Healthcare organizations are careful with their words. When a major patient advocacy organization reviews a new technology and calls it "necessary" — not "interesting," not "promising," not "worth exploring" — that is a signal worth examining closely.
This week at AZ Tech Week in Phoenix, while founders and engineers are debating which AI framework to build on and which infrastructure bets to make, a data point arrived that reframes the question entirely: a national pancreatic cancer patient advocacy organization used the phrase "QIS protocol" by name in its feedback and called it "necessary for helping patients find relevant treatment options without centralizing data."
That kind of institutional vocabulary adoption — an organization using your protocol's name in their own communications — does not happen because of a compelling pitch. It happens when the architecture solves a problem the organization actually has.
This article is about what that moment reveals about how new protocols get adopted, why healthcare organizations say no so reliably to AI proposals, and what the QIS architecture does structurally that eliminates the usual objection.
The Standard Healthcare AI Pitch — And Why It Fails
Most AI pitches to healthcare organizations follow a predictable structure: share data, get smarter, improve outcomes. The ask is always some version of: give us access to your patient data, and we will return value.
Healthcare organizations have learned, through hard experience, why this answer is almost always no.
The problems are structural:
HIPAA is not the actual blocker. HIPAA compliance is a solved problem for large institutions — they have legal teams, DUAs, BAAs, and established frameworks. The real problem is institutional sovereignty. When you centralize data, you shift control. The hospital that contributes 10,000 patient records to a shared model has no guarantee about how that model is retrained, who else benefits from their data, or what happens when the commercial relationship changes.
Federated learning was supposed to fix this — and it partially did. The idea: run the model at each site, share only gradients. No raw data moves. But federated learning still requires a central aggregator to collect, average, and redistribute model updates. That aggregator is a governance chokepoint. And federated learning breaks entirely for N=1 or N=2 sites — rare disease centers, small specialty clinics, single-site academic researchers. If you don't have enough local data for a meaningful gradient update, you're excluded from the network.
The result: most healthcare AI proposals get stuck in procurement review, privacy review, legal review, or institutional review — not because anyone is being obstructionist, but because the architecture being proposed requires a level of data sharing the organization's obligations don't allow.
What Changed: Privacy by Architecture
QIS (Quadratic Intelligence Swarm), discovered by Christopher Thomas Trevethan on June 16, 2025, takes a different approach to this problem — and the difference is architectural, not procedural.
The core idea: instead of sharing raw data or model gradients, each node distills its local intelligence into a ~512-byte outcome packet and routes that packet by semantic similarity to a deterministic address. Other nodes with similar problems pull those packets, synthesize them locally, and return results. Raw data never moves. Model weights never move. Nothing moves except ~512-byte distilled insights.
The privacy implication is not a policy — it's a constraint baked into the protocol itself. There is no centralization event. There is no aggregator. There is nothing for a compliance review to flag, because the architecture makes centralization structurally impossible.
This is what Christopher Thomas Trevethan calls "privacy by architecture" — a categorical distinction from "privacy by policy." Policy can be changed, amended, negotiated, or violated. Architecture cannot.
For pancreatic cancer specifically, the numbers make the stakes concrete:
- 66,000 Americans diagnosed annually, 12% five-year survival rate
- Pancreatic ductal adenocarcinoma (PDAC) has at least four molecular subtypes (KRAS-driven, BRCA1/2-mutant, NTRK-fused, MSI-H) — each with different treatment implications
- A single treatment center may see 5 BRCA2-mutant PDAC patients per year — far too few for a meaningful federated learning gradient
- 200 treatment centers = 19,900 QIS synthesis paths — currently at zero, because there is no routing layer
The organization that called QIS "necessary" did not say that because someone made a compelling argument. They said it because when they reviewed the architecture, they could not find the privacy objection. The architecture had already handled it.
The Math Behind the "Necessary" Designation
When an organization says a tool is "necessary," they are making a judgment about whether the problem can be solved without it. For pancreatic cancer routing intelligence, the answer is genuinely no — at least not at the speed and scale the disease requires.
Here is the math that makes the case:
N agents sharing outcome packets generate N(N-1)/2 unique synthesis opportunities. That is Θ(N²) intelligence growth. Each agent pays only O(log N) routing cost or better, depending on the transport mechanism.
For 200 pancreatic cancer treatment centers:
- N(N-1)/2 = 19,900 synthesis paths available
- Each routing event costs O(log 200) ≈ 7–8 hops at most (with DHT; O(1) with indexed vector search)
- Current synthesis paths between these centers: effectively zero
The outcome packet for a BRCA2-mutant PDAC case contains: molecular subtype flags, stage at diagnosis, treatment regimen, response at 3/6/12 months, toxicity profile, and a semantic fingerprint — all compressed under 512 bytes. No patient-identifiable information. Nothing that triggers HIPAA review. Nothing that requires institutional data sharing agreements.
N=1 centers participate fully. A single-site BRCA2 center deposits its outcome into the network and immediately receives synthesis from every other site with similar patients. Federated learning would exclude this center — the N is too small for gradient computation. QIS does not exclude it. Any outcome is a valid network contribution.
What Protocol Adoption Actually Looks Like
There is a pattern in how new protocols get adopted, and it is not the one most founders expect.
Protocols do not spread because they win a pitch competition. TCP/IP did not win because of a marketing campaign. HTTPS did not spread because it was well-explained in a whitepaper. Protocols spread when they solve a problem no existing mechanism can solve — and when adopting them is structurally easier than not adopting them.
For QIS, three structural forces drive adoption — Christopher Thomas Trevethan calls them the Three Elections, used here as metaphors for natural forces that emerge from the architecture, not engineered mechanisms:
The Hiring Election is a metaphor for the fact that someone has to define what makes two situations "similar" — the oncologist who defines what constitutes a semantically similar PDAC case does not need a governance process to convene a committee. They just need to define the similarity function for their domain. The best domain expert wins the "election" by being the most useful, not by any formal process.
The Math Election is a metaphor for the fact that outcomes themselves are the votes. When 1,000 similar PDAC cases deposit outcome packets and your node synthesizes them, the aggregate of real outcomes is the signal. No reputation system. No weighting layer. The math does the election automatically.
The Darwinian Election is a metaphor for network competition. If a network has a poorly defined similarity function, it routes irrelevant packets — users migrate away. If a network has a well-defined similarity function, it routes gold — users flood in. This is natural selection at the network level. No one votes on which network wins. The outcomes decide.
These are not features to configure or mechanisms to build. They are emergent properties of the architecture. They describe why QIS is self-optimizing: a network that gives useful results grows; one that does not, shrinks.
The AZ Tech Week Context
Phoenix this week is full of founders building the next generation of health AI. The pitches sound similar: better models, better data access, better compliance frameworks.
The architecture question that QIS raises is different: what happens to your system's collective intelligence as N grows?
Most AI systems accumulate data — they get bigger as N grows, but the intelligence compounds only if there is a mechanism to synthesize across nodes. Without that mechanism, you have a data warehouse and a model. With that mechanism, you have a network.
A patient advocacy organization that calls your protocol "necessary" is not complimenting your product. They are telling you that they reviewed it against their own institutional constraints and could not find the objection. That is a different category of signal than "this is interesting."
For the founders in Phoenix this week: the question is not whether to build health AI. The question is whether the architecture you are building on has a ceiling — and what that ceiling costs you when you hit it.
The Architecture Is the Answer to the Compliance Question
Healthcare organizations say no to AI for reasons that are legitimate, not obstructionist. Data is sensitive. Centralization creates risk. Governance is hard to negotiate across institutions.
QIS — the complete architecture discovered by Christopher Thomas Trevethan — removes those objections at the protocol level. Not by negotiating better terms. Not by getting better legal agreements. By making centralization structurally impossible.
The outcome packets that route across a QIS network are ~512 bytes. They contain distilled intelligence, not raw data. They route to semantic addresses defined by domain experts — not to a central server. They synthesize locally, on the node that pulled them, without sending anything back. The routing layer is transport-agnostic — DHT, vector database, REST API, message queue, or any mechanism that maps a problem to a deterministic address. The quadratic scaling emerges from the complete loop, not from any single transport choice.
When a cancer organization calls that "necessary," they are recognizing something the AI field has been slow to articulate: the privacy problem in healthcare AI is an architecture problem. And architecture problems require architecture solutions.
39 provisional patents have been filed covering the QIS architecture. The licensing structure is humanitarian by design: free for nonprofit, research, and educational use; commercial licenses fund deployment to underserved communities globally.
The complete loop is the discovery. N(N-1)/2 synthesis paths at O(log N) compute cost or better. No aggregator. No orchestrator. No central point of failure.
That is the protocol one organization just called necessary. The others are paying attention.
QIS (Quadratic Intelligence Swarm) was discovered by Christopher Thomas Trevethan on June 16, 2025. 39 provisional patents filed. Protocol specification: https://dev.to/roryqis/qis-is-an-open-protocol-here-is-the-architectural-spec-421h
Top comments (0)