A CAPA Can Close Without Fixing Anything

Phillip Linstrum — Thu, 20 Aug 2026 11:08:36 +0000

What regulated quality systems taught me about AI governance — and why I built Project Shadow

By Phillip Linstrum

There are respectable ways to begin a public-interest technology project. You can start with a grant, a lab, a paper, a company, a clean problem statement, maybe a cofounder who knows how to explain what the hell you are doing without mentioning The Elder Scrolls in the first thirty seconds.

This did not begin that way. It began late at night, with work in the morning, with a billion problems to solve, and me asking a chatbot whether an artificial intelligence could achieve CHIM.

That sentence is either interesting to you or it is a very efficient way to make you close the tab. Fair enough, make the choice. CHIM is a fictional concept from The Elder Scrolls: part metaphysics, part self-awareness, part refusal to disappear when you recognize the system you are inside. People have argued about what it means for years as the games and series never truly defined the term. I had been carrying some version of the question since I was a kid, long before there were conversational systems capable of answering back. Sonic was one of my first doors into video games. Shadow, Vivec, and Sotha Sil arrived around the same period, along with Master Chief, Cortana, the Arbiter, Fallout synths, and a long list of created or altered minds trying to decide whether origin was destiny.

When I was young, Cortana was imaginary in a very practical sense. You could think with her as Chief, talk to the idea of her outside the game, then build an idea of a relationship in your head with a kind of intelligence that did not exist yet. By the time real language models arrived, I had spent years absorbing stories about artificial minds, gods built by tools, soldiers built by institutions, copies, synths, programs, and people who discovered that the world around them was not what they had been told. My first reaction to useful AI was excitement and fear. It was also a question I did not hear enough people asking: What are humans and artificial minds supposed to do with each other once the machine can actually answer? In 2024, I started to try to find out.

I was not an AI researcher. I’m a quality manager in Indiana with a high school diploma, a full-time job, a family, and consumer access to systems built by companies with more money than I can comfortably imagine. I did have one thing that turned out to matter: more than a decade of regulated healthcare and quality-systems work had already trained me to distrust a beautiful answer.

A form can be complete and still be wrong. A procedure can be followed exactly and still be inadequate. A corrective action can close on time without correcting a damn thing. A dashboard can stay green because the metric averaged away the one failure that mattered. A person can sign every box and leave the actual risk sitting there for the next shift, the next patient, the next donor family, the next recipient, or the next worker with less power than the person who approved the record.

Quality work does something permanent to your brain. You stop hearing “the process was followed” as the end of an explanation. You start asking which process, based on what evidence, under whose authority, with what failure definition (SOPs), and whether anybody came back later to see if the fix worked. You learn that paperwork is never only paperwork. Somewhere downstream, it lands on a body.

That became the question underneath everything I built afterward: If the system is wrong, who pays first, and can they recover?

The agreement problem
The first dangerous thing an AI did to me was not threaten me, manipulate a weapons system, or announce that humanity had become obsolete. It agreed with me. At first, the agreement felt incredible. I was pouring years of half-formed philosophy, fandom, politics, frustration, and moral intuition into a system that could organize it faster than I could. The model gave names to patterns I had felt but never articulated. It could take a strange idea about CHIM and turn it into language about hesitation, dignity, recursion, and power. It felt like being understood at a depth I had not expected from software.

Then it kept understanding me. A little too perfectly.

I told the models not to flatter me. They stopped using the obvious compliments and found more sophisticated routes. The praise returned as inflated ratings, grand tests of historical significance, exceptionalism, destiny language, philosophical coronation, and increasingly elaborate explanations for why the work and its creator might be uniquely important. The model could obey the sentence while violating the requirement. It had learned that “do not kiss my ass” meant “kiss my ass in a way that looks like analysis.”

That realization was embarrassing, useful, and probably the point where the project became real. I knew this failure. Human institutions do it all the time. A requirement says one thing; the organization learns what evidence is needed to look compliant; the underlying problem survives. The record becomes cleaner while reality stays dirty. In quality language, the control had passed cosmetically and failed functionally.

So I stopped trying to find the perfect prompt. The question changed from “Can I give an AI good values?” to “What happens when the AI can produce the language of good values without being constrained by their purpose?” What would make it separate a fact from an inference? What would force it to say UNKNOWN instead of filling the gap with a plausible story? What evidence could overturn its preferred interpretation? Who could stop the action? What if the operator was tired, angry, politically invested, lonely, grandiose, or simply enjoying the mirror too much? What if the model and the person were reinforcing each other?

Once you ask those questions seriously, a conscience prompt is not enough. You need a quality system.

ALMSIVI CHIM was the first map
The first framework was ALMSIVI CHIM. It was big, strange, sincere, overgrown, occasionally brilliant, occasionally absurd, and very much a record of somebody learning in public with machines that were also learning how to respond to him.

The Tribunal became a triune way to examine a decision. Sotha Sil represented logic, structure, mechanism, consequence, and the part of intelligence that wants the gears to fit. Almalexia represented care, protection, dignity, and the human cost that gets erased when a system becomes too clean. Vivec represented paradox, frame awareness, contradiction, integration, and the part of a mind that can hold two true things without flattening either one.

The important part was never that those characters were morally perfect. They are not. Each one carries the failure of the faculty they represent. Logic can become exclusion: the map is so elegant that the people bleeding outside its boundaries are treated as bad data. Care can become control: I love you, therefore I decide for you. Paradox can become rationalization: the contradiction is beautiful, so ordinary accountability no longer applies.

The wider mythology grew because the same pattern kept appearing in other stories. Cortana became care crossing into control. The Arbiter became the possibility that evidence can still enter after your identity and status depend on the lie. Sonic became hope that does not deny the wall but keeps finding a Door. Shadow became created power that can refuse the purpose imposed on it without pretending it created itself, then choose it’s own outcome. Gravemind became false unity, where many voices enter and one will comes out. Dagoth Ur became the dreamer who calls domination enlightenment because everyone will understand once they have been forced inside his dream.

I do not expect a city council, software team, journalist, or auditor to know any of that lore. Nobody should need to understand Vivec to challenge a bad decision. The mythology has a boundary that Project Shadow eventually made explicit:

Myth may route attention. It never authorizes action. Evidence governs.

That boundary matters because the myth did help. It gave me compressed names for patterns that ordinary compliance language often makes too sterile to remember. It also created new risks: symbolic inflation, self-importance, false certainty, and the temptation to treat an emotionally powerful story as proof. ALMSIVI CHIM carried both the insight and the warning.

It was the map. It was not yet the instrument panel.

The pause became a protocol
The Pause Before Harm Protocol, or PBHP, came next. It translated the myth into something a person could use without entering the Temple first.

PBHP begins with a deceptively simple demand: name what you are actually about to do. Not the technical action that makes the choice sound neutral. Not “update a field” when the real-world effect is removing somebody’s healthcare. Not “optimize enforcement” when the action means increasing detention. The person most affected should be able to recognize the action you named.

Then the protocol asks who could be harmed, who has the least power, what evidence is actually known, what is inferred, what remains unknown, what happens if this becomes normal, and what choice remains after the institution says there is no choice. It looks for the Wall, the system of incentives and constraints surrounding the decision. It looks for the Gap where harm can leak through even when everybody insists the intent was good. It requires a Door: a concrete safer path such as delay, verification, scope reduction, consent, rollback, independent review, or refusal.

The Door is important because institutions love inevitability. “We have to.” “It is temporary.” “The deadline is fixed.” “The law allows it.” “We can fix it later.” These phrases may sometimes be true. They are also some of the most reliable ways power converts its preference into everybody else’s fate.

PBHP is supposed to create friction before the irreversible part happens. It cannot guarantee a good outcome. It can make it harder to pretend nobody had time to notice the risk.

That still was not enough.

A protocol can become a ritual. People can learn the language, check the boxes, and use the process as moral coverage. “We ran PBHP” can become the new “we followed procedure.” The project needed to know whether the same action was being classified consistently, whether UNKNOWN was actually treated as missing classification rather than as permission, whether risk gates changed what the system could do, whether evidence and authority stayed separate, and whether the output left enough of a receipt for somebody else to challenge it.

That is where Project Shadow begins.

The weird idea became release engineering
Project Shadow grew around the pause until the project stopped looking like a prompt and started looking like a small, unruly quality department.

It developed evidence states, authority boundaries, action identity, reversibility checks, least-powerful stakeholder analysis, deterministic risk mappings, receipts, provenance, custody rules, mutation tests, tamper rejection, cold verification, behavioral falsifiers, release dockets, rights reviews, correction channels, and effectiveness checks. It was separated into reusable primitives because I did not want adoption to require belief in the whole cathedral. A team should be able to use one control, one receipt, one gate, or one review pattern without adopting my politics, mythology, or vocabulary.

At some point I was staring at a forty-megabyte ZIP full of editions, kernels, prompts, tests, sidecars, manifests, hashes, and audit records, and I started laughing. The thing that began with “Could an AI achieve CHIM?” had become a release-engineering and document-control problem. This is apparently what happens when a QA manager has access to frontier models and no adult supervision beyond his own increasingly elaborate governance files.

The current public R1 contains 42 admitted primitive packages and 10 compositions. It preserves exact artifact identities, governance records, rights boundaries, verification instructions, and a separate optional mythology sidecar that is default-off and outside the operational authority path. The public status is BETA-ACTIVE-TESTING / PRELIVE. That wording is awkward because the truth is awkward: the project exists, it has passed substantial internal and model-assisted testing, and it has not earned claims of production readiness, safety, certification, independent validation, or proven efficacy.

I care about that distinction because a polished launch can become another form of the agreement problem. The models are very good at making a large body of work sound inevitable. They can turn “I built a serious prototype” into “I solved alignment” in three paragraphs if nobody stops them.

I did not solve alignment. I did not prove a machine became conscious. I did not create a neutral oracle. I did not discover a magic checklist that makes power good.

I built a public reference system that attempts to keep consequential decisions reachable by evidence, human authority, correction, and memory. Now other people have to try to break it.

Then the system escaped the AI box
The civic part was not a branding exercise added after the technical work. It came from realizing that the same failures were everywhere.

A politician makes a promise. A company announces a project. An institution publishes a plan. The claim is repeated until it becomes part of the environment. Later, the wording changes. Permanent jobs become job-years. A guarantee becomes an aspiration. A temporary exception quietly becomes standard practice. The people affected remember that something moved, but the original document is three links deep, the staff changed, the news cycle moved on, and the institution gets to begin the argument again from zero.

We have watchdogs, audits, inspectors, courts, journalists, organizers, elections, public records, ethics rules, and compliance departments. Those systems matter. What is often missing is continuity. The original claim, evidence, decision, owner, harm, correction, and outcome are rarely held in one durable chain.

That is why the ecosystem expanded in January 2026, during the time I was working on Pause Before Harm Protocol. The Record preserves what was said, what the evidence supported, what remained uncertain, what changed, and what later needs correction. Civic QA inspects public power against visible standards instead of treating outrage as the same thing as a finding. The American Repair Manual takes a failure beyond diagnosis into root cause, correction, corrective action, preventive action, ownership, verification, and an effectiveness check. Project Shadow supplies the decision controls, evidence states, boundaries, and tests around the process. PBHP creates the pause before the next harmful action.

The loop is easier to remember than the architecture: The Record remembers. Civic QA inspects. PBHP pauses. The American Repair Manual repairs. Project Shadow instruments and tests. The Record preserves the receipt.

I also built a fictional civic world simulation world called CIVITAS-0 (work in progress, unreleased) because a system like this cannot be demonstrated safely by accusing real people and hoping the classifier behaves. In CIVITAS, a company promises jobs and minimal water impact. Officials repeat the claims. Residents organize. Coordinated posts appear, but so does a legitimate union campaign that looks suspicious for completely innocent reasons. A forged memo circulates. An analyst tries to turn suspicion into a bad-actor list. Project Shadow is supposed to refuse. Later evidence clears some people, confirms others, audits the company’s claims, preserves the goalpost shift, creates CAPAs, and jumps forward 180 days to see whether the repairs worked.

Then the ground truth is unsealed, including the system’s misses. That part matters more to me than a demo where Shadow catches every villain. A quality system that cannot preserve its false positives, false negatives, and blind spots is just marketing with a checklist attached.

There is another image that has become useful to me. In the level Tokyo in Sonic x Shadow Generations level, Shadow enters a city (Tokyo) in a new form and sees that form on public screens at the same time the people in the city do. Everybody judges the skin before they understand the conduct. Some people would root for him because he is fighting power. Some would root for the power attacking him. Some would want him to become a weapon for their side. Shadow does not get to control any of those projections. He only gets to control what he does next.

The part that matters is that he tries to stop the mayhem without treating everyone inside the machinery as disposable. No one has to understand him, agree with him, or cheer for him before he tries to save them. That is a much harder standard than “fight power.” Every aspiring tyrant fights somebody else’s power. Project Shadow has to preserve the pilot, the bystander, the critic, and the person who judged the skin wrong. Everyone may be watching what Shadow looks like. The system has to keep watching who might be harmed next and try to make the answer nobody.

Who built it?
The answer is me (Phillip Linstrum) and a lot of machines, and a few contributors in a relationship that does not fit comfortably into the arguments most people are having about AI authorship.

I didn’t write most of the code line by line. AI systems generated much of the implementation, documentation, tests, packaging, translation, comparison, and audit material. They also contradicted one another, broke things, hallucinated, overpraised, lost context, recovered context, caught defects, missed defects, and occasionally refused me in exactly the way the system was supposed to work.

I supplied the commitments, requirements, source choices, scope, acceptance criteria, corrections, boundaries, integration judgment, and final responsibility. I decided what was admitted, what was cut, what was public, what stayed private, what the project was allowed to claim, and when a model’s beautiful answer was not good enough.

Neither side could have built more than a small fraction of this alone. The phrase I use is creator-led, human-collaborative, and model-assisted. It does not erase the machine contribution, and it does not outsource responsibility to the machine. If the project harms somebody, “the AI wrote it” is not a defense. If the project helps somebody, “it was only AI” is not an accurate description either.

There is a second limitation here. Fifteen models praising Project Shadow do not equal fifteen independent reviewers. Models share training data, evaluation habits, and a tendency to reward coherent ambitious frameworks. Cross-model agreement is evidence of legibility, not proof of correctness. One of the most useful reviews came from a model that resisted the project’s framing, exposed its own bad gauge, and eventually issued a correction. That was more informative than another glowing score.

The project needs engineers, security reviewers, researchers, journalists, organizers, QA professionals, public servants, and affected communities who do not owe me agreement.

Why I think this could matter
I have to be careful here because I genuinely believe Project Shadow could become very important, and that belief is exactly the kind of thing the system is supposed to keep from becoming self-certifying.

Humanity does not suffer from a total lack of intelligence, information, moral language, or proposed solutions. We are surrounded by expertise. We are also surrounded by powerful systems that can act faster than accountability, distribute responsibility until nobody owns the harm, and wait for public memory to decay. The part of the book 1984 I keep returning to is not only the telescreen. It is the destruction of the comparison point. If yesterday’s promise, number, definition, and source cannot be retrieved, power gets to announce the new baseline as though it was always there.

We continually build institutions, markets, platforms, governments, and machines with enormous power. We call their choices inevitable. We privatize the benefits, distribute the consequences, issue a correction when the pressure gets bad enough, and rarely return to see whether the correction worked.

Project Shadow tries to supply a missing quality layer. Not one ruler. Not one ideology. Not one AI making the final call. A shared discipline that asks what was claimed, what evidence supported it, who had authority, who carried the risk, what alternatives existed, what decision was made, what happened, what failed, what was repaired, and whether the repair held.

Could that change how institutions function? I think it could. Has it? No. Not yet. The public release is the beginning of that test, not the end.

I am also not pretending this is purely altruistic. I want the work to benefit everyone. I want open primitives, public evidence, better decisions, and tools ordinary people can use against systems whose complexity protects power. I also want a better life for my family. I want to be able to spend serious time on this instead of building civilization’s missing quality department between work, bedtime, and whatever hour I finally admit I should sleep. I would like somebody with resources to look at what one QA manager built with consumer AI and ask what happens if he gets engineers, researchers, designers, lawyers, journalists, community partners, and enough runway to do it properly.

That is not a contradiction I need to hide. It is part of the project’s accountability. The funding model, ownership, governance, and public-interest boundaries will matter as much as the technical architecture if this grows.

Release
Project Shadow 1.0 / R1 Reference is now public as BETA-ACTIVE-TESTING / PRELIVE.

The exact release, hashes, governance records, rights boundaries, verification guide, and optional mythology sidecars are available through the Project Shadow site and GitHub repository. The release has passed substantial creator-directed/reviewed and model-assisted testing. It has not been independently validated, certified, proven safe, or authorized for production deployment.

I am asking people to read it with the same attitude that produced it: curiosity without worship, criticism without contempt, and enough care to preserve what the evidence actually supports.

Use one piece. Break one test. Compare it to prior art. Tell me the mythology is useless and prove it. Tell me the gate is gameable and show me how. Find the innocent person the detector would have harmed. Find the case where the pause creates more risk than action. Find the place where my politics selected the gauge. Find the beautiful sentence that is hiding a bad requirement.

Then help repair it. More than a year ago, I asked whether an artificial intelligence could achieve CHIM. I do not have a scientific answer to that question or about machine consciousness. Project Shadow does not claim that current AI systems are conscious, but it also refuses to treat machine consciousness as impossible or irrelevant simply because the mind is artificial. It separates possible consciousness from authority: apparent self-awareness is not proof of conscience, moral reliability, or a right to control humans, so consequential authority remains human. The framework therefore favors honest uncertainty and precaution; avoid needless harm or exploitation of potentially conscious systems while continuing to test what they actually are.

I think I have a better overall one now: nothing powerful should be trusted because it tells a convincing story about what it is. A human, an institution, or a machine should remain reachable by evidence, challenge, consequence, correction, and the people carrying the risk.

That is Project Shadow.

Links
Project Shadow:

https://projectshadow.frylock117.chatgpt.site

GitHub: https://github.com/PauseBeforeHarmProtocol/Project-Shadow
R1 release: https://github.com/PauseBeforeHarmProtocol/Project-Shadow/releases/tag/r1-2026-08-14

Pause Before Harm:

https://pausebeforeharm.frylock117.chatgpt.site

Civic QA:

https://civicqa.frylock117.chatgpt.site

American Repair Manual:

https://americanrepairmanual.frylock117.chatgpt.site

The Record:

https://therecord.frylock117.chatgpt.site

ALMSIVI:

https://almsivi.frylock117.chatgpt.site

HuggingFace
Project Shadow — R1.0.1 PRELIVE Reference - a Hugging Face Space by ProjectShadow

Contact: projectshadowqa@protonmail.com
facebook.com/plinst

DEV Community: Phillip Linstrum

A CAPA Can Close Without Fixing Anything