Declared / v1 / 2026-07-29

What this claims, and what would prove it wrong

Principle 4 of this work says a system that states no expectations cannot be caught missing. That applies to the work itself, so here is what it expects, before you read it.

The load-bearing claim

A miss is a gap between a declared expectation and an observed outcome, not between an outcome and anyone's standard of the good. Five reviewers say this is the weakest part of the essay and one says it breaks outright. It is kept in this draft because it can be run tomorrow, and it is published alongside the objection rather than instead of it.

What five reviewers found

Before publishing this draft I put it through a structured adversarial review. Five research agents were each assigned a different specialism, told to find a real authority in that field, and instructed to attack this essay from that authority's documented published positions: frontline safety reporting, independent accident investigation, safety-case and requirements critique, incident analysis in the resilience engineering tradition, and the design of oversight institutions.

I want to be exact about what that is and is not. These are not endorsements. No living person has read this essay and signed off on any of it, and I am not going to imply otherwise by borrowing anyone's name. What the method produced was sourced, checkable objections drawn from real published work, several of which corrected outright factual errors in an earlier version of this text. The findings below rest on their own evidence and on nobody's authority, which is how they should be read.

The verdict was that the architecture survives and this draft does not yet. The objections are unresolved in this version.

The obvious objection to this review, before anyone else makes it.

Five prompts to one model are not five independent checkers. By principle 7 of this very essay, checkers only compose when their failures are uncorrelated, and five instances of one system are correlated by construction: approximately one checker, wearing five hats. So the validation procedure I ran here is weaker than the standard the essay sets, and I am not going to describe it as a panel or a peer review, because it was neither.

What it was good for is what a single well-read checker is good for: it caught factual errors, it found an internal contradiction between two of my own principles, and it supplied a counterexample I could verify myself. What it cannot do is tell me whether the argument holds. That needs people, and the docket is where their objections will go when they arrive.

The keystone does not catch the cases that matter most. Measuring against declarations catches violations of what was promised and misses harm nobody promised against, which in most working lives is the larger category. A reviewer supplied the reductio from a real fatality: where a system's perception had no category for the hazard, its declaration that the path was clear was vacuously true, no declared expectation was violated, and a person died. Under this definition that is not a miss.
The declaring party writes the ruler. Under pressure the honest move and the profitable move diverge: declare less, later, narrower, and measured performance improves while the system degrades. Calling the result neutral is generous; auditable is the accurate word, and only while a custodian other than the declarer holds an unalterable register.
Principle 4 manufactures the correlated checkers principle 7 forbids. Two parties reading one declaration set inherit the same categories and the same omissions, however independent their interests. Independence of interest is not independence of representation.
The architecture is closed. Nothing in it requires findings to reach anyone outside the loop, so a deciding layer can simply decline to act and no principle does anything. A missing thirteenth principle.
Nothing in it says to exercise it. An untested channel and an unused stop authority are indistinguishable from healthy ones from the inside, which is this essay's own thesis, unapplied to its own architecture.
Principle 8 is underspecified. Biology enforces limits of scope, not limits of power: an immune cell destroys the target it convicts. A catcher that cannot compel evidence is a suggestion box, and reviewers disagreed about whether the power to compel buys effectiveness or invites elimination. Both are probably right about different institutions.
The confidence labels on the principles are wrong in both directions. The one graded weakest has substantial literature behind it and the one graded adequate is the least defensible in the set.
The essay asserts that one primitive holds across ten orders of magnitude and does not argue it. A fast reflex loop and a slow interpretive reporting loop have opposite requirements, and the human sections run on the second while borrowing authority from the first.
The biology does not vouch for the keystone. There is no declared expectation anywhere in a reflex arc, in proofreading, or in antigen presentation. The one claim carrying the structure has no natural instance.
Principles 2 and 9 are in tension, and 2 and 10 more so. Minimisation is optimised for the adjudicator and starves the learner: what changes another team's behaviour is the context that de-identification strips.
No cost figures, no funding instrument, no host organisation, and no answer to what stops a catching layer being folded back into the operating division in year three.

What would falsify the whole thing

A domain where a catching layer is demonstrably independent of its deciding layer, receives protected reports, and still fails at the rate of a system with neither. Or a working definition of a miss that does not reduce to a declared expectation and does not require agreement about ends. Either one would take the argument apart rather than dent it.

What this does not claim

That a catching layer prevents harm nobody declared against.
That the architecture is novel. Nature arrived at all of it; the contribution is stating it in one place.
That the coherence research mentioned at the end is established. It is an open question with a published null.