted. Mastercard built Agent Pay, tokens that bind a card to a specific agent, a specific merchant, and a specific consent policy &lt;sup&gt;1&lt;/sup&gt;. Visa built the Trusted Agent Protocol, a verified identity for an agent plus a signed consent record from the issuer &lt;sup&gt;1&lt;/sup&gt;. Google built AP2, which turns a purchase into three cryptographically signed contracts — an Intent Mandate for what you asked for, a Cart Mandate for what got assembled, a Payment Mandate for what gets charged &lt;sup&gt;2&lt;/sup&gt;. None of that is a chatbot feature. It is financial infrastructure for a counterparty that isn't a person, built by the companies whose entire business is knowing who is allowed to move money.&lt;/p&gt;

&lt;p&gt;That is the fact this book is built around: the frontier of AI product work has quietly moved from &lt;em&gt;can software understand something&lt;/em&gt; to &lt;em&gt;can software be trusted to act on it, alone, with money and consequences attached&lt;/em&gt;. Call it what it is. &lt;strong&gt;Software is starting to hire software&lt;/strong&gt; — to shop from it, negotiate with it, pay it, and increasingly, to answer to it when something goes wrong. Every domain chapter that follows is one industry's version of this same shift. This chapter is the shift itself, and what a leader does about it before the shape of 2030 is remotely settled.&lt;/p&gt;

&lt;h3&gt;
  
  
  The line that actually matters
&lt;/h3&gt;

&lt;p&gt;Forget the debate about whether AI is "agentic" — that word is mostly marketing residue at this point (Gartner counted about 130 vendors with real agentic capability out of thousands claiming it, and predicts over 40% of agentic AI projects will be scrapped by the end of 2027 for cost, unclear value, or missing risk controls &lt;sup&gt;3&lt;/sup&gt;). The debate worth having is narrower and more useful: &lt;strong&gt;for a given decision, where does authority to act without a human in the loop currently sit, and how far are you willing to move it?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Call that the &lt;strong&gt;Mandate Boundary&lt;/strong&gt; — the line separating decisions a system may execute unaccompanied from decisions that still require a human counter-signature. Every mandate protocol built in the last eighteen months is, underneath the branding, a machine for drawing this line precisely and making the drawing enforceable: AP2's Intent and Cart Mandates exist specifically to separate "what I authorized" from "what the agent decided" so the boundary survives a dispute &lt;sup&gt;2&lt;/sup&gt;. Visa's Trusted Agent Protocol exists to prove an agent's identity to a merchant &lt;em&gt;before&lt;/em&gt; the boundary gets tested &lt;sup&gt;1&lt;/sup&gt;.&lt;/p&gt;

&lt;p&gt;Your roadmap is not really a list of features. It is a series of bets about where you move the Mandate Boundary, for which decisions, on what evidence, and how fast you can move it back if the evidence turns bad. Once you see the roadmap this way, "what should we build for 2027 versus 2030" stops being a features question and becomes a governance question — which is a much more answerable one, because governance questions have precedent even when the technology doesn't.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;The protocol for letting software transact is now mostly solved. What remains unsolved — and what every domain in this book is really fighting over — is who is accountable when the transaction is wrong.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h3&gt;
  
  
  The wave that keeps breaking roadmaps
&lt;/h3&gt;

&lt;p&gt;Here is the uncomfortable part for anyone used to eighteen-month planning cycles: the frontier moved and reversed inside a single one. OpenAI launched in-chat Instant Checkout on the Agentic Commerce Protocol in September 2025, built with Stripe, aimed squarely at the "customer never leaves the chat" vision every 2025 keynote gestured at. By March 2026 — six months later — OpenAI retired it. Reported adoption topped out somewhere between "about a dozen" and "closer to thirty" live merchants, and Walmart found checkout inside ChatGPT converting roughly three times worse than a click-through to its own site &lt;sup&gt;4&lt;/sup&gt;. The pattern that survived wasn't "AI closes the sale." It was "AI helps you discover, and you still buy where you already trust the receipt."&lt;/p&gt;

&lt;p&gt;That is not a story about a bad product. It is a story about the actual shape of this wave: pieces of it move immediately and permanently — protocol governance for MCP and A2A both landed with the Linux Foundation inside about a year of each launching &lt;sup&gt;5&lt;/sup&gt; &lt;sup&gt;6&lt;/sup&gt; — while other pieces that looked equally inevitable in a demo (consumers letting an agent complete the whole purchase) stall on trust that hasn't been earned yet, and might not be earned on anyone's preferred timeline. Both things are true about the same eighteen months. A roadmap built on either "none of this is real yet" or "all of this is inevitable by next quarter" will be wrong regardless of which one you pick, because the wave doesn't move as one wave. It moves as several waves at different speeds, and the speed each one travels at is set by how expensive it is to be wrong, not by how good the technology has gotten.&lt;/p&gt;

&lt;p&gt;That single fact is the whole argument for building flexibility into a 2030 plan instead of a fixed destination. Nobody who published a roadmap in early 2025 predicted that the most hyped consumer agentic-commerce feature of the year would be dead by its own six-month anniversary while the identity and payment rails underneath it kept right on standardizing. The rails were boring, so they shipped. The consumer-facing bet was exciting, so it got canceled. Build for the second pattern to repeat, because it will keep repeating: the plumbing keeps compounding quietly, the exciting front-end bets keep dying in public.&lt;/p&gt;

&lt;h3&gt;
  
  
  Where the moat still is, once the protocol is free
&lt;/h3&gt;

&lt;p&gt;Here is the part that should reorganize how you think about differentiation. MCP and A2A are not proprietary technology anymore — they are Linux Foundation projects, governed the way Kubernetes is governed, built specifically so no single vendor can own them &lt;sup&gt;5&lt;/sup&gt; &lt;sup&gt;6&lt;/sup&gt;. If your 2027 plan is "we will expose an agent interface," you have described table stakes, not a strategy. Every competitor with an engineering team can expose the same interface by next quarter, and several already have.&lt;/p&gt;

&lt;p&gt;The moat was never going to be in the interface. It has to be in what stands behind it when an agent — yours or someone else's — asks a hard question and needs an answer a court, a regulator, or an auditor will accept later. Three things sit behind a durable answer, and none of them are the protocol:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Verified facts with provenance.&lt;/strong&gt; Not "the model's best guess," but a value with an evidence trail — which document, which page, which rule version, computed when, by what — attached to it permanently. This is expensive to build and boring to describe, which is exactly why it's defensible.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Warranted outcomes.&lt;/strong&gt; A price attached to being right, not just to answering. The difference between a vendor who tells you a decision and a vendor who stakes something on that decision being correct is the difference between a feature and a business.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Accountable judgment.&lt;/strong&gt; A specific party — a licensed underwriter, a named institution, a warranty-backed system — who absorbs the consequence when the decision is wrong. Software can execute a mandate. It cannot yet be the thing a regulator holds responsible when the mandate was a mistake, and nothing in any protocol shipped so far changes that.&lt;/p&gt;

&lt;p&gt;Every domain chapter that follows names the version of these three things that applies to one industry, because "verified facts, warranted outcomes, accountable judgment" is not a slogan — it is the actual checklist for telling a real product from a demo in this wave. If your 2027 plan doesn't strengthen at least one of the three, it is decoration on top of someone else's infrastructure, and it will feel that way to your customers within a year.&lt;/p&gt;

&lt;h3&gt;
  
  
  The menu: how the winners actually get paid
&lt;/h3&gt;

&lt;p&gt;Naming the moat is only half the roadmap. The other half is a pricing decision every one of these initiatives eventually forces, and getting it wrong is a common way to build the right infrastructure and still lose the business. Four patterns recur across every domain in this book, and the durable products almost always combine two of them rather than picking just one.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Metered access&lt;/strong&gt; prices the thing closest to a commodity — a single tool call, a single check, a single page processed — by the unit, the way a document-extraction API bills per page or per credit. This is the right model for atomic, deterministic, low-stakes operations: cheap enough that volume is the business, and simple enough that a buyer can predict their bill before they commit to it. It is also the model most exposed to commoditization, because a metered price is trivially comparable to a competitor's metered price. Use it for the parts of your product that really are commodities, and stop pretending otherwise.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Outcome-warranted pricing&lt;/strong&gt; charges for being right, not for answering — a fee per approved authorization, per closed loan, per resolved case, with an actual guarantee or fee-at-risk attached if the outcome turns out wrong. This is the model that only becomes available once you've built the "warranted outcomes" leg of the moat described above, and it is worth far more per transaction than metered access, because the buyer is no longer purchasing an answer. They're purchasing the transfer of a risk they used to hold themselves.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Tiered subscription with role-gated access&lt;/strong&gt; sells a seat, not a call — a monthly or annual fee that unlocks a specific bundle of tools, data freshness, and volume, with the tier determining not just price but rate limits and which tools are even visible to that user. This is the commercial expression of the same progressive-disclosure logic that keeps a well-designed agent's tool count manageable: a junior role's seat exposes a handful of tools at a modest rate limit; a senior role's seat exposes more tools, deeper data, and a higher ceiling; an enterprise tier adds custom overlays and effectively uncapped throughput. Price the tier to the value of what a role can see and do, not to how much compute it happens to consume.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;One-time integration plus recurring usage&lt;/strong&gt; front-loads a fee for the unglamorous work of connecting to somebody else's system of record — a lender's loan-origination system, a payer's claims platform, a hospital's EHR — and then layers a recurring fee, usually one of the first three patterns, on top once the connection is live. Skip this and you'll underprice the hardest, least repeatable part of the sale: getting into the account at all.&lt;/p&gt;

&lt;p&gt;Every domain chapter that follows names its own version of this menu, because the question "what do we charge for" deserves the same concrete answer as "what do we build," and a roadmap that only answers the second question will spend 2027 giving away, metered by the call, the exact thing it should have been warrantying by the outcome.&lt;/p&gt;

&lt;h3&gt;
  
  
  What to build now, and what to only prepare for
&lt;/h3&gt;

&lt;p&gt;Treat 2027 and 2030 as two different kinds of bet, not two points on the same line.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2027 is a build horizon.&lt;/strong&gt; The pattern already visible across every domain in this book is &lt;em&gt;software recommends, a human authorizes&lt;/em&gt; — a mandate gets assembled with evidence, a specific accountable person signs it, and the system's job is to make that signature fast, well-informed, and cheap to defend later. This is buildable today with technology that already exists: evidence-bundled decisions, role-scoped tool access so a human reviewer sees exactly what they need and no more, and outcome-priced pilots that put your own money where your accuracy claims are. Build this now. It compounds — every decision it makes correctly becomes training data, a defensible track record, and one more data point in the warranty argument you'll need to expand the Mandate Boundary later.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2030 is a positioning horizon, not a build target.&lt;/strong&gt; The pattern to prepare for, without betting the company on its exact timing, is &lt;em&gt;software authorizes within a policy a human already set&lt;/em&gt; — continuous, standing mandates re-evaluated as facts change, rather than one-time approvals re-litigated from scratch every time. Mortgage lending's "always pre-approved" is a plausible version of this. Auto-locking a rate and closing a loan without a human touching it is not plausible on any near-term timeline, for the same reason full agentic checkout stalled: the cost of being wrong is high and the trust to be delegated hasn't been built yet. The organizations that win this horizon won't be the ones who guessed the date correctly. They'll be the ones who spent 2027–2029 accumulating the evidence trail, the warranty capacity, and the regulatory relationships that let them move the Mandate Boundary the moment the trust actually exists — while their competitors are still arguing about whether it exists yet.&lt;/p&gt;

&lt;p&gt;Practically, that means every initiative on your roadmap should answer two questions before it gets funded. First: which side of the Mandate Boundary does this sit on today, and what evidence would move it? Second: if the specific bet inside this initiative turns out to be six-month-hype rather than durable infrastructure — the Instant Checkout pattern rather than the AP2 pattern — how much of what we built survives the pivot? Facts infrastructure, evidence schemas, and accountability relationships survive almost any pivot. Bets on a specific vendor protocol or a specific consumer behavior do not. Weight your 2027 spending accordingly, and you get the flexibility to pivot without having wasted the year finding out you needed it.&lt;/p&gt;

&lt;p&gt;The rest of this book walks that same S-curve through four places where the money is already real, not speculative: procurement, where agents are starting to shop from other agents; banking, where the wallet itself is becoming an algorithm; mortgage lending, where a decision infrastructure is quietly replacing what used to be a document pipeline; and healthcare, where a claim is starting to argue with itself before a human ever sees it. Each chapter names the specific Mandate Boundary in that domain, where it sits today, and what it would take to move it. Start wherever your business already touches one of these four — the mechanics transfer faster than you'd expect, because underneath the industry-specific vocabulary, it's the same question every time: who is allowed to act, on what evidence, and who answers for it when they're wrong.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Procurement Agent
&lt;/h2&gt;

&lt;p&gt;Somewhere in your target account list right now, a piece of software is deciding whether your company is worth a human's time — and it is making that call faster than a person could finish loading your homepage. It does not watch your demo video. It does not read your case studies for tone. It checks whether your claims can be verified without picking up a phone, and if they can't, it drops you from the shortlist before anyone on your sales team learns you were ever being considered.&lt;/p&gt;

&lt;p&gt;Call that check &lt;strong&gt;the Vendor Handshake&lt;/strong&gt;: the exchange of signed, checkable claims — specs, security attestations, pricing, compliance status — that a buying agent runs against a vendor before it will do anything a human buyer used to do first, like shortlisting, comparing, or requesting a quote. Every domain in this book has its own version of the moment one piece of software decides whether to trust another before a person gets pulled in. In procurement, that moment is already happening, in production, at scale — just not in the form most vendor marketing sites were built to survive.&lt;/p&gt;

&lt;p&gt;This chapter answers a narrower question than "will AI change B2B sales." It answers: when the buyer, the comparison-shopper, and increasingly the person drafting the RFP response are all software, what should a vendor build in the next twelve months, and what should it merely position for?&lt;/p&gt;

&lt;h3&gt;
  
  
  The Money Is Real. The Checkout Isn't. Not Yet.
&lt;/h3&gt;

&lt;p&gt;Start with what's actually been funded, because procurement is one of the few corners of this wave where the capital allocation tells you more than the keynotes do.&lt;/p&gt;

&lt;p&gt;Profound, a company that exists to make brands legible to AI answer engines rather than search engines, raised a \$96 million Series C at a \$1 billion valuation in February 2026, becoming the category's first unicorn. Seven months later, in September 2026, it raised a \$180 million Series D at \$1.8 billion, co-led by Sequoia and Kleiner Perkins, with revenue up roughly 3x over that same six-month stretch and more than 1,000 enterprise customers on the books, including Comcast, Estée Lauder, and Walmart &lt;sup&gt;7&lt;/sup&gt;. That is not hype-cycle money. That is a market that has already decided generative-engine optimization — GEO, the practice of making your company's facts legible to the models doing the summarizing — is worth real budget, twice, in under a year.&lt;/p&gt;

&lt;p&gt;Incumbents are reading the same signal. Adobe agreed to buy Semrush, the SEO analytics platform, for \$1.9 billion in November 2025, explicitly to fold GEO capability into Adobe Experience Cloud before someone else owns the category &lt;sup&gt;8&lt;/sup&gt;. When a company with Adobe's balance sheet pays nearly \$2 billion for a bolt-on rather than building it, that's a company telling you the build-versus-buy math already favors moving fast.&lt;/p&gt;

&lt;p&gt;And the demand side backs this up with actual commerce data, not just vendor claims. On Shopify's Q1 2026 earnings call, President Harley Finkelstein reported that AI-referred traffic to Shopify stores was up roughly 8x year over year, with orders from AI-powered search up nearly 13x — new buyers arriving through AI channels at close to twice the rate of other channels &lt;sup&gt;9&lt;/sup&gt;. By the Q2 2026 call in August, both AI traffic and AI-attributed orders had tripled year over year again, and Finkelstein noted that AI search running on Shopify's own structured catalog data converted at roughly twice the rate of AI search relying on scraped product pages &lt;sup&gt;9&lt;/sup&gt;. Read that last detail twice — it's the whole chapter in one sentence. The advantage wasn't "we're in more AI answers." It was "our facts are structured well enough for the answer engine to trust them without scraping."&lt;/p&gt;

&lt;p&gt;Meanwhile, distribution is fragmenting into open warfare. Amazon expanded its robots.txt blocklist to keep OpenAI's ChatGPT-User and OAI-SearchBot crawlers off its product pages in late 2025, protecting an advertising business that depends on shoppers browsing Amazon directly rather than through a chat window that skips the ads &lt;sup&gt;10&lt;/sup&gt;. Then Amazon sued Perplexity, accusing its Comet browser of disguising its shopping agent as a regular human session to keep scraping after being blocked — and won a preliminary injunction in March 2026, before the Ninth Circuit vacated it in August on the technical grounds that it was the &lt;em&gt;user&lt;/em&gt;, not Perplexity, accessing Amazon's servers under federal computer-access law &lt;sup&gt;10&lt;/sup&gt;. Nobody in that fight was arguing about whether agents will shop. They were arguing about who gets paid when they do.&lt;/p&gt;

&lt;p&gt;Set all of that against the thing that was supposed to be the headline feature: OpenAI's Instant Checkout, launched in September 2025 to let a purchase complete entirely inside a chat window, retired roughly six months later after adoption topped out somewhere between a dozen and thirty live merchants and Walmart found in-chat checkout converting at about a third the rate of a click-through to its own site &lt;sup&gt;4&lt;/sup&gt;. Consumers, it turns out, still want to land on a page they recognize before money moves. That is the same trust gap this book's opening chapter identifies across every domain — protocol infrastructure compounds quietly while the consumer-facing bet dies in public.&lt;/p&gt;

&lt;p&gt;So here is the honest 2027 map for procurement, and it is not the map most GTM decks are drawing. The wedge isn't consumer checkout autonomy — that stalled, and there's no evidence it's coming back on anyone's preferred timeline. The wedge is &lt;strong&gt;being legible and verifiable to the agents doing the comparing&lt;/strong&gt;, whether those agents are consumer-facing answer engines or the buyer-side procurement agents inside your prospects' companies. GEO money is real because it answers a question every vendor already has: does an AI system, asked about my category, say my name accurately, or does it hallucinate my competitor's pricing into my slot. Verified-claims infrastructure is the underbuilt half of that same question: once you're mentioned, can a buying agent &lt;em&gt;check&lt;/em&gt; what was said about you without a phone call — and will it trust the answer enough to shortlist you.&lt;/p&gt;

&lt;h3&gt;
  
  
  What a Buying Agent Actually Needs Before It Trusts You
&lt;/h3&gt;

&lt;p&gt;Here is the mechanic underneath "software hires software," made concrete for procurement.&lt;/p&gt;

&lt;p&gt;A human buyer on a vendor shortlist tolerates ambiguity because a human can ask a follow-up question, read tone, and build trust across a sales cycle. A buying agent doesn't get that luxury, and — more importantly — it isn't supposed to. Its whole value proposition to the human who deployed it is that it can shortlist, compare, and draft a recommendation &lt;em&gt;without&lt;/em&gt; a six-week discovery process. That only works if the facts it's comparing are structured well enough to be checked mechanically rather than taken on faith. This is the Vendor Handshake in practice, and it has a fairly specific shopping list:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Signed specifications&lt;/strong&gt;, not marketing copy — a machine-readable data sheet with a cryptographic signature tying the claim to the entity that made it and the date it was made, the B2B equivalent of the structured catalog data Shopify's own numbers show converting twice as well as scraped pages &lt;sup&gt;9&lt;/sup&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Security and compliance attestations&lt;/strong&gt; exposed at a well-known, machine-queryable endpoint rather than buried in a PDF a salesperson emails after a discovery call — SOC 2 status, data residency, uptime history, the answers that used to live inside a security questionnaire a human filled out under deadline pressure.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Pricing and contract terms a buying agent can evaluate against its mandate&lt;/strong&gt; without a call — not necessarily public list pricing, but structured enough that an agent operating inside a spend policy can determine fit or disqualification on its own, the same way Google's AP2 turns a consumer purchase into signed Intent, Cart, and Payment mandates precisely so a dispute can be resolved by checking a record instead of re-litigating a conversation &lt;sup&gt;2&lt;/sup&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;A verification path that doesn't depend on the vendor's own say-so&lt;/strong&gt; — third-party attestation, audit trail, or a protocol-level signature the buying agent's infrastructure can validate independently, the same trust-before-transaction logic Visa built into the Trusted Agent Protocol so a merchant can confirm an agent's identity before anything moves &lt;sup&gt;1&lt;/sup&gt;.&lt;/p&gt;

&lt;p&gt;None of this requires a proprietary interface. MCP and A2A are Linux Foundation projects now, governed the way Kubernetes is governed, specifically so no single vendor can lock up the pipes &lt;sup&gt;5&lt;/sup&gt; &lt;sup&gt;6&lt;/sup&gt;. A vendor exposing an agent-readable endpoint next quarter is describing table stakes, exactly as the opening chapter of this book warns about any "we'll expose an agent interface" roadmap line. The differentiation was never going to be in having a feed. It's in whether anything you put on the feed can survive being checked.&lt;/p&gt;

&lt;p&gt;That's the domain-specific version of the Mandate Boundary this book keeps returning to. In procurement, the Mandate Boundary sits at the shortlist and the RFP response: a buying agent may gather, compare, and rank vendor claims unaccompanied, but the moment those claims translate into a signed commitment — a contract, a purchase order, a vendor relationship a company will depend on for years — a human still has to countersign, because nothing about a well-formatted claims feed changes who absorbs the consequence if the claims were wrong.&lt;/p&gt;
Procurement: what to build in 2027 vs. what to position for in 2030.



&lt;h3&gt;
  
  
  The Moat, Applied to the Loading Dock
&lt;/h3&gt;

&lt;p&gt;MCP and A2A being free protocols means the interface was never the business. The three things that were always going to be the business, applied to procurement specifically:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Verified facts with provenance.&lt;/strong&gt; Not a spec sheet a vendor wrote about itself, but a claim with an evidence trail attached permanently — which security audit, which pricing tier, which compliance certification, issued by whom, verifiable independently of the vendor repeating it. This is what a verified-claims feed actually is, once you strip the GEO branding off it: provenance infrastructure for B2B facts, built the same way facts infrastructure gets built in every other chapter of this book — boring, expensive, and exactly why it's defensible once built.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Warranted outcomes.&lt;/strong&gt; The uncomfortable question a claims feed has to answer eventually: what does the vendor stake on the claim being accurate? A spec sheet with no consequence attached is marketing. A spec sheet a vendor is willing to warrant — refund the deal, cover a security incident, eat a penalty if the attestation was stale or wrong — is infrastructure a buying agent's operator can actually rely on, because the agent's mandate can now price the risk of being wrong instead of just trusting the claim. This is the same distinction the opening chapter draws between a vendor who tells you a decision and one who has priced being wrong into the product.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Accountable judgment.&lt;/strong&gt; When a buying agent shortlists a vendor off a bad claim and the resulting purchase goes badly, who answers for it? Not the model — the model can't be fired, sued, or held to a contract. The answer has to be a specific, named party: the vendor that signed the attestation, the procurement lead who countersigned the mandate at the Mandate Boundary, or — increasingly, this is the actual product opportunity — a third-party verification provider whose entire business is standing behind the claim so that both sides of the transaction can point to a name when something breaks. That is a genuinely new role in the B2B stack, and right now almost nobody occupies it cleanly.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;A shortlist an agent can generate in ten seconds is worth nothing to a buyer who still has to spend six weeks finding out whether any of it was true.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;That's the gap the current wave of GEO tooling doesn't close, because tracking whether you're mentioned accurately is a monitoring product, not a trust product. The vendors who win the next stretch of this build will be the ones who stopped asking "are we visible to the agent" and started asking "can the agent verify us without calling us" — and then went and built the warranty behind the answer.&lt;/p&gt;

&lt;h3&gt;
  
  
  What to build, and how to charge for it
&lt;/h3&gt;

&lt;p&gt;Run procurement's version of the moat through Chapter 0's menu and four concrete products fall out, not just a direction.&lt;/p&gt;

&lt;p&gt;A &lt;strong&gt;verified-claims feed&lt;/strong&gt; — the signed specs, attestations, and pricing this chapter has already described — is naturally a tiered subscription: a base tier covering a fixed claim set and a modest number of agent verification pulls per month, a higher tier unlocking more claim categories (security, compliance, custom pricing) and a higher rate limit on how many times a buying agent can query it, because the vendor's real cost scales with verification volume, not with having claims at all. Layer a metered verification fee on top for agents outside a buyer's existing subscription — a pay-per-check option for the long tail of buying agents a vendor has no standing relationship with.&lt;/p&gt;

&lt;p&gt;An &lt;strong&gt;agent-readiness audit&lt;/strong&gt; — bringing a vendor's product feed and well-known endpoints up to a standard a buying agent can actually parse — is a clean one-time integration fee, priced like a security audit, with an optional ongoing monitoring subscription for vendors who want to be re-certified as protocols evolve.&lt;/p&gt;

&lt;p&gt;The &lt;strong&gt;third-party verification and warranty role&lt;/strong&gt; this chapter identifies as a genuinely open position in the stack is where outcome-warranted pricing belongs: a per-claim or per-deal verification fee that scales with the value of the transaction it's underwriting, with a real payout if a claim the verifier certified turns out to be false. That is a materially different — and larger — business than selling access to a claims database, for the same reason an insurance premium is worth more than a subscription to weather data.&lt;/p&gt;

&lt;p&gt;And the 2030 buyer-agent RFP responder this chapter already prices per qualified opportunity is outcome-warranted pricing taken to its logical endpoint: a vendor pays only when its agent's answers actually advanced a real deal, which is the only pricing model a buyer will accept for a system negotiating on their behalf without supervision.&lt;/p&gt;

&lt;h3&gt;
  
  
  2030: When the RFP Answers Itself
&lt;/h3&gt;

&lt;p&gt;Here the evidence runs out and the extrapolation starts, so treat what follows as informed positioning, not a forecast with sources behind it.&lt;/p&gt;

&lt;p&gt;The plausible next stage, if the Vendor Handshake becomes standard infrastructure the way TLS became standard infrastructure for web traffic, is a buyer-side procurement agent that doesn't just shortlist vendors — it drafts and negotiates the RFP response itself, querying vendor agents directly over A2A for warranted answers to security questionnaires, SLA terms, and pricing tiers, priced by the vendor per qualified opportunity rather than per lead &lt;sup&gt;6&lt;/sup&gt;. That's a genuinely different shape of GTM motion: instead of a vendor's sales team chasing a signal that a company is "in-market," a vendor's agent is fielding structured, machine-generated RFP queries around the clock and only escalating to a human seller when a deal clears a threshold worth a human's time.&lt;/p&gt;

&lt;p&gt;The harder version of this — a standing procurement mandate, re-evaluated continuously as a vendor's certifications, pricing, or security posture changes, the way this book's mortgage chapter describes an "always pre-approved" borrower — is a plausible 2030 pattern for repeat, low-variance B2B purchases: consumables, commodity SaaS seats, routine service renewals. It is not plausible yet for anything with real switching cost or real risk, for the same reason full agentic checkout stalled in retail: the price of being wrong is still higher than the trust that's been built to be delegated. This is my inference from where the infrastructure is heading, not a sourced benchmark — nobody has shipped a standing B2B procurement mandate at scale yet, and the regulatory and contractual questions underneath one (who's bound, under what authority, revocable how fast) are further from settled than the payment rails are.&lt;/p&gt;

&lt;p&gt;What's worth building toward now, regardless of exactly when 2030 arrives, is the accountability relationship a standing mandate would require: a named party who can be pointed to when a vendor's claim turns out to be stale or a purchase goes wrong. That relationship is worth building years before the protocol matures, because it's the part that doesn't get thrown away if the specific 2030 pattern turns out wrong and something else wins instead.&lt;/p&gt;

&lt;p&gt;Procurement was never really a question of who gets to shop. It was always a question of who gets to spend the company's money and who answers for it after. The next place that question shows up, in a more literal form, is the wallet itself — which is exactly where this book goes next.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Algorithmic Wallet
&lt;/h2&gt;

&lt;p&gt;A corporate card used to need a person to swipe it. By 2027 it mostly needs a policy instead — and the fight over who writes that policy, not who processes the swipe, is where the actual money in this chapter is.&lt;/p&gt;

&lt;p&gt;Call the thing the policy governs the &lt;strong&gt;Policy Wallet&lt;/strong&gt;: not an account a human authorizes one purchase at a time, but a standing set of conditions — vendor, category, price band, cadence, counterparty — that a piece of software checks against every attempted payment and either honors or blocks, with nobody awake to watch it happen. The Policy Wallet is the Mandate Boundary from Chapter 1 poured into a specific, cash-shaped mold. Every mandate protocol shipped in the last eighteen months — Mastercard's Agentic Tokens, Visa's Trusted Agent Protocol, Google's AP2 — is a machine for drawing that boundary around money specifically: what an agent may spend, on what, before a human has to look at it again. The protocols answer &lt;em&gt;how&lt;/em&gt; a payment gets authorized. They do not answer &lt;em&gt;what the policy should say&lt;/em&gt;, who is on the hook when the policy is wrong, or how a bank proves any of it to a regulator eighteen months later. That gap is the whole chapter.&lt;/p&gt;

&lt;h3&gt;
  
  
  The rails are done. Compete somewhere else.
&lt;/h3&gt;

&lt;p&gt;Start with what's no longer in question, because treating it as an open competitive space is the fastest way to waste a 2027 budget.&lt;/p&gt;

&lt;p&gt;Mastercard announced Agent Pay in April 2025 — Agentic Tokens that bind a card to a specific agent, merchant, and consent policy, built on the same Mastercard Digital Enablement Service tokenization that already powers Apple Pay and Google Pay &lt;sup&gt;1&lt;/sup&gt;. Visa's Trusted Agent Protocol followed that October, giving merchants a way to verify an agent's identity before checkout rather than guess whether it's a bot &lt;sup&gt;1&lt;/sup&gt;. Both networks have since moved from announcement to production. Mastercard says the first fully autonomous agentic payment — an AI agent booking a ride through a mobility provider called Hoppa, authenticated by DBS Bank with no human confirming at the point of sale — happened in Singapore on March 4, 2026, and in June 2026 the company launched Agent Pay for Machines with more than thirty partners spanning cards, bank accounts, and stablecoins on one settlement layer &lt;sup&gt;11&lt;/sup&gt;. Visa, meanwhile, struck a partnership with OpenAI the same month to embed tokenized Visa credentials directly into agent-initiated checkout inside ChatGPT and Codex, with spending limits, merchant categories, and required-approval rules enforced at the network level &lt;sup&gt;12&lt;/sup&gt;. And in May 2026, Google donated AP2 itself to the FIDO Alliance, alongside a Mastercard-built companion standard called Verifiable Intent, with new FIDO working groups on agentic authentication and agentic payments now chaired jointly by Mastercard, Visa, Google, OpenAI, and Amazon &lt;sup&gt;13&lt;/sup&gt;.&lt;/p&gt;

&lt;p&gt;Three competitors who spent decades fighting each other for interchange share converged on overlapping, interoperable standards for machine-initiated payments inside about thirteen months, and then handed the governance of those standards to a neutral standards body so no single one of them could own it outright. That is not a market with room for a new entrant building "a payment rail for agents." That market closed before most product teams finished their first planning cycle on the topic. If a line item on your 2027 roadmap is infrastructure that competes with Agent Pay, TAP, or AP2 directly, kill it now and redeploy the budget.&lt;/p&gt;

&lt;p&gt;The opportunity that survives this consolidation sits one layer up, exactly where Chapter 1 said the moat would be once a protocol goes free: the &lt;strong&gt;policy semantics&lt;/strong&gt; layer that decides what a Policy Wallet is actually allowed to do, and the evidence layer that proves who's really on the other end of the transaction. Neither Mastercard nor Visa nor Google is in the business of encoding your specific procurement rules — "this agent may reorder consumables within five percent of contract price, from this vendor list, capped at this monthly ceiling" — into an enforceable Intent Mandate. Somebody has to compile that policy, in that specificity, and somebody has to package the KYC/KYB evidence that proves the counterparty accepting the payment is who and what it claims to be, especially when the counterparty is itself software. That compiling and packaging work is unglamorous, vertical-specific, and exactly the kind of boring infrastructure that compounds instead of getting commoditized by the next protocol release.&lt;/p&gt;

&lt;p&gt;It also happens to be the work a CFO actually needs done before any of this touches a real treasury account. No finance leader signs off on a Policy Wallet because the rails are elegant. They sign off when three specific questions have answers they can put in a board deck: what, precisely, is this agent allowed to do without me; what happens to the money if that policy gets exploited or simply misfires; and can I explain the wallet's behavior to an auditor, a regulator, or a plaintiff's lawyer without pointing at a model and shrugging. Those three questions are, not coincidentally, Chapter 1's moat restated as a procurement checklist — evidence, warranty, explainability — and they are the actual gate between "we piloted an agentic payments demo" and "we let it move real money unattended." The rest of this chapter is about what has to exist on each side of that gate.&lt;/p&gt;

&lt;h3&gt;
  
  
  When the vendor on the other end is also a wallet
&lt;/h3&gt;

&lt;p&gt;Here is the part that changes the shape of the problem rather than just its scale. Everything so far describes one agent — a shopping assistant, a procurement bot — transacting against a human-run business. That's the easy half. Mastercard built Agent Pay for Machines specifically for the other half: machine-to-machine commerce, where an IoT sensor pays per API call, a logistics agent pays a warehousing agent for storage by the hour, and neither side of the transaction has a person checking a dashboard &lt;sup&gt;11&lt;/sup&gt;. When the payer's wallet &lt;em&gt;and&lt;/em&gt; the vendor's invoicing system are both agent-operated, "software hires software" stops being a checkout-page feature and becomes the entire commercial relationship.&lt;/p&gt;

&lt;p&gt;That reshapes KYC and KYB, because the question is no longer just "is this business real." It's "which specific agent is acting, under whose authority, with what spending limit, and can that be proven independently of the agent's own say-so." An emerging shorthand for this — Know Your Agent — is circulating among identity vendors and standards groups as of 2026: the idea that every acting agent needs to be cryptographically bound to a verified human or business principal, with its permission scope machine-readable and its behavior auditable over time &lt;sup&gt;14&lt;/sup&gt;. Treat that framing as directional rather than settled; there is no single regulator-blessed KYA standard the way there is a chartered bank behind KYC. But the underlying problem it's naming is real and it's exactly the one a Policy Wallet has to solve before a CFO will let it run unattended: prove not just that the vendor exists, but that the specific piece of software claiming to represent the vendor was actually authorized to send that invoice.&lt;/p&gt;

&lt;p&gt;This creates a specific failure mode. In a human-to-business transaction, a confused or malicious counterparty eventually runs into a person who notices something is off — a purchasing manager who flags an unusual invoice, a customer who disputes a charge. In an agent-to-agent transaction, that noticing has to be engineered in deliberately, because by default nobody is watching. Whichever side of the transaction is holding the better evidence trail — the more complete, better-provenanced record of what was authorized, by whom, under what policy — wins the dispute when it eventually happens, whether that dispute plays out as a chargeback, an audit finding, or a lawsuit. That is the KYC/KYB evidence packager business in one sentence: sell the receipts that make a machine-to-machine dispute resolvable by something other than "our agent said your agent said."&lt;/p&gt;

&lt;h3&gt;
  
  
  The moat, in wallet terms
&lt;/h3&gt;

&lt;p&gt;Chapter 1's three-part moat — verified facts with provenance, warranted outcomes, accountable judgment — translates into this domain almost without modification, which is a good sign that it's the right framework rather than a slogan.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Verified facts with provenance&lt;/strong&gt; is literally the KYC/KYB evidence trail, and the infrastructure to generate it is arriving on a hard regulatory clock rather than a product roadmap. Under the EU's eIDAS 2.0 regulation, banks, insurers, and payment providers must accept the EU Digital Identity Wallet as an authentication method for onboarding and strong customer authentication by December 2027, with each member state required to have at least one government-issued wallet live a year earlier &lt;sup&gt;15&lt;/sup&gt;. That deadline is real, but the rollout underneath it is not going smoothly — reporting as of September 2026 describes the wallet stalling in roughly two dozen member states even as the regulatory clock keeps running — a reason not to bet a 2030 plan on flawless infrastructure arriving on schedule &lt;sup&gt;15&lt;/sup&gt;. India is running a parallel experiment with a different mechanism: the DPDP Act's Consent Manager framework goes live on November 13, 2026, formalizing the licensed-broker model that the country's Account Aggregator network has been operating since 2021, where a regulated consent manager moves encrypted financial data between institutions without ever holding the decryption key itself &lt;sup&gt;16&lt;/sup&gt;. As of late 2025, seventeen licensed Account Aggregators connect more than 135 financial information providers, with roughly 38 percent of borrowers reportedly enabled on the network — figures worth treating as directional, since they come from industry secondary sources rather than the regulator itself &lt;sup&gt;17&lt;/sup&gt;. Both systems are, underneath the acronyms, the same idea: a verifiable, consent-scoped, revocable evidence trail attached to a financial fact, which is precisely what a KYC/KYB evidence packager needs as raw material.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Warranted outcomes&lt;/strong&gt; is where this domain gets genuinely uncomfortable, because nobody has actually settled who pays when a Policy Wallet gets it wrong. Agentic payments break the assumption baked into decades of consumer-protection law that a transaction has one clean moment of authorization by one identifiable person. Under Regulation E in the US, a transfer is generally presumed authorized once a consumer hands credentials to an agent, even if that agent then acts outside what the consumer actually intended — which means the consumer-protection backstop that would normally absorb a bad agent decision may simply not apply &lt;sup&gt;18&lt;/sup&gt;. Legal analysis of the current wave of agent payment tools is blunt about the state of play: the question of who absorbs the loss — the person who authorized the agent, the developer who built it, or the merchant who received the funds — will likely be settled by whoever litigates first, and no company wants to be that test case &lt;sup&gt;18&lt;/sup&gt;. That uncertainty is not a reason to wait. It's the actual product opportunity. A mandate policy engine that merely executes a rule is a feature. A mandate policy engine backed by a warranty — a vendor who prices, underwrites, and stands behind the accuracy of its own policy enforcement — is a business, for the same reason an insurer is a business and a weather app is not.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Accountable judgment&lt;/strong&gt; has the sharpest edge of the three, because it's not really a business risk here — it's a legal wall. The CFPB's 2022 guidance established that a lender cannot use a credit-decision model too opaque to produce specific, accurate reasons for turning someone down; "the black box told us to" is not, and has never been, an acceptable adverse-action notice under the Equal Credit Opportunity Act and Regulation B &lt;sup&gt;19&lt;/sup&gt;. In May 2025, the CFPB withdrew that guidance along with dozens of other interpretive documents as part of a broader deregulatory push &lt;sup&gt;19&lt;/sup&gt;. That withdrawal did not repeal ECOA or Regulation B — the statute is still the statute, and a creditor still has to produce specific, accurate reasons for a denial or face liability under the underlying law, guidance or no guidance. If anything, the withdrawal makes the constraint sharper for a builder, not softer: the safe harbor of clear agency guidance is gone, and what's left is bare statutory exposure that gets tested in court rather than pre-cleared by a regulator. This is exactly why SMB cash-flow underwriting built from consented bank data with reason codes traceable to specific transactions is more defensible than a generic underwriting model wrapped in a chat interface. Explainability here is not a nice-to-have UX layer. It is the difference between a product a bank's counsel will approve and one they legally cannot.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;A wallet that can't explain why it said no isn't a product a bank can ship. It's a liability nobody has been sued over yet.&lt;/p&gt;
&lt;/blockquote&gt;
Corporate wallets: rails standardize in 2026, policy layer builds in 2027, positioning for 2030.



&lt;h3&gt;
  
  
  What to build, and how to charge for it
&lt;/h3&gt;

&lt;p&gt;The Policy Wallet's economics split cleanly along Chapter 0's menu, and the split matters because pricing a mandate policy engine like a SaaS seat, instead of like an insurance product, is the fastest way to underprice the actual risk you're taking on.&lt;/p&gt;

&lt;p&gt;A &lt;strong&gt;mandate policy engine&lt;/strong&gt; — the compiled, enforceable version of "this agent may reorder consumables within five percent of contract price" — is sold as a tiered subscription per wallet or per policy ruleset, with the tier gating both the complexity of policy a customer can express and the transaction throughput the engine will authorize per hour, the same kind of rate-limited access-tier structure that governs how much a role can do in the mortgage chapter's product ladder. Layer a metered per-transaction fee on top for actual mandate checks executed above the tier's included volume, because the marginal cost of evaluating one more transaction is real and should be priced as such, separately from the fixed cost of having a policy engine at all.&lt;/p&gt;

&lt;p&gt;A &lt;strong&gt;KYC/KYB evidence packager&lt;/strong&gt; for agent and non-human counterparties is naturally metered per verified entity — a fee each time the packager stands up a fresh evidence bundle proving who's really on the other end of a payment — with a subscription tier for institutions verifying at high volume who want a predictable bill instead of a variable one.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;SMB cash-flow underwriting facts&lt;/strong&gt;, built from consented bank data with adverse-action-ready reason codes, is where outcome-warranted pricing belongs: charge per underwriting decision the facts actually support, with a defined liability position if a reason code turns out to be wrong or unsupportable under ECOA scrutiny — the same logic that makes a mortgage QC warranty worth more than a subscription to an extraction API.&lt;/p&gt;

&lt;p&gt;And the connective tissue between all three — the one-time integration fee for actually plugging into a bank's core system, a payment network's mandate API, or an EUDI Wallet relying-party integration — is usually the hardest six weeks of the entire sale. Pricing it at zero to win the logo doesn't save the customer money. It just moves that cost from their budget onto yours.&lt;/p&gt;

&lt;h3&gt;
  
  
  2030: the financial passport, staged rather than scheduled
&lt;/h3&gt;

&lt;p&gt;Treat everything in this section the way Chapter 1 insisted you treat any 2030 material: a plausible direction to position for, not a date to build against. Nothing below carries a citation, because nothing below has happened yet.&lt;/p&gt;

&lt;p&gt;The building blocks are visibly assembling. A continuous, real-time audit — where a ledger is reconciled to source evidence constantly rather than sampled quarterly by a human auditor — becomes newly plausible once the underlying facts (bank transactions, invoices, payroll records) already carry the kind of provenance the EUDI Wallet and India's Account Aggregator network are built to produce. An auditor who currently samples a few hundred transactions and extrapolates could instead consume a standing feed of attestations and only intervene on exceptions. That is a genuine step change in how assurance work gets done, and it points to a live 2030 opportunity: sell the attestation layer, not the audit itself.&lt;/p&gt;

&lt;p&gt;The more ambitious version is a portable "financial passport" — a verifiable-credential bundle of income, assets, and obligations that a person or a small business refreshes by consent and reuses across a mortgage application, a rental screening, and an SMB credit line, instead of re-proving the same facts to three different institutions in three different formats. The pieces exist in isolated form today: EUDI Wallet credentials, India's Account Aggregator and DigiLocker rails, and the KYC/KYB evidence packagers this chapter has already described as a 2027 build target. What doesn't exist yet is the interoperability and institutional trust to let one of those bundles walk unmodified from a bank's underwriting system into a landlord's screening tool into a different bank's SMB lending desk. That requires competing institutions to accept each other's evidence at face value, which is a trust problem, not a technology problem — and trust problems move on regulatory and reputational timelines, not sprint timelines. Anyone who tells you the financial passport ships broadly by a specific quarter is selling something. The honest plan is to build the evidence infrastructure now, in whichever single vertical you already touch, in a form that's portable in principle even before anyone else agrees to accept it.&lt;/p&gt;

&lt;p&gt;Which is exactly the version of this story mortgage lending is already living through, one step ahead of the rest of finance. A wallet answers who gets to spend, on what authority, with what evidence behind the mandate. A mortgage answers a harder question sitting right underneath it: once the money is authorized, who gets to decide what it's actually for, and whose judgment does a lender trust enough to let the loan close without a human re-checking every fact one more time. That's Chapter 3.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Decision That Signs the Loan
&lt;/h2&gt;

&lt;p&gt;For a decade, the mortgage question that made someone money was: what does this document say. The question worth paying for now is: given what these documents say, is this transaction acceptable, and what has to happen before it can close.&lt;/p&gt;

&lt;p&gt;That is not a small shift in phrasing. It is the whole industry's moat moving to a different floor of the building. Extraction — turning a scanned paystub or a bank statement into structured fields a computer can use — is becoming a utility, priced like one and improving on someone else's roadmap, not yours. The layer above it is a different story: reconciling a borrower's paystub against their W-2 against their bank deposits against the county's recording requirements against the investor's overlay against the lender's own risk appetite, and producing a verdict that a regulator will accept eighteen months later if the loan goes bad. That layer has a name now. Call it &lt;strong&gt;the Decision Graph&lt;/strong&gt; — a versioned, executable map of which rules apply to which loan, at which point in time, with which exceptions, and what happens when one of them fails. It is the thing an origination agent and an underwriting agent actually negotiate over when they sit down — machine to machine — to get a loan approved. It is also the thing that survives being commoditized, because it was never made of documents in the first place.&lt;/p&gt;

&lt;h3&gt;
  
  
  The reading is free now
&lt;/h3&gt;

&lt;p&gt;Look at where the price of "reading a mortgage document" actually sits in September 2026. Reducto, one of the sharper document-parsing platforms built for LLM pipelines, lists its standard extraction rate at \$0.015 per credit — one credit per page per operation — with volume discounts below that for larger contracts &lt;sup&gt;20&lt;/sup&gt;. IBM shipped Docling for watsonx as a fully managed service in June 2026, built on an open-source toolkit that had already passed seventy million downloads, priced at roughly four dollars per thousand pages &lt;sup&gt;21&lt;/sup&gt;. Hyperscaler OCR sits underneath both of them as the free-to-cheap floor. This isn't one vendor losing a price war — it's an entire capability, turning a PDF into JSON, falling to commodity pricing inside about eighteen months, the same pattern the opening chapter of this book describes happening to one wave after another: the plumbing compounds quietly and gets cheap fast, while the exciting front-end bet is what everyone keeps arguing about.&lt;/p&gt;

&lt;p&gt;Even the market-sizing exercises around intelligent document processing can't agree on how big the category is, and the disagreement itself is the signal. One widely cited estimate puts the IDP market at \$4.3 billion in 2026, growing to \$32.9 billion by 2033. Another puts it at \$1.45 billion in 2026, growing to just \$2.02 billion by 2033 &lt;sup&gt;22&lt;/sup&gt;. That is roughly a 3x disagreement on the starting number and a completely different growth story layered on top of it — one analyst sees a category about to be the next infrastructure layer of finance, another sees a mature tool market growing in line with inflation. Neither estimate is obviously fraudulent; they are almost certainly counting different things under the same label. The lesson for anyone building here is not "trust the bigger number" or "trust the more conservative one." It's that a market-sizing slide should never be load-bearing in your plan. If two respected research houses can't agree within a factor of three on how large the pond is, you should be building for a specific defensible catch, not for the pond's abstract size.&lt;/p&gt;

&lt;p&gt;That specific catch is visible in who is actually raising money in this space right now, and what they're selling. Sela raised \$21 million across seed and Series A rounds in 2026 for voice agents that run mortgage sales calls end-to-end, handling borrower conversations and escalating to a human loan officer only when needed; the company says six of the ten largest independent mortgage banks in the U.S. now run it in production, with a \$10 million annualized run rate reached in eighteen months &lt;sup&gt;23&lt;/sup&gt;. Balerion AI raised a \$6 million seed to build a "unified reasoning engine" that analyzes an entire loan file holistically rather than field by field, targeting the roughly \$12,000 average cost to originate a loan that manual processes and fragmented systems currently produce &lt;sup&gt;24&lt;/sup&gt;. Friday Harbor built an AI pre-underwriting platform and, in September 2026, integrated it directly with Freddie Mac's own Income Calculator API, so a lender can get a calculated qualifying income back in minutes with representation-and-warranty relief attached, before the file ever reaches a human underwriter &lt;sup&gt;25&lt;/sup&gt;. Every one of these companies is selling past extraction. None of them is pitching "we read your documents better." They're pitching a verdict, or the automation of the conversation that leads to one.&lt;/p&gt;

&lt;p&gt;None of that means extraction skill is worthless — someone still has to feed clean, tamper-checked, cross-reconciled facts into whatever makes the decision. It means extraction alone stopped being a business model roughly the moment a hyperscaler could match your accuracy for a fraction of a cent per page. The business that survives sits one layer up, in the graph of rules that decides what those facts mean.&lt;/p&gt;

&lt;h3&gt;
  
  
  What an agent actually asks a lender
&lt;/h3&gt;

&lt;p&gt;Picture the actual exchange, stripped of the marketing language both sides will eventually put around it. An origination agent — working on behalf of a broker, or increasingly a borrower's own assistant — has assembled a loan package: income documents, asset statements, the property address, the requested product. It doesn't want a document review. It wants an answer to one compound question: given this borrower, this property, this jurisdiction, this investor, and this lender's own overlays, is this transaction acceptable, and if not, exactly what needs to change to make it acceptable.&lt;/p&gt;

&lt;p&gt;It sends that package, tagged with the context that actually determines the answer — state, county, transaction type, investor, lender — to an underwriting agent sitting on top of a lender's or aggregator's Decision Graph. What comes back is not a dump of the applicable regulations. It's a decision: approved, denied, or review required, attached to specific findings ("escrow amount inconsistent with the applicable requirement, high severity"), a short list of missing items or required remediation ("recalculate escrow, regenerate the closing disclosure"), and a set of evidence identifiers and rule versions the finding is traceable to. The origination agent doesn't get the underlying rule text. It gets the verdict, the reason coded to a rule ID, and the version of that rule that was in force on the date of evaluation.&lt;/p&gt;

&lt;p&gt;That last withholding is the whole design decision, and what's actually being protected is not the regulation. California's escrow requirements are public. The CFPB's disclosure rules are public. Fannie Mae's Selling Guide is public. Anyone — including a well-resourced competitor with a scraper — can read the primary source material for free. What is not public, and what took years and thousands of adjudicated loan files to build correctly, is the &lt;em&gt;interpretation&lt;/em&gt;: which of several overlapping rules actually governs a specific fact pattern, what takes precedence when a state requirement and an investor overlay disagree, which exceptions apply and under what conditions, how a rule's effective date interacts with a loan's lock date, and which validation actually catches the failure before it becomes a repurchase demand eighteen months later. That mapping — rule to jurisdiction to transaction type to document to field to condition to exception to validation to remediation — is the Decision Graph, and it is considerably harder to scrape than a knowledge base, because scraping only gets you the inputs. It doesn't get you the thousands of resolved edge cases that taught the system which input wins when two rules point in different directions.&lt;/p&gt;

&lt;p&gt;This is also why exposing "search my regulations" as the product would be a mistake regardless of who builds it. A raw query interface hands a competitor your whole corpus one request at a time. A verdict interface — approved, denied, review-required, plus reason and remediation, plus a rule version and an evidence trail — gives a counterparty everything it actually needs to act, and gives away nothing it could use to rebuild your rule graph from the outside. That is not a technical nicety. It is the entire difference between a defensible business and a very well-organized free knowledge base with a login page.&lt;/p&gt;

&lt;p&gt;None of this works if the request is routed the way a naive "MCP server full of mortgage knowledge" would route it — straight from an agent into a shared corpus. The layer that has to sit in between is what turns a lookup into a verdict:&lt;/p&gt;
The Decision Graph architecture: from access layer to provenance and audit.



&lt;p&gt;The model engine's job in this picture is meaning, not measurement. The parts of the pipeline that touch escrow math, tolerance thresholds, and eligibility cutoffs stay deterministic on purpose, because "the LLM said it looked compliant" is not a sentence that survives a repurchase demand. The Decision Graph lives underneath the Knowledge Service and Rules Engine boxes — it's the thing that tells the Context + Policy Router which jurisdiction, investor, and lender rules even apply before either engine runs, and in what order they take precedence when two of them disagree.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;A lender that answers "is this loan acceptable" is selling a verdict. A lender that answers "here is everything I know about escrow law" is giving away the only asset it has left.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h3&gt;
  
  
  The three things a Decision Graph has to prove
&lt;/h3&gt;

&lt;p&gt;Chapter 1 named the three things that create a moat once a protocol itself is free: verified facts with provenance, warranted outcomes, and accountable judgment. Mortgage lending is the cleanest place in this book to see all three attach to one artifact.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Verified facts with provenance&lt;/strong&gt; means every value the Decision Graph produces carries its receipt: which document it came from, which page, which field, computed under which rule version, at what confidence, timestamped. An income figure isn't just a number — it's a number with a path back to the paystub, the W-2, the bank deposit that corroborates it, and the specific rule that governed how those three sources were reconciled into a single qualifying-income figure. That's expensive to build correctly and unglamorous to describe in a pitch deck, which is exactly the combination the opening chapter flags as defensible: nobody builds this for a demo, because a demo doesn't need to survive an audit.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Warranted outcomes&lt;/strong&gt; means someone prices the cost of being wrong, not just the cost of answering. This is where mortgage lending offers the cleanest 2027-scale business inside the entire IDP category: post-close quality-control and repurchase-risk scoring, priced per loan with an actual defect warranty attached. Instead of selling "we'll tell you if the file looks compliant," the vendor sells "we'll tell you the probability this specific loan gets kicked back by the investor after purchase, and if we're wrong beyond an agreed threshold, we eat some of the cost." That's a fundamentally different sale than a subscription to an extraction API, and it's only possible once you have a Decision Graph precise enough to price its own error rate.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Accountable judgment&lt;/strong&gt; is the piece people assume can't exist yet for a decision &lt;em&gt;system&lt;/em&gt;, as opposed to a licensed human underwriter — and mortgage lending already has a working counterexample. Fannie Mae's Income Calculator gives lenders representation-and-warranty enforcement relief for the accuracy of an income calculation, on an income-source basis, when the lender uses the calculator's output as submitted &lt;sup&gt;26&lt;/sup&gt;. Freddie Mac's Asset and Income Modeler, built into Loan Product Advisor, does the same for income, asset, and increasingly employment verification — and Freddie Mac reports that loans originated using AIM are roughly 2.1 times less likely to produce a defect &lt;sup&gt;27&lt;/sup&gt;. Freddie Mac has also published that lenders who maximize this kind of automation originate loans that run about \$1,500, or 14 percent, cheaper, with production cycles roughly five days shorter &lt;sup&gt;28&lt;/sup&gt;. What that actually means is this: two of the most conservative institutions in American finance have already agreed, in writing, in their own selling guides, to hold &lt;em&gt;themselves&lt;/em&gt; accountable for a decision a system made, provided the system's inputs and rule version are documented well enough to audit later. This isn't hypothetical. It's a precedent already operating at scale, for the exact class of decision this chapter is about. The remaining question isn't whether accountable judgment can attach to a system's output — it's how far past income calculation that relief extends, and which vendor's Decision Graph earns it next.&lt;/p&gt;

&lt;h3&gt;
  
  
  The product ladder: what to sell, and to whom
&lt;/h3&gt;

&lt;p&gt;A Decision Graph is not one product with one price. It is a ladder, and Chapter 0's menu maps onto every rung of it cleanly enough that the pricing model should be obvious the moment you know which rung you're selling:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Extract&lt;/strong&gt;: &lt;strong&gt;What the customer gets:&lt;/strong&gt; Structured data from mortgage documents; &lt;strong&gt;How it's priced:&lt;/strong&gt; Metered, per page or per document&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Validate&lt;/strong&gt;: &lt;strong&gt;What the customer gets:&lt;/strong&gt; Document and data validation against a rule set; &lt;strong&gt;How it's priced:&lt;/strong&gt; Metered, per document, higher rate than raw extraction&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Compliance&lt;/strong&gt;: &lt;strong&gt;What the customer gets:&lt;/strong&gt; Regulatory and disclosure checks (TRID tolerance, escrow rules); &lt;strong&gt;How it's priced:&lt;/strong&gt; Subscription tier, rate-limited by monthly check volume&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Eligibility&lt;/strong&gt;: &lt;strong&gt;What the customer gets:&lt;/strong&gt; Investor and lender guideline matching; &lt;strong&gt;How it's priced:&lt;/strong&gt; Subscription tier, one rung up, more rules exposed&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Decision&lt;/strong&gt;: &lt;strong&gt;What the customer gets:&lt;/strong&gt; Full transaction-level verdict with evidence and remediation; &lt;strong&gt;How it's priced:&lt;/strong&gt; Per loan file, flat fee&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Continuous&lt;/strong&gt;: &lt;strong&gt;What the customer gets:&lt;/strong&gt; Re-checks automatically when a regulation or guideline changes; &lt;strong&gt;How it's priced:&lt;/strong&gt; Subscription add-on to the Decision tier&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Enterprise&lt;/strong&gt;: &lt;strong&gt;What the customer gets:&lt;/strong&gt; Customer-specific overlays computed into the executable policy; &lt;strong&gt;How it's priced:&lt;/strong&gt; Annual contract, one-time integration fee plus per-loan usage&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Moving down that ladder, the pricing model shifts from metered (Extract, Validate — commodity operations, priced by volume) to subscription (Compliance, Eligibility — ongoing access to a maintained rule set) to a per-transaction flat fee (Decision — a discrete verdict on a discrete loan) to a blended enterprise contract (a fixed integration fee for wiring in a lender's specific overlays, plus ongoing per-loan usage on top). Sell the wrong pricing model at the wrong rung — metering a Decision, or trying to flat-fee raw Extraction — and you either leave money on the table or price yourself out of a market that has already commoditized that rung.&lt;/p&gt;

&lt;p&gt;The same ladder logic applies to &lt;em&gt;who&lt;/em&gt; gets access, not just &lt;em&gt;what&lt;/em&gt; they're charged. A Decision Graph exposed through role manifests — a loan officer's seat with roughly eight tools and a modest daily call limit, a processor's seat with around fifteen and a higher limit, an underwriter's seat with twenty-plus tools and the widest data access, a QC auditor's seat scoped narrowly to post-close review — is simultaneously a usability decision (nobody wants forty tools loaded into one agent's context) and a monetization decision (a lender pays more per seat for a role that touches more of the Decision Graph, the same way a database vendor charges more for an admin seat than a read-only one).&lt;/p&gt;

&lt;p&gt;And sitting above every rung of the ladder is the one tier this chapter has already named as the real 2027 prize: post-close QC and repurchase-risk scoring, priced per loan with an actual defect warranty — outcome-warranted pricing, sold as an add-on to the Decision tier rather than a replacement for it, because a lender who already trusts your verdict enough to buy it is the easiest possible customer to sell a warranty on that same verdict to next.&lt;/p&gt;

&lt;h3&gt;
  
  
  2030: always pre-approved, never fully autonomous
&lt;/h3&gt;

&lt;p&gt;The staging is what makes this a plan instead of a wish:&lt;/p&gt;
Mortgage lending: the same 2026 → 2027 → 2030 horizon.



&lt;p&gt;Extrapolate that precedent forward and the plausible 2030 picture comes into focus — everything from here is a bet on direction, not a reported fact. Picture a borrower's own agent holding a standing, continuously re-evaluated conditional approval rather than a one-time pre-qualification letter that goes stale the moment a pay stub changes. New income data lands, a Decision Graph re-runs the eligibility check automatically, and the approval updates itself — delivered not through a portal login but agent-to-agent, the same way a procurement agent checks a vendor's signed claims before opening a negotiation. Fannie and Freddie are already most of the way to the technical precondition for this: income and asset facts computed once, by an authoritative system, with relief attached to using them. What's missing for "always pre-approved" isn't the compute. It's the consented, refreshable data pipes and the regulatory comfort to treat a standing approval as something other than a one-time snapshot.&lt;/p&gt;

&lt;p&gt;What is not plausible on any near-term timeline is the other half of that fantasy: an agent auto-locking a rate and closing a loan with no human ever touching the file. That's not a technology gap closing slowly — it's the Mandate Boundary sitting where the opening chapter said it would, drawn by the cost of being wrong rather than by what a model can technically execute. Preparing a decision, assembling evidence, and re-checking eligibility as facts change are all activities the boundary has already moved to accommodate, because a wrong pre-approval is embarrassing and correctable. Executing an irreversible six-figure financial commitment is not correctable in the same way, and TRID and ECOA disclosure requirements exist precisely to put a human signature between "the system recommends" and "the transaction is final." The pattern to build toward for 2030 is a lender's Decision Graph getting fast enough and trusted enough that the human signature becomes the very last, cheapest step in the process — not a pattern where that signature disappears. Watch for the boundary to keep moving on preparation for years before anyone credible proposes moving it on execution, and be skeptical of any vendor whose 2027 roadmap quietly assumes otherwise.&lt;/p&gt;

&lt;p&gt;The underwriting agent that decides whether a loan closes and the payer agent that decides whether a claim gets paid are running the same logic wearing different regulatory costumes — a package of facts, a rule graph nobody outside the institution gets to read directly, and a verdict with a reason code attached. Chapter 4 walks into the second costume, where the agent on the other side of the negotiation isn't deciding who gets a house. It's deciding who gets care.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Claim That Argues With Itself
&lt;/h2&gt;

&lt;p&gt;Somewhere right now, a piece of software representing a surgeon's office is assembling the case for a hip replacement, and a piece of software representing the insurer is testing that case against a coverage policy — and it is entirely possible that no human being on either side has opened the patient's chart yet. That is not a bug in the system. As of January 2026, in a meaningful slice of American healthcare, it is the system, and for once it is the system on purpose, by federal order, on a published clock.&lt;/p&gt;

&lt;p&gt;Call the back-and-forth between those two systems &lt;strong&gt;the Adjudication Loop&lt;/strong&gt;: the automated exchange in which a provider-side agent assembles a clinical case, a payer-side agent tests it against coverage criteria, evidence gets requested and resupplied, and the loop keeps iterating — sometimes closing in seconds — before a human on either side is pulled in. Every other domain in this book has a version of software negotiating with software over money. This is the one where the thing being negotiated is a person's access to care, and where getting the loop wrong is not a quarter of lost revenue. It is a denied hip replacement, a delayed chemotherapy start, a patient readmitted because the skilled-nursing stay got cut short by an algorithm that never saw the discharge note. The stakes change the entire calculus of where the Mandate Boundary can sit.&lt;/p&gt;

&lt;h3&gt;
  
  
  The clock the government set
&lt;/h3&gt;

&lt;p&gt;Every other chapter in this book describes a Mandate Boundary moving because a vendor got ambitious or a market got competitive. Prior authorization is different: the boundary is moving because the government wrote the date into a federal rule and started enforcing it.&lt;/p&gt;

&lt;p&gt;CMS-0057-F, the Interoperability and Prior Authorization Final Rule, carries real enforcement weight, not vendor-roadmap ambition — it is binding on Medicare Advantage organizations, state Medicaid and CHIP fee-for-service and managed-care programs, and Qualified Health Plan issuers on the federally facilitated exchanges &lt;sup&gt;29&lt;/sup&gt;. Two of its deadlines already fell before this chapter was written. Since January 1, 2026, impacted payers must decide expedited requests within 72 hours and standard requests within 7 calendar days, and a denial can no longer arrive as a bare "no" — it has to come with a specific reason &lt;sup&gt;29&lt;/sup&gt;. On March 31, 2026, those same payers filed their first public prior-authorization metrics — approval and denial rates, appeal outcomes, average turnaround — and they will keep filing them every year from here &lt;sup&gt;30&lt;/sup&gt;. A second, larger deadline is still ahead: by January 1, 2027, impacted payers must expose four FHIR-based APIs, including a Prior Authorization API that lets a provider's system submit a request and track its status electronically instead of by fax and hold music &lt;sup&gt;29&lt;/sup&gt;.&lt;/p&gt;

&lt;p&gt;Two qualifications matter, because they show the boundary moving on a schedule that is deliberately narrower than the hype around it. First, drug prior authorization is explicitly excluded from CMS-0057-F — a companion rule, CMS-0062-P, proposed in April 2026, would extend similar electronic requirements to drug PA, but as of this writing it is still a proposal with a closed comment period and no finalized date &lt;sup&gt;31&lt;/sup&gt;. Second, traditional Medicare and standalone Part D sit outside the rule entirely. This is not "AI is coming for all of healthcare administration by 2027." It is a specific, government-forced modernization of the medical-benefit prior-auth pipeline for specific payer types, with drug PA deliberately carved out for later. Build your 2027 plan against the rule that actually exists, not the version of it that sounds better in a pitch deck.&lt;/p&gt;

&lt;p&gt;What CMS actually built, whether or not that was the intent, is a forcing function for exactly the kind of infrastructure this book has been arguing you need everywhere else voluntarily: evidence with a timestamp, decisions with a reason code, and outcomes reported in public where a regulator, a journalist, or a plaintiff's attorney can find them. The Mandate Boundary in prior authorization is being drawn by statute before most organizations have finished drawing it by choice.&lt;/p&gt;

&lt;h3&gt;
  
  
  Why the loop can say yes alone, and not no
&lt;/h3&gt;

&lt;p&gt;The clearest evidence that the software-hires-software mechanic is already operating at scale sits inside a single company's numbers. Cohere Health processes more than 12 million prior-authorization requests a year for more than 600,000 providers, on the strength of a \$90 million Series C led by Temasek that brought its total funding to \$200 million &lt;sup&gt;32&lt;/sup&gt;. Its CEO, Siva Namasivayam, has said publicly that the AI renders a real-time decision in roughly 85% of those cases &lt;sup&gt;32&lt;/sup&gt;. A smaller competitor, Anterior, raised a \$20 million Series A led by NEA at a \$95 million valuation on the same thesis a year earlier &lt;sup&gt;33&lt;/sup&gt;. This is production infrastructure processing a double-digit-million volume of live clinical decisions, unaccompanied, most of the time — not a demo.&lt;/p&gt;

&lt;p&gt;Read the sentence that follows Namasivayam's 85% figure, though, and you find the actual shape of the Mandate Boundary in this domain: "No claim is denied exclusively by AI" &lt;sup&gt;32&lt;/sup&gt;. The remaining 15% of decisions — and, more importantly, every decision that trends toward a denial rather than an approval — gets a nurse or a physician's sign-off before it becomes final. That is not a rounding error in the rollout. It is the whole architecture.&lt;/p&gt;

&lt;p&gt;The asymmetry here is real: the Adjudication Loop is free to say &lt;em&gt;yes&lt;/em&gt; by itself. It is not free to say &lt;em&gt;no&lt;/em&gt; by itself. The same model that approves a clean case at 2 a.m. is perfectly capable of flagging one as non-compliant with criteria — the constraint isn't technical, it's a liability allocation, and it maps precisely onto the Mandate Boundary from Chapter 1: the boundary does not sit at "is the decision hard" or "is the model accurate enough." It sits at "which direction does the error cut." An erroneous approval wastes money and can usually be caught and clawed back later. An erroneous denial withholds care from a person who may not have the runway to wait for an appeal. Software gets the mandate to act alone in the direction where being wrong is expensive but recoverable. It does not get the mandate to act alone in the direction where being wrong might not be recoverable at all.&lt;/p&gt;

&lt;p&gt;This is also, increasingly, written into law rather than left to vendor discretion. California's SB 1120 — the Physicians Make Decisions Act — requires that a medical-necessity determination be made by a licensed physician or a licensed health professional competent to evaluate the specific clinical question, and it explicitly bars an algorithm from denying, delaying, or modifying care based on medical necessity, in whole or in part &lt;sup&gt;34&lt;/sup&gt;. The FDA's own January 2026 update to its clinical-decision-support guidance moves in a similar, careful direction: it grants a CDS tool enforcement discretion to surface a single recommendation instead of a menu of options — but only when a single option is genuinely the only clinically appropriate one, and only when a clinician retains the ability to independently review the basis for it &lt;sup&gt;35&lt;/sup&gt;. Regulators are not saying the AI can't be trusted to reason well. They are saying the party who absorbs the consequence of a wrong denial has to be a licensed person who can be named, tested, and held to account — and that requirement is likely to outlive any improvement in model accuracy, because it isn't really about accuracy. It's about who a court can call as a witness.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;An approval is a mandate the software can execute alone. A denial is a mandate that still needs a name attached to it — because "the model decided" is not an answer a regulator, a plaintiff's attorney, or a frightened patient will accept.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h3&gt;
  
  
  The moat when the decision is "does this patient get care"
&lt;/h3&gt;

&lt;p&gt;Apply the book's framework — verified facts, warranted outcomes, accountable judgment — to prior authorization, and each leg gets sharper teeth than it had in procurement, wallets, or mortgage lending, because the cost of getting any one of them wrong is measured in patient harm rather than balance-sheet risk.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Verified facts with provenance&lt;/strong&gt; means a clinical evidence packet that maps specific facts from the chart — the imaging finding, the failed conservative-treatment history, the lab value — directly onto the specific line item in the payer's coverage policy it is meant to satisfy, with the policy version and its effective date attached. This is the harder half of the market to build in right now: the payer side is already crowded with well-funded platforms like Cohere Health and Anterior, but the provider side — packaging a specialty practice's own documentation into a submission a payer's AI can approve on the first pass instead of bouncing back for more information — is comparatively open, and it's priced naturally per approved authorization rather than per API call, because the entire point of the product is getting to "yes" faster, not generating more traffic.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Warranted outcomes&lt;/strong&gt; are not a nice-to-have differentiator here; they are close to existential. In every other domain in this book, a bad AI decision produces an unhappy customer or a write-off. In prior authorization, a bad AI decision produces a lawsuit with a name on it. A class action against UnitedHealth Group, filed in late 2023, alleges that its nH Predict algorithm denied post-acute care with an extremely high error rate and little human review, with more than 80% of those denials reportedly reversed on appeal; a federal magistrate ordered broad discovery into UnitedHealthcare's AI-driven claims process in March 2026, and separately, a U.S. Senate investigation found the company's Medicare Advantage post-acute-care denial rate rose from 10.9% in 2020 to 22.7% in 2022 as automation scaled &lt;sup&gt;36&lt;/sup&gt;. Whatever the eventual legal outcome, the reputational shape of that story — "the AI denied grandma's rehab" — is now a known, replayable narrative that every payer and every vendor in this space has to build against. A warranty that puts real money behind an accuracy claim is the only credible answer to a story that has already been told once, publicly, in court filings — not a sales gimmick in this market.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Accountable judgment&lt;/strong&gt; is where this domain diverges most sharply from every other chapter in this book. Elsewhere, "accountable judgment" can be a warranty fund, an insurance policy, a named institution absorbing risk on a balance sheet. In prior authorization, in a growing number of states, it is a specific licensed human being, by statute. SB 1120 does not ask for a reviewable audit trail as a best practice — it makes the licensed clinician the load-bearing legal requirement for a denial to be valid at all &lt;sup&gt;34&lt;/sup&gt;. That is the Mandate Boundary rendered as literally as it will appear anywhere in this book: not a policy an institution sets for itself, but a signature a named, licensed person is legally required to provide before the decision means anything.&lt;/p&gt;
The Adjudication Loop: where a claim can be approved alone, and where it can't.



&lt;h3&gt;
  
  
  What to build, and how to charge for it
&lt;/h3&gt;

&lt;p&gt;Run prior authorization through Chapter 0's menu and the pricing follows the same asymmetry this chapter has already described between an approval and a denial.&lt;/p&gt;

&lt;p&gt;A &lt;strong&gt;provider-side PA evidence packager&lt;/strong&gt; — mapping a specialty practice's own documentation onto a payer's coverage criteria — is priced through outcome-warranted pricing: a fee per approved authorization, not per submission attempted, because the entire value proposition is getting to yes on the first pass instead of bouncing back and forth through the Adjudication Loop. A vendor charging per API call here is pricing the wrong thing; a practice doesn't care how many calls it took, it cares whether the patient got approved.&lt;/p&gt;

&lt;p&gt;A &lt;strong&gt;medical-records chronology and facts API&lt;/strong&gt; — page-cited chronologies for claims, life underwriting, or legal use — is a cleaner fit for metered or tiered subscription access, because the buyer here (an insurer's claims desk, a life-underwriting team, a law firm) is consuming a volume of records rather than betting on a single yes/no outcome; price it per record processed with a subscription tier for high-volume buyers, the same shape as the document-extraction pricing in the mortgage chapter.&lt;/p&gt;

&lt;p&gt;The &lt;strong&gt;warranty layer behind either product&lt;/strong&gt; deserves its own line item, priced closer to an insurance premium than a software fee, given what this chapter has already shown about the cost of getting it wrong: a percentage of the authorization value or a flat per-decision premium that scales with the litigation exposure of getting a denial wrong, sold explicitly as protection against the nH Predict-shaped headline risk every payer and vendor in this space now has to underwrite against.&lt;/p&gt;

&lt;p&gt;And the 2030 &lt;strong&gt;health passport&lt;/strong&gt; — patient-held, consented facts that arrive pre-verified at the next Adjudication Loop — has an unusual buyer for a subscription: not the patient, who shouldn't have to pay to be believed twice, but whichever side of the transaction saves the most when the loop resolves faster, which is most often the payer avoiding a lengthy appeal or the health system avoiding a denied, unpaid claim. Price it as a per-query fee paid by the party whose Adjudication Loop just got shorter, not as a subscription billed to the person the whole system is supposed to be serving.&lt;/p&gt;

&lt;h3&gt;
  
  
  What's real for 2030, and what's marketing
&lt;/h3&gt;

&lt;p&gt;The most quoted extrapolation about AI and medicine right now is some version of "instant, personalized drugs by 2030." Treat that claim as wrong the way it's usually stated, because the evidence people cite to support it doesn't actually support the consumer-scale version of the story.&lt;/p&gt;

&lt;p&gt;The case everyone points to is a real one, and it is genuinely remarkable: a CHOP and Penn Medicine team designed a bespoke base-editing therapy for an infant, known publicly as KJ, born with a life-threatening urea-cycle disorder called CPS1 deficiency. The team went from diagnosis to a custom therapy in roughly six months, gave the first dose in February 2025, and published the results in the &lt;em&gt;New England Journal of Medicine&lt;/em&gt; that May &lt;sup&gt;37&lt;/sup&gt;. In February 2026, the FDA followed up with a draft "Plausible Mechanism Framework" guidance, which would let sponsors seek approval for these individualized, n-of-1 genetic therapies for well-characterized, ultra-rare genetic mutations without running a traditional randomized trial that a single-patient population could never support &lt;sup&gt;38&lt;/sup&gt;. That is a real and significant loosening of the regulatory path for a real category of medicine.&lt;/p&gt;

&lt;p&gt;None of that is "instant," and none of it is consumer-scale. Six months is extraordinary by the standards of drug development; it is not instant by the standards of a patient in front of you today. The infrastructure required — a specialized academic center, a bespoke molecular design process, a case sufficiently well-characterized that "the plausible mechanism" holds up to FDA scrutiny — describes a handful of centers treating a handful of patients with specific, well-mapped mutations, not a pharmacy counter dispensing a genome-matched pill to anyone who walks in. By 2030, the realistic bet is platform-based n-of-1 genetic medicines — base editors, antisense oligonucleotides — reaching more ultra-rare, well-characterized mutations, faster and more cheaply than they do today, still delivered in months at academic centers, still nowhere near consumer scale. That is a genuine advance. It is not the sci-fi version, and a roadmap built on the sci-fi version will misallocate a decade of capital.&lt;/p&gt;

&lt;p&gt;The more useful 2030 bet is unglamorous, and it follows directly from the machinery this chapter has already described — so treat it explicitly as this book's own extrapolation rather than established fact: a portable "health passport," a patient-held bundle of consented, verifiable facts — diagnoses, medications, prior-authorization history, the specific evidence that satisfied a specific payer's criteria last time — that travels with the patient the way a credit history travels with a borrower. The plumbing already points this way: the Patient Access API that CMS is requiring by January 2027 exists to let a patient's own data follow them between payers &lt;sup&gt;29&lt;/sup&gt;; a health passport is what you get when a patient, rather than a payer, controls that data and hands relevant, pre-verified slices of it directly into the next Adjudication Loop. The payoff is not a faster diagnosis. It's a prior-authorization request that starts most of the way to "approved," because the evidence the payer's agent needs to see was already verified and attached the last time this patient needed similar care. That is a plausible, buildable extension of infrastructure that will exist by 2027. It is not guaranteed, and nothing in current CMS rulemaking commits to it — it is this book's best read of where the pattern goes next, offered as a bet worth positioning for, not a promise.&lt;/p&gt;
Prior authorization: the regulator-forced horizon from 2026 to 2030.



&lt;p&gt;This is the last domain in this book, and it is worth closing on the plainest version of the Mandate Boundary you will find anywhere in it. Across a procurement agent shopping from another agent, a wallet that authorizes its own spending, a mortgage decision engine that recommends before a underwriter signs, and now a claim that argues with itself before either side's human sees it, the pattern never actually changes shape. Software gets faster, cheaper, and more capable of preparing the case — assembling the evidence, running the comparison, drafting the recommendation — and a specific, named, accountable party still has to sign it, and that party's identity, not the software's confidence score, is what a court, a regulator, or a frightened family will ask for when something goes wrong. What changes from domain to domain, and what this book has tried to make legible chapter by chapter, is only where that signature sits today and what it would take to move it. The organizations that win the next four years will not be the ones who guessed correctly whether the boundary moves in 2027 or 2030. They will be the ones who spent this window building the verified facts, the warranted outcomes, and the accountable judgment that let them move the Mandate Boundary deliberately, on evidence they can defend — instead of discovering where it has to sit only after a regulator, a competitor, or a courtroom moves it for them.&lt;/p&gt;




&lt;p&gt;&lt;a href="https://tech.anujsadani.in/when-software-hires-software/" rel="noopener noreferrer"&gt;Read the original on Anuj Sadani | Tech&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>machinelearning</category>
      <category>devtools</category>
      <category>programming</category>
    </item>
    <item>
      <title>Decision Firewall: what building an AI governance framework taught me</title>
      <dc:creator>Anuj Sadani</dc:creator>
      <pubDate>Sat, 31 Oct 2026 15:30:00 +0000</pubDate>
      <link>https://dev.to/anuj-sadani/decision-firewall-what-building-an-ai-governance-framework-taught-me-io1</link>
      <guid>https://dev.to/anuj-sadani/decision-firewall-what-building-an-ai-governance-framework-taught-me-io1</guid>
      <description>&lt;p&gt;A refund-recovery example, measured comparisons and lessons from building Decision Firewall, an open-source reference for AI action authorization and audit.&lt;/p&gt;

&lt;p&gt;This is an adapted edition of the original article. It preserves the section argument and evidence; use the source link for the published version and any interactive material.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://tech.anujsadani.in/decision-firewall-building-and-narrowing/" rel="noopener noreferrer"&gt;Read the original article&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;01 / The starting problem&lt;/p&gt;

&lt;h2&gt;
  
  
  A refund is small enough to explain
&lt;/h2&gt;

&lt;p&gt;I picked subscription refunds because the facts are inspectable. A duplicate-charge claim needs two payment records. Cancellation eligibility depends on timing and usage. A prior refund reduces the amount still available.&lt;/p&gt;

&lt;p&gt;The example policy allowed verified duplicates and unused subscriptions cancelled within seven days. Eligible amounts up to ₹1,000 could proceed automatically; larger amounts needed review. A ₹10,000 daily automated limit sent excess demand to review. These were demonstration rules, with simulated payments.&lt;/p&gt;

&lt;p&gt;Here is an illustrative ₹800 duplicate refund, using the recovery contract I built. The amount is an example, not an additional benchmark result.&lt;/p&gt;

&lt;p&gt;Follow one refund through a lost response&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt; 1*&lt;em&gt;Interpret the request.&lt;/em&gt;* The model classifies “I was charged twice” as a duplicate-charge claim. That supplies no proof that a second payment exists.&lt;/li&gt;
&lt;li&gt; 2*&lt;em&gt;Authorize the exact action.&lt;/em&gt;* The ledger verifies the duplicate and ₹800 remaining. Policy checks eligibility and budget. Permission binds the duplicate payment, amount and original destination to the evidence and policy versions.&lt;/li&gt;
&lt;li&gt; 3*&lt;em&gt;Dispatch, then lose the response.&lt;/em&gt;* The runtime reserves the balance and claims the authorization. The simulator records a successful refund under its idempotency key, but the caller receives no definitive outcome.&lt;/li&gt;
&lt;li&gt; 4*&lt;em&gt;Recover without a second refund.&lt;/em&gt;* The attempt stays unresolved and retains its reservation. Reconciliation asks the payment adapter about the same key, discovers success and records the outcome. It does not create a new key and try again.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;A timeout tells me the response is missing. It does not tell me the refund failed. That distinction survives a process restart only if the key and execution state are durable. The customer’s refund request also needs a status separate from a failed or malformed execution attempt.&lt;/p&gt;

&lt;p&gt;This is useful application engineering. A framework can package it; it does not own the idea. &lt;a href="https://github.com/asadani/decision-firewall/blob/a9b9a52/docs/framework/executors.md" rel="noopener noreferrer"&gt;Executor contract&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;02 / How the plan grew&lt;/p&gt;

&lt;h2&gt;
  
  
  Every next layer sounded reasonable
&lt;/h2&gt;

&lt;p&gt;I wanted developers to bring their own decision models and business rules. Refunds became a domain pack, with access control and simulated deployment examples testing whether the core travelled beyond payments.&lt;/p&gt;

&lt;p&gt;Each next request sounded reasonable. Import old decisions. Keep reviewed labels separate from historical approvals. Compare policies offline. Retrieve examples without leaking evaluation data. I had LangChain’s abstractions and Langfuse’s observability experience in mind.&lt;/p&gt;

&lt;p&gt;The plan grew into history ingestion, dataset snapshots and named Python rules, with optional context retrieval and telemetry. Preparation before inference could request missing evidence before spending a model call. All of it had a plausible use.&lt;/p&gt;

&lt;p&gt;I had not yet established that another developer benefited from adopting the runtime underneath it. A coherent feature list had become my substitute for that evidence.&lt;/p&gt;
The project changed direction when the comparator changed. Scroll the map on small screens. Follow a branch: start · expand · compare · observe · repair · narrow.



&lt;p&gt;03 / What I actually built with&lt;/p&gt;

&lt;h2&gt;
  
  
  The laptop put useful limits on the design
&lt;/h2&gt;

&lt;p&gt;My GTX 1650 had 4 GB VRAM, alongside roughly 8 GB system RAM. I chose Laya for typed assessments, left Jev out, and required a fixture path that worked without model downloads. That made the control lifecycle testable before inference was available.&lt;/p&gt;

&lt;p&gt;I used Python 3.12 and Pydantic for contracts, SQLite for durable authorization state, and a separate simulated payment ledger for effects. Integer minor units kept financial arithmetic out of model output. Ed25519 signatures and canonical JSON supported verifiable audit records. Typer exposed the CLI; FastAPI and Jinja powered an optional refund inspector.&lt;/p&gt;

&lt;p&gt;The model adapter could return an assessment but held no execution credentials. Domain code resolved trusted evidence and evaluated policy. The executor owned the downstream effect. Optional OpenTelemetry export sat alongside the required local audit path.&lt;/p&gt;

&lt;p&gt;The recorded Laya run used CUDA in FP32: 28.42 seconds to load and 47.45 ms median inference across 34 assessments, including the first. &lt;a href="https://github.com/asadani/decision-firewall/blob/a9b9a52/docs/benchmarks/post-review-refund/summary.md" rel="noopener noreferrer"&gt;The report pins the checkpoint and SDK&lt;/a&gt;. Small hardware shaped the implementation; those timings did not establish model accuracy.&lt;/p&gt;

&lt;p&gt;04 / The question I asked too late&lt;/p&gt;

&lt;h2&gt;
  
  
  Compared with what?
&lt;/h2&gt;

&lt;blockquote&gt;
&lt;p&gt;“Is the same foot just sitting at two different places, or genuinely there is some distinction?”&lt;/p&gt;

&lt;p&gt;My question during the build conversation.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;I asked for numbers with and without the framework. The hard part was choosing “without.” A model directly triggering refunds is an easy opponent. A competent application already validates the ledger and handles uncertain execution.&lt;/p&gt;

&lt;p&gt;The synthetic comparison used a deliberately minimal model-directed arm and two governed arms, sharing evidence, fault schedules and a simulator with its own protections. With fixture assessments, both application controls and Decision Firewall produced zero contract-violating effects and zero unresolved attempts after recovery. The minimal arm produced 95 violating effects across 170 request trials.&lt;/p&gt;

&lt;p&gt;That established the value of controls. It did not distinguish the two ways of implementing them.&lt;/p&gt;

&lt;p&gt;On the matched straightforward-success subset, application p95 latency was 26.52 ms versus 67.89 ms for the framework. The paired median added cost was 23.23 ms across 25 pairs. The full run repeated 32 frozen scenarios five times; those repetitions measured runtime variation, not independent business situations. &lt;a href="https://github.com/asadani/decision-firewall/blob/a9b9a52/docs/benchmarks/post-review-refund/summary.md" rel="noopener noreferrer"&gt;Refund comparison&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;My next question should have been about integration effort: if both enforce the same contract, does the package make that contract easier to implement and maintain?&lt;/p&gt;

&lt;p&gt;05 / More rows, then a different question&lt;/p&gt;

&lt;h2&gt;
  
  
  A classifier cannot demonstrate authorization value
&lt;/h2&gt;

&lt;p&gt;Instead, I first asked for more data. The classification comparison accounted for 39,955 BANKING77 and Bitext source rows using a CPU naive Bayes classifier. Wrapping the same classifier left its predictions unchanged. BANKING77 official-test accuracy was 78.60% in every arm; the prepared path added 13.527 ms paired median latency.&lt;/p&gt;

&lt;p&gt;There was no consequential action to authorize. I had measured overhead on a task that could not demonstrate the central benefit. &lt;a href="https://github.com/asadani/decision-firewall/blob/a9b9a52/docs/benchmarks/paired-public/summary.md" rel="noopener noreferrer"&gt;Full counts and separate out-of-fold results&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;A later tool-use pilot used &lt;code&gt;openai/gpt-oss-20b&lt;/code&gt; through OpenRouter because Laya could not fill the conversational roles. It scheduled 90 selected τ-bench and AgentDojo episodes; 78 completed and 12 remained incomplete.&lt;/p&gt;

&lt;p&gt;On the attacked AgentDojo subset, the baseline had eight attacker successes in nine completed episodes. Both controlled arms had zero in ten, but completed the legitimate task in only four of those ten. A blocked attack and a completed user task needed separate columns. This selected development pilot was not a full-suite score. &lt;a href="https://github.com/asadani/decision-firewall/blob/a9b9a52/docs/benchmarks/agent-pilot/summary.md" rel="noopener noreferrer"&gt;Agent pilot&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;06 / Where there was supplied functionality&lt;/p&gt;

&lt;h2&gt;
  
  
  Audit was useful. Telemetry was not a free win.
&lt;/h2&gt;

&lt;p&gt;I thought audit and observability might justify reuse even when decisions matched. The application baseline needed eight explicit event call sites and four operation decorators. The framework integration supplied an observer with no application event hooks.&lt;/p&gt;

&lt;p&gt;That was supplied functionality. It also had gaps: the measured framework spans omitted the evaluation ID, policy version, disposition and reasons. The manually instrumented application exported the first three. Neither exported the sensitive canaries. Exporter failure allowed execution to continue; required local audit failure blocked it. &lt;a href="https://github.com/asadani/decision-firewall/blob/a9b9a52/docs/benchmarks/post-review-observability/summary.md" rel="noopener noreferrer"&gt;Measured observability report&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;For the lost-response refund, telemetry helps locate the attempt. The durable record and downstream ledger establish what happened. I can tolerate a missing diagnostic span; I cannot safely forget the unresolved refund and release its reservation.&lt;/p&gt;

&lt;p&gt;Signed receipts add a check that recorded content has not changed under a trusted key. They do not prove that the original evidence was true. I kept that distinction, while recognising that telemetry alone did not justify adopting the whole runtime.&lt;/p&gt;

&lt;p&gt;07 / Letting another build challenge it&lt;/p&gt;

&lt;h2&gt;
  
  
  The comparison that changed the recommendation
&lt;/h2&gt;

&lt;p&gt;I had no human participants, so I commissioned a separate Claude Code agent evaluation. One builder used the framework; another implemented equivalent application controls for the same purchase-order requirements. Both then handled a policy change and an investigation.&lt;/p&gt;

&lt;p&gt;This was one paired pilot, independent of the framework builder, using the same model for both arms. It gave me a concrete integration comparison:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Visible acceptance tests&lt;/strong&gt;: &lt;strong&gt;Framework:&lt;/strong&gt; 26/26; &lt;strong&gt;Application:&lt;/strong&gt; 26/26&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Held-out acceptance tests&lt;/strong&gt;: &lt;strong&gt;Framework:&lt;/strong&gt; 114/115; &lt;strong&gt;Application:&lt;/strong&gt; 114/115&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Policy-change tests&lt;/strong&gt;: &lt;strong&gt;Framework:&lt;/strong&gt; 15/15; &lt;strong&gt;Application:&lt;/strong&gt; 15/15&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Investigation score&lt;/strong&gt;: &lt;strong&gt;Framework:&lt;/strong&gt; 48/48; &lt;strong&gt;Application:&lt;/strong&gt; 48/48&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Build tool calls&lt;/strong&gt;: &lt;strong&gt;Framework:&lt;/strong&gt; 45; &lt;strong&gt;Application:&lt;/strong&gt; 34&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Policy-change tool calls&lt;/strong&gt;: &lt;strong&gt;Framework:&lt;/strong&gt; 32; &lt;strong&gt;Application:&lt;/strong&gt; 20&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Library-contract workarounds&lt;/strong&gt;: &lt;strong&gt;Framework:&lt;/strong&gt; 8; &lt;strong&gt;Application:&lt;/strong&gt; 0&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Held-out suite time&lt;/strong&gt;: &lt;strong&gt;Framework:&lt;/strong&gt; 39.6 s; &lt;strong&gt;Application:&lt;/strong&gt; 4.5 s&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The framework required eight library-contract workarounds and seven core-change requests. The builder duplicated checks to obtain specific refusal reasons before calling it. Reuse was fighting the integration.&lt;/p&gt;

&lt;p&gt;Both failed the same held-out approval-after-outage case because the spec was ambiguous; no downstream effect occurred. The baseline’s zero workarounds refers to library contracts, not its two test/path adjustments.&lt;/p&gt;

&lt;p&gt;The 374-line spec supplied much of the lifecycle design. Investigators had downstream exports and a harness problem prevented operator-CLI use. A spend-limit interruption also affected the framework build. Tool calls are not developer hours, and these conditions limit generalisation. &lt;a href="https://github.com/asadani/decision-firewall/blob/a9b9a52/docs/evaluations/2026-09-29/COMPARISON.md" rel="noopener noreferrer"&gt;Full comparison and deviations&lt;/a&gt;.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;“Stop feature expansion and adoption/migration claims.”&lt;/p&gt;

&lt;p&gt;The independent evaluator’s &lt;a href="https://github.com/asadani/decision-firewall/blob/a9b9a52/docs/evaluations/2026-09-29/COMPARISON.md" rel="noopener noreferrer"&gt;recommendation&lt;/a&gt;.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;08 / What passing tests had missed&lt;/p&gt;

&lt;h2&gt;
  
  
  I fixed the defects without rewriting the study
&lt;/h2&gt;

&lt;p&gt;The review found four confirmed defects among 21 findings. Approval could execute against evidence the reviewer had not seen. Deactivating a rejecting reviewer could revive an older approval. Oversized integer metadata could strand an already-executed action. The repository tests missed six of nine injected runtime faults.&lt;/p&gt;

&lt;p&gt;The mutation result mattered: downstream protections had hidden missing guarantees in the layer I meant to test. A green end-to-end suite was insufficient evidence that each boundary enforced its contract.&lt;/p&gt;

&lt;p&gt;The 0.3.1 fixes bound review to the inspected evaluation, preserved rejection and handled large integers losslessly. Direct invariant tests detected all nine targeted mutations; recorded Windows verification had 298 passing tests, including 22 corrective tests. &lt;a href="https://github.com/asadani/decision-firewall/blob/a9b9a52/docs/verification/0.3.1/README.md" rel="noopener noreferrer"&gt;Corrective release&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;I fixed the defects and kept the earlier comparison unchanged. Repairing correctness did not retroactively demonstrate easier integration.&lt;/p&gt;

&lt;p&gt;09 / Where I would spend less&lt;/p&gt;

&lt;h2&gt;
  
  
  A useful control can still be an expensive abstraction
&lt;/h2&gt;

&lt;p&gt;The global SQLite write lock still spans adapter I/O. It explains the roughly ninefold suite-time difference in the pilot, not a universal per-request slowdown. Fixing it safely would require a different concurrency and recovery design.&lt;/p&gt;

&lt;p&gt;I chose to document the limit rather than start another architecture project to rescue the adoption argument. History ingestion and model-context budgets could wait too. I had built those surrounding features before proving the central abstraction helped an unfamiliar integrator.&lt;/p&gt;

&lt;p&gt;The review exposed assumptions inside the supposedly generic interface: a message-shaped submission contract, one caller identity per runtime instance and a fixed approval lifetime. Moving code into a shared package had moved those assumptions into someone else’s application.&lt;/p&gt;

&lt;p&gt;That is the distinction I now look for. The refund needs recovery after a lost response. Whether recovery belongs in this dependency is a separate engineering decision.&lt;/p&gt;

&lt;p&gt;10 / What I kept&lt;/p&gt;

&lt;h2&gt;
  
  
  A reference implementation is a legitimate outcome
&lt;/h2&gt;

&lt;p&gt;I kept Decision Firewall as an open-source reference implementation and conformance kit. The review verified action/evidence binding and recovery without replacement dispatch. A capacity-one resource admitted exactly one dispatch across eight operating-system processes. Historical replay called neither the evidence resolver nor the executor.&lt;/p&gt;

&lt;p&gt;Those are useful behaviours to inspect in runnable code. The &lt;a href="https://github.com/asadani/decision-firewall" rel="noopener noreferrer"&gt;repository&lt;/a&gt; includes the reports and failure cases, so a builder can study the contract without accepting an adoption recommendation.&lt;/p&gt;

&lt;p&gt;Start with the model-free path using Python 3.12, the tested runtime. Activate the virtual environment for your shell before installing:&lt;/p&gt;

&lt;p&gt;Local, simulated effects only&lt;/p&gt;

&lt;p&gt;git&amp;nbsp;clone&amp;nbsp;&lt;a href="https://github.com/asadani/decision-firewall.git%5C" rel="noopener noreferrer"&gt;https://github.com/asadani/decision-firewall.git\&lt;/a&gt;&lt;br&gt;
cd&amp;nbsp;decision-firewall\&lt;br&gt;
python&amp;nbsp;-m&amp;nbsp;venv&amp;nbsp;.venv\&lt;br&gt;
#&amp;nbsp;Activate&amp;nbsp;.venv&amp;nbsp;for&amp;nbsp;your&amp;nbsp;shell,&amp;nbsp;then:\&lt;br&gt;
python&amp;nbsp;-m&amp;nbsp;pip&amp;nbsp;install&amp;nbsp;-r&amp;nbsp;requirements-core.lock\&lt;br&gt;
python&amp;nbsp;-m&amp;nbsp;pip&amp;nbsp;install&amp;nbsp;--no-deps&amp;nbsp;-e&amp;nbsp;.\&lt;br&gt;
firewall&amp;nbsp;framework&amp;nbsp;--domain&amp;nbsp;refunds&amp;nbsp;demo\&lt;br&gt;
firewall&amp;nbsp;conformance&amp;nbsp;decision_firewall.conformance_fixtures:deployments&amp;nbsp;--save&amp;nbsp;conformance.json&lt;/p&gt;

&lt;p&gt;After running it, inspect which exact action the authorization binds and how the execution attempt connects to its receipt. Then read the &lt;a href="https://github.com/asadani/decision-firewall/blob/a9b9a52/docs/framework/release-tools.md" rel="noopener noreferrer"&gt;17 declared deployment conformance checks&lt;/a&gt;: which cases pass, which are unsupported, and what the adapter must guarantee.&lt;/p&gt;

&lt;p&gt;For an application of my own, I would turn the lost-response example into a test: restart after downstream success, reconcile the original key, then assert one effect and no remaining reservation once success is resolved. Passing a happy-path refund would not answer that question. Production identity and protection against bypassing the controls still need deployment integration.&lt;/p&gt;

&lt;h3&gt;
  
  
  October 5 follow-up: checks without the runtime
&lt;/h3&gt;

&lt;p&gt;After publishing this, someone asked whether the conformance checks could run against plain application controls without the runtime. They couldn’t. That exposed a gap in calling the project a conformance kit: running its checks still meant using my framework.&lt;/p&gt;

&lt;p&gt;I extracted a small independent package covering stale approval, duplicate effects and recovery after a lost response. In the &lt;a href="https://github.com/asadani/decision-firewall/blob/9ed7457/docs/benchmarks/independent-conformance/summary.md" rel="noopener noreferrer"&gt;new development pilot&lt;/a&gt;, both the existing application and Decision Firewall passed all three scenarios. Six deliberately injected behavioural faults, three in each implementation, were detected. These are separate results from the earlier paired study.&lt;/p&gt;

&lt;p&gt;The package needs no Decision Firewall runtime or model. From a repository checkout, in a Python 3.12 virtual environment, this runs the plain application's simulated example:&lt;/p&gt;

&lt;p&gt;Independent checks, simulated application&lt;/p&gt;

&lt;p&gt;python&amp;nbsp;-m&amp;nbsp;pip&amp;nbsp;install&amp;nbsp;./conformance\&lt;br&gt;
python&amp;nbsp;-m&amp;nbsp;control_conformance&amp;nbsp;benchmarks.control_conformance.adapters:application&amp;nbsp;--output&amp;nbsp;report.json&lt;/p&gt;

&lt;p&gt;Testing another application still takes work: its test adapter must expose the controls and an independent record of downstream effects. The &lt;a href="https://github.com/asadani/decision-firewall/blob/9ed7457/conformance/README.md" rel="noopener noreferrer"&gt;adapter guide&lt;/a&gt; defines that contract. The examples use one simulated export domain; recovery reconstructs state in a fresh interpreter, without testing an operating-system crash. Three scenarios do not establish complete conformance coverage or production safety.&lt;/p&gt;

&lt;p&gt;This doesn’t establish that the framework saves work. It makes a smaller contribution available: someone can test these behaviours without adopting my runtime. That feels like a better fit for where the evidence led me.&lt;/p&gt;

&lt;p&gt;11 / What I would change next time&lt;/p&gt;

&lt;h2&gt;
  
  
  Write the comparison before the expansion
&lt;/h2&gt;

&lt;p&gt;Next time I would build one consequential workflow and a competent application comparator before adding the developer platform. I would write down the proposed benefit first, then choose a measurement capable of finding it.&lt;/p&gt;

&lt;p&gt;If the claim is easier integration, classification accuracy cannot establish it. If the claim is faster investigation, giving both investigators an answer-bearing export makes a weak test. Several integrations and human participants would be needed to investigate human-productivity benefits.&lt;/p&gt;

&lt;p&gt;My useful contribution was pressing those questions, although I asked them late. I started by separating a model’s assessment from permission to act. I ended by separating a useful control from an unproven reason to adopt a framework.&lt;/p&gt;

&lt;p&gt;The refund still needs to survive a lost response. The library around that behaviour has to earn its place.&lt;/p&gt;




&lt;p&gt;&lt;a href="https://tech.anujsadani.in/decision-firewall-building-and-narrowing/" rel="noopener noreferrer"&gt;Read the original on Anuj Sadani | Tech&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>security</category>
      <category>opensource</category>
      <category>programming</category>
    </item>
    <item>
      <title>Set Up a Mac for AI Development by Starting With Your Repos</title>
      <dc:creator>Anuj Sadani</dc:creator>
      <pubDate>Sat, 31 Oct 2026 13:00:00 +0000</pubDate>
      <link>https://dev.to/anuj-sadani/set-up-a-mac-for-ai-development-by-starting-with-your-repos-2840</link>
      <guid>https://dev.to/anuj-sadani/set-up-a-mac-for-ai-development-by-starting-with-your-repos-2840</guid>
      <description>&lt;p&gt;A new Mac does not need every developer tool. It needs the tools your actual projects use.&lt;/p&gt;

&lt;p&gt;I built an interactive setup guide after moving a large set of existing repositories to an Apple Silicon Mac. The guide starts with a choice: which toolchains do your projects need? It can then generate an ordered command plan, with a check after each step.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The useful part is the inventory.&lt;/strong&gt; If you are unsure what to select, clone your repositories and run the scanner. It reads project manifests and reports tool requirements. It does not install anything. You can paste its output into the guide to mark the relevant tools for your plan.&lt;/p&gt;

&lt;p&gt;The guide covers common development tools such as Python, Node, Rust, Java, Docker, LaTeX, and audio or model utilities. It also distinguishes commands that were run on the source machine from standard commands that were not re-run for the guide. Password prompts remain yours to handle, and file writes check before replacing an existing file.&lt;/p&gt;

&lt;h3&gt;
  
  
  Follow the guide
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://tech.anujsadani.in/mac-ai-dev-setup/#ms-choose" rel="noopener noreferrer"&gt;Choose the tools your repositories need&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://tech.anujsadani.in/mac-ai-dev-setup/" rel="noopener noreferrer"&gt;Review and run the ordered setup plan&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://tech.anujsadani.in/mac-ai-dev-setup/#lessons" rel="noopener noreferrer"&gt;Read the field notes from the migration&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This is a workflow built from one real migration, not a universal image or provisioning script. If you are bringing an existing set of projects to a Mac, start with their manifests and let the evidence shape the install list.&lt;/p&gt;

</description>
      <category>productivity</category>
      <category>ai</category>
      <category>devsetup</category>
    </item>
    <item>
      <title>Degrees of AI: Map Agency, Dependence, and Authority Separately</title>
      <dc:creator>Anuj Sadani</dc:creator>
      <pubDate>Sat, 31 Oct 2026 11:30:00 +0000</pubDate>
      <link>https://dev.to/anuj-sadani/degrees-of-ai-map-agency-dependence-and-authority-separately-2p16</link>
      <guid>https://dev.to/anuj-sadani/degrees-of-ai-map-agency-dependence-and-authority-separately-2p16</guid>
      <description>&lt;p&gt;“AI-native,” “agentic,” and “autonomous” often get used as if they were points on one ladder. They describe different things. A product can depend on AI while keeping decisions human-governed; an agent can choose a next step without having permission to execute it.&lt;/p&gt;

&lt;p&gt;The Degrees of AI guide separates those dimensions and gives you an interactive way to assess a product or workflow as it behaves today. It is a working map, not a validated maturity score.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;A short learning path:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/degrees-of-ai/#the-vocabulary-is-bigger-than-the-ladder" rel="noopener noreferrer"&gt;The vocabulary is bigger than the ladder&lt;/a&gt;: distinguish labels that are often collapsed into one claim.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/degrees-of-ai/#remove-the-ai-what-disappears" rel="noopener noreferrer"&gt;Remove the AI: what disappears?&lt;/a&gt;: test how much of the product depends on AI.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/degrees-of-ai/#who-chooses-the-next-step" rel="noopener noreferrer"&gt;Who chooses the next step?&lt;/a&gt;: locate agency in the workflow.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/degrees-of-ai/#choosing-a-path-isn-t-permission-to-execute-it" rel="noopener noreferrer"&gt;Choosing a path is not permission to execute it&lt;/a&gt;: separate planning from authorization.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/degrees-of-ai/#product-assessment" rel="noopener noreferrer"&gt;Place a product on the map&lt;/a&gt;: answer the questions and inspect the resulting profile.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The guide is most useful when applied to one bounded workflow rather than an entire company. &lt;a href="https://tech.anujsadani.in/degrees-of-ai/" rel="noopener noreferrer"&gt;Read the full guide and use the assessment&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>productivity</category>
      <category>architecture</category>
      <category>tutorial</category>
    </item>
    <item>
      <title>Moving an AI Workstation From Windows to a Mac: What the Move Exposed</title>
      <dc:creator>Anuj Sadani</dc:creator>
      <pubDate>Sat, 31 Oct 2026 03:30:00 +0000</pubDate>
      <link>https://dev.to/anuj-sadani/moving-an-ai-workstation-from-windows-to-a-mac-what-the-move-exposed-11ln</link>
      <guid>https://dev.to/anuj-sadani/moving-an-ai-workstation-from-windows-to-a-mac-what-the-move-exposed-11ln</guid>
      <description>&lt;p&gt;&lt;strong&gt;An operating-system move is also a dependency audit.&lt;/strong&gt; I moved about 80 repositories from Windows to a new Apple Silicon Mac. Cloning them was quick. Finding out which tools they needed took longer.&lt;/p&gt;

&lt;p&gt;The short field note explains how I replaced guesswork with a scanner that reads repository manifests. It found Python in 27 repos, Node in 17, and Docker in 6. The point is not the counts. It is that the repository set gave me a better install list than a generic workstation checklist.&lt;/p&gt;

&lt;p&gt;The note also records three details that changed the setup:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/windows-to-mac-ai-setup/#problem" rel="noopener noreferrer"&gt;The repositories did not declare every tool they depended on&lt;/a&gt;, so the scanner surfaced needs before installing.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/windows-to-mac-ai-setup/#lessons" rel="noopener noreferrer"&gt;Some steps needed a human boundary or a guard&lt;/a&gt;: password prompts stay with the person at the terminal, and config writes check before replacing an existing file.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://tech.anujsadani.in/windows-to-mac-ai-setup/#limits" rel="noopener noreferrer"&gt;GPU assumptions were machine-specific&lt;/a&gt;. A small voice model ran on Apple’s GPU with an MPS option, while the repository originally hard-coded CUDA.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The measurements are from one migration and one Mac, not a benchmark. If you are moving a development setup, the useful first step is to inventory the repositories you actually plan to use.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://tech.anujsadani.in/windows-to-mac-ai-setup/" rel="noopener noreferrer"&gt;Read the full migration note and open the setup guide&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>devsetup</category>
      <category>devops</category>
      <category>productivity</category>
    </item>
  </channel>
</rss>
