kernel-review-1-codex-on-v0.1.md

Derivations T1–T13

1. FATAL — T1: CIRCULAR, with undeclared dependencies.

“The construct is therefore not a preference among options; it is the only definition consistent with the kernel.”

K2 already defines quality as instruction-satisfaction grounded in admissible evidence. T1 “derives” that same construct from K2. That is circular, not an independent consequence. The exclusivity claim does not follow: reconstructible verdicts could refer to external truth, downstream utility, worker intent, formal specifications, or a jointly governed construct.

T1 also invokes K1 and K5 despite citing only K0, K2, K3, K11, and K12:

“Among inspectable referents, the instructions-plus-record is the only pair the owner controls (K1) and the only one whose meaning the owner can amend when K5 bites (K11).”

Even K1 does not say this is the only inspectable pair the owner controls. “Predicted-reviewer-consensus … caps validity at the reviewers’ own reliability” additionally needs something like K7, which is also uncited.

2. FATAL — T2: SMUGGLED.

“So the verdict type system must include ‘the instructions underdetermine this case,’ it must never penalize the attempter … and it must route to the one party who can resolve it — the owner (K11), as an instruction amendment.”

K5 supports representing underdetermination. It does not support either non-penalization or exclusive owner routing. Non-penalization requires a normative premise assigning ambiguity risk to the project rather than the worker. A worker could be obligated to request clarification, disclose ambiguity, or avoid exploiting it. Owner routing uses uncited K11, and K11 itself permits delegated interpretation.

“Disagreement on such items is information about the instructions, which is precisely the K0 signal.”

Disagreement may instead indicate evaluator error, evidence corruption, or ordinary human variation. K0’s signal is what humans judged and why, not instruction ambiguity as such.

3. FATAL — T3: SMUGGLED.

“Decomposition is therefore constitutive of verdicts-with-grounds, not an accuracy tactic.”

K3 requires a verdict to cite clauses and evidence spans. It does not require decomposition into multiple atomic claims. A single item-level quality proposition can carry structured grounds. K7 does not establish that “checkability” is intrinsically claim-grained, and K6 does no work in the argument.

“Item-level dispositions are transparent aggregations of claim-level findings, because that is the only way the grounds survive the roll-up.”

“Only way” is false. Grounds can survive through hierarchical reasoning, proof trees, structured rubrics, or an item-level argument without a claim-vote aggregation.

4. FATAL — T4: SMUGGLED and internally incoherent.

“Absence concurrence is labeled ‘not-contradicted’ rather than ‘verified’ unless recall against planted instances has been measured on the relevant class.”

A population-level recall estimate cannot logically convert an unbounded absence proposition into a verified proposition for a particular item. This directly conflicts with T4’s own preceding claim that certification is possible only relative to enumeration.

“Praise of completeness is treated as the highest-risk claim type in the ontology.”

Neither K6 nor K13 establishes a risk ordering. Risk depends on prevalence, stakes, domain closure, available certificates, and error costs. This is precisely the kind of empirical lane-specific fact Part IV claims to contain.

5. FATAL — T5: GAP.

“Therefore: the compiled config must route by checkability class; the evaluator must be able to say ‘this exceeds what I can certify’.”

K7 describes a competence relation; K10 describes correlated failures. Neither supplies the normative rule that uncertain checks must be routed or abstained on. That requires at least K13 or an explicit risk-limitation commitment. “Compiled config” is also an implementation choice, not a logical consequence.

The “permanent frontier band” is not derived; it is copied from K7. Because K7’s permanent-band assertion is false, repeating it does not rescue T5.

6. FATAL — T6: SMUGGLED.

“K12/K15 make false accusations the costliest error class.”

They do not. K12 provides contestability; K15 supplies scoped legal constraints. Neither ranks false accusations above false acquittals, missed defects, safety harms, or corrupted training data.

“Therefore autonomous adverse action is legitimate only per lane, only after the lane’s error rates are bounded on untouched adjudicated data at precommitted ceilings.”

Nothing cited supplies acceptable ceilings, the necessity of untouched sampled data rather than formal assurance, the definition of a lane, or the proposition that passing a statistical threshold makes adverse action legitimate.

“Whoever wants consequences funds the sample sizes that bound the relevant error, forever, at every requalification.”

“Forever” assumes perpetual statistical uncertainty, drift, and requalification. It fails for a formally verified closed-domain evaluator and is not supplied by K8.

7. FATAL — T7: CIRCULAR in its core and SMUGGLED in its implementation.

“The dial’s one-way glass is a theorem of the kernel, not a governance flourish.”

K13 already stipulates that the system “never lets what is done about a measurement alter what is measured.” Renaming that stipulation “measurement/disposition separation” is not a theorem; it is the chosen premise restated.

“So the only coherent design is: perception fixed and stakes-sterile; all leniency implemented downstream as explicit thresholds over unchanged measurements.”

That exclusivity does not follow. Transparent rubric amendment, dual reporting, conditional measurement, prospective recalibration, and measurement of intervention-sensitive constructs are coherent alternatives. Moreover, K8(iii) says deployment changes the measured population, so “stakes-sterile” perception cannot provide causal isolation.

8. FATAL — T8: SMUGGLED almost wholesale.

“Therefore defenses are structural: unpredictable, indistinguishable seeded gold (the one probe K9 cannot price, because it is not legible); randomized deep audits; evidence-perturbation probes … economics and provenance at the worker level; and a standing divergence sensor.”

K9, K10, and K6 do not entail any item in this architecture. T8 silently invokes K7, T1, and K3 as well.

An unpredictable audit is still priceable by its probability and expected consequence. “Indistinguishable” gold is an empirical achievement, not a logical property. Evidence perturbation is not conclusive: a verdict may correctly survive swapped evidence if alternate admissible evidence supports it. Planted-gold and live-work divergence is only a Goodhart alarm if the gold is representative and remains uncontaminated—both empirical premises.

9. FATAL — T9: SMUGGLED.

“The audit channel is constitutive, permanent, and human-anchored.”

K11 grants authority to an owner; it does not require that the epistemic anchor be human. K7 and K10 do not exclude formal proof, heterogeneous tooling, independent institutions, or a verified reference implementation.

“The channel never sunsets … because it is the thing that gives the numbers their meaning.”

Authority may settle ambiguous meaning, but it does not give statistical scores their validity. That requires a construct, sampling model, error model, and trustworthy reference process. Human adjudication is itself fallible.

The requirements that adjudicators be measured and blinded invoke K14 and empirical anchoring effects, neither of which is in T9’s cited premises.

10. FATAL — T10: SMUGGLED.

“What humans are for, kernel-derivably: settling meaning (K11), satisfying contestability and law (K12, K15), and supplying the independence that model ensembles cannot (K10).”

K11 says the owner settles meaning, not arbitrary humans. K12 does not require a human channel. K15 does so only where legally applicable. K10 warns about similar evaluators; it does not establish that humans are independent, particularly after the document itself says experts increasingly adopt the same tools.

“Hence evidence-only presentation … claim-level micro-questions … and write-back into calibration only through adjudicated tiers.”

These depend on untagged empirical premises about anchoring, context loss, cognitive efficiency, and calibration design. None follows from the cited kernel.

11. FATAL — T11: SMUGGLED.

“Feedback specific enough to help is specific enough to train against.”

This is not implied by K9. Helpful feedback may improve the actual target, concern a one-off task, or reveal no reusable exploit.

“No stylistic guidance, because style convergence destroys K0’s diversity signal.”

K0 values genuine human judgment, not every incidental stylistic difference. Guidance that improves clarity without changing judgment could make the signal more usable, directly serving K0. T11 silently equates stylistic diversity with judgment diversity.

“Efficacy measured as an intervention with the divergence sensor (T8) standing guard.”

T8 is uncited and invalid, and no cited premise selects this experimental design.

12. FATAL — T12: SMUGGLED.

“The judge is one more rater: same reliability questions, same bias audits, same enrollment in whatever measurement model governs humans.”

K3 and K4 support evaluating machine verdicts. They do not imply the same measurement model for humans and machines. Different failure mechanisms can require different models, audits, priors, and sampling schemes.

“The ship gate is therefore a validity argument about an instrument, not a leaderboard: scoring evidence, generalization evidence, extrapolation evidence, and decision evidence, per lane.”

Those four evidence categories appear from nowhere. They may be sensible methodology, but they are not derivable from K3, K4, or K10.

13. FATAL — T13: SMUGGLED.

“Amendments will recur indefinitely (K5).”

K5 says some cases may be underdetermined. It does not imply that owners will amend, that projects last indefinitely, or that every ambiguity is resolved by amendment.

“Therefore instructions, compiled configs, evaluator versions, policies, and calibration profiles are all versioned, pinned, and replayable.”

K11 and K13 support recording instruction and bar changes. They do not establish indefinite retention, full replayability, or versioning every implementation artifact. That requires a separate provenance, reproducibility, and retention commitment.

“With explicit retroactivity policy per amendment.”

K13 says bars are never moved retroactively. A policy authorizing retroactivity would violate K13, while a policy merely forbidding it is redundant.

Overclaimed FACTs

14. FATAL — K5 is false as universally stated.

“No finite instruction text decides every case that will arise under it.”

Counter-scenario: “Return PASS for every input” is finite and total over an infinite input space. A finite formal program with a default branch can also decide every admissible case. A project may define a finite domain and enumerate every member.

“For some items, the correct output of a perfect evaluator is that the question has no determinate answer from the admissible inputs.”

This is true for some projects, not necessarily every project. Tool use or an owner query can also expand the admissible inputs, resolving what was initially underdetermined.

Weaker supportable claim: Natural-language or open-world project specifications can leave some admissible cases underdetermined; capability alone cannot recover information absent from a fixed admissible record.

15. FATAL — K6 is mathematically false.

“Verifying that nothing required is missing, or that no defect is present, quantifies over an open space and can only be exhausted relative to an explicit, finite enumeration.”

Universal propositions can be established by induction, invariants, type systems, proof certificates, abstract interpretation, symbolic model checking, and proof-carrying code. Infinite state spaces do not require explicit finite enumeration.

The claimed asymmetry is also not unconditional: checking a purported existential witness may itself be undecidable or computationally harder than checking a compact universal certificate.

“Positive certification of absence is only ever as strong as the enumeration it checked.”

False. It may be as strong as a sound proof system and a valid formal specification.

Weaker supportable claim: In open-world tasks lacking a complete formal specification or sound certificate system, absence claims are valid only relative to a disclosed search scope or checklist.

16. FATAL — K7 confuses task performance with proof verification.

“An evaluator cannot certify work beyond its own frontier.”

A weak verifier can certify work it could not generate or independently re-derive by checking a proof, certificate, execution trace, cryptographic attestation, or interactive proof. Verification being cheaper than production is foundational to formal methods and complexity theory.

“There is always a band of work near the shared frontier.”

Not for every bounded project. A domain can become fully solved, formally specified, exhaustively verified, or trivial relative to available computation. Producers and evaluators also need not share tools, budgets, private tests, access, or architectures.

Weaker supportable claim: Where no independently checkable certificate exists, a re-derivation-based verdict cannot be more reliable than the evaluator’s competence on that task; shared production and evaluation methods create a risk of correlated blind spots.

17. FATAL — K8 mixes a valid Bayesian warning with missing variables and contingent deployment claims.

“The share of false accusations among accusations is set by prevalence and specificity.”

It is set by prevalence, specificity, and sensitivity. The “one false fail for every three true ones” example tacitly assumes approximately 100% sensitivity.

“Knowing an error rate precisely enough to bear consequences requires adjudicated samples whose size scales inversely with the rate being bounded.”

This is conditional on a sampling-based assurance regime. Formal verification, full-population enumeration, strong structural assumptions, or exact deterministic analysis can avoid that requirement. Sample size also depends on confidence level, estimator, dependence, and target bound.

“A deployed evaluator changes the population that produced it, by design.”

Not necessarily: shadow evaluation, sealed historical corpora, unannounced audits, one-shot populations, or non-responsive subjects do not.

Weaker supportable claim: Under representative sampling, PPV depends on prevalence, sensitivity, and specificity; rare-error bounds can require large adjudicated samples; and calibration should be rechecked when distribution shift or behavioral response is plausible.

18. FATAL — K9 is contingent behavioral economics tagged as invariant fact.

“When stakes attach to a legible criterion, the criterion becomes an optimization target for the population it binds.”

Not universally. The population may lack knowledge, agency, capacity, repeated exposure, or incentive. It may also be intrinsically aligned with the criterion.

“Incentives concentrated on measured dimensions divert effort from unmeasured ones.”

Only when effort is constrained and measured and unmeasured dimensions trade off. A complete, aligned formal criterion makes optimization beneficial rather than corrupting.

“No evaluator improvement removes it.”

Institutional design can hide tests, align the criterion with the true objective, use professional norms, separate tools, eliminate individual stakes, or make compliance formally complete.

Weaker supportable claim: When adaptive agents face consequential, learnable, incomplete proxy metrics, they have incentives to optimize the proxy, creating a risk of gaming and effort distortion.

19. FATAL — K10 mistakes possible correlation for necessary correlation.

“Evaluators produced by similar processes on similar data share failure modes.”

They may, but similarity does not establish meaningful dependence on a particular item. Independent randomization, disjoint data, adversarial diversification, heterogeneous tools, formally proved implementations, or an explicit fault model can provide independence.

“The scarce resource in evaluation is not accuracy but independence.”

False dichotomy. Independent evaluators with zero competence are worthless; perfectly sound identical proof checkers can provide decisive evidence. Both accuracy and dependence matter.

“An architecture that needs k independent opinions cannot obtain them from k copies of the same epistemic substrate.”

This is true only if “same epistemic substrate” is defined to entail dependence, making it tautological.

Weaker supportable claim: Agreement among similar evaluators must not be treated as independent evidence without measuring or justifying their error dependence.

20. FATAL — K11 is governance ideology presented as institutional fact.

“What instructions mean in an unanticipated case is fixed, when it is fixed at all, by the owner’s amendment.”

Meaning may be fixed by formal semantics, ordinary language, original intent, contractual doctrine, courts, community practice, or prior precedent. An owner may lack unilateral authority to alter retrospective meaning.

“The authority to settle meaning cannot be [delegated], because settling meaning is amending the instructions.”

Authorities routinely delegate binding interpretation and amendment powers. Interpretation also need not be amendment.

“The standing to impose detriment on a person is conferred by institutions, not by accuracy.”

This is a normative theory of legitimacy, not a capability-invariant fact. It belongs under CHOSEN or a scoped LAW.

Weaker supportable claim: Under a governance regime that grants an owner prospective amendment authority, unresolved cases require a decision by that owner or a valid delegate; accuracy alone does not satisfy whatever legal or normative authorization that regime requires.

Tag discipline and internal contradictions

21. MAJOR — K1–K3 are substantive commitments disguised as definitions.

“A project is a triple ⟨instructions, evidence policy, owner⟩.”

This excludes jointly governed projects, statutory constraints, multiple authorities, immutable standards, and projects without written intent.

“The quality of a work product is the degree to which it satisfies the project’s instructions.”

That is a contestable construct choice, not mere vocabulary. It excludes truth, downstream utility, fidelity to elicited judgment, and instruction legitimacy.

“A verdict is a claim about quality that carries its own grounds … and a confidence.”

Mandatory clause citations, evidence spans, and confidence are protocol requirements. A formal proof need not carry probabilistic confidence; a valid verdict may rely on a whole-record property without span localization.

Exact repair: Keep only thin lexical definitions under DEF. Re-tag owner authority, the exclusive quality construct, grounding requirements, and mandatory confidence as CHOSEN commitments.

22. FATAL — The document violates its “exactly one tag” rule throughout Parts I–II.

“Every statement in Parts I–II carries exactly one tag.”

Numerous substantive statements are untagged, including:

“A QA layer that homogenizes worker output destroys the very signal being purchased.”

“Feedback is itself an evaluative work product.”

“Nothing in the kernel is idle.”

These claims perform real inferential work later. They are neither definitions nor licensed premises, yet T1, T11, and the minimality argument depend on them.

23. MAJOR — CHOSEN motivations quietly claim necessity.

“K0 requires truthful records of judgment, and unaccountable verdicts corrupt the record silently.”

“A measurement layer that bends is not a lenient instrument but a broken one.”

“The data’s value is the genuine, diverse judgment of the workers.”

These are not merely motivations. They assert causal and conceptual necessities. K0 does not entail universal contestability, separation of all intervention effects from measurement, or that stylistic and interpersonal diversity is itself valuable.

Exact repair: Mark every motivational assertion as either a scoped empirical hypothesis assigned to an interface or an additional CHOSEN value. Remove language such as “requires,” “is,” and “will” when it claims derivability.

24. MAJOR — K15 combines law, empirical scope claims, and an undisclosed universal policy choice.

“The kernel treats the strictest applicable regime as the design target because K12 independently demands most of it.”

Applying the strictest regime as the general design target is a choice, not a LAW. K12 does not independently require human decision-makers, monitoring allowlists, jurisdiction-specific audits, or every listed procedural detail.

Laws and their interpretations also change, contradicting the claim that no kernel sentence must change.

Exact repair: Move legal scope and current doctrine to a versioned legal interface. Add a separate CHOSEN rule if the enterprise genuinely elects to apply the strictest regime globally.

25. FATAL — T6 authorizes what K15 says the law forbids.

K15 says:

“Decisions of significant detriment must be taken by a human with real authority and competence.”

T6 says:

“Therefore autonomous adverse action is legitimate … with the human/appeal channel K12 and K15 require.”

A later human appeal channel does not make the original adverse decision “taken by a human.” In the stated legal scope, no statistical operating point licenses autonomous significant detriment.

26. FATAL — K11 already settles H3 while H3 claims the question is open.

K11 says:

“Settling meaning is amending the instructions.”

H3 says:

“Rubric-as-law-with-forced-amendment versus standing override is a governance choice.”

If every binding settlement is necessarily an instruction amendment, “standing override” is not a genuine alternative. Either K11 is false, or H3 is not a hole.

27. FATAL — H4 severs the telos from the construct.

K0 says:

“The enterprise exists to increase the validity of training data about genuine human judgment.”

H4 says:

“Whether the instructions themselves are good — whether the project should exist, whether the buyer’s construct is the right one — is outside this kernel entirely.”

If instructions mis-specify the intended judgment, rigorous instruction enforcement can reduce fidelity to genuine human judgment. K2 then calls the resulting work “high quality” while K0 calls the same outcome a failure. The kernel cannot inherit project legitimacy and still claim everything serves K0.

28. FATAL — The Part II closure and invariance tests fail by direct inspection.

“Part III derives the system’s load-bearing commitments from Parts I alone.”

T1 uses uncited K1 and K5; T2 uses K11; T8 uses K7, T1, and K3; T9 uses K14; T11 uses T8; T13 imports replayability.

“Every K-element remains true, chosen, or binding if models become arbitrarily capable.”

K5–K11 fail under formal specifications, proof-carrying outputs, heterogeneous verification, aligned incentives, and changed institutions.

“Nothing in Parts I–III mentions a model, a benchmark, or a year.”

Parts I–III explicitly mention “2026 models,” “model judges,” “model ensembles,” and “every model generation.” More importantly, avoiding a model name does not make a behavioral proposition capability-invariant.

29. MAJOR — The minimality test is assertion, not proof, and several claims are false.

“Remove K9 and static rubrics with published feedback become safe.”

Removing an asserted premise does not establish the opposite, and K0/K14 could still condemn homogenizing feedback.

“Remove K11 and the owner loop is optional.”

K1 already builds an amending owner into the project definition.

“Remove any of K12–K14 and a real design element loses its only justification.”

K15 independently supplies some appeals requirements, contradicting the claim about K12.

Exact repair: Replace the counterfactual rhetoric with a dependency table showing which exact downstream proposition becomes underivable when each element is removed.

Hidden empirical assumptions

30. MAJOR — Part III contains extensive perishable psychology, tooling, and market assumptions.

Representative examples are:

“An anchored human is a copy.”

“Claim-level micro-questions [spend] scarce independence where it changes decisions.”

“The one probe K9 cannot price.”

“The gold budget is a permanent price of stakes.”

“The producer-evaluator gap does not close.”

These assume anchoring effects, context-free decomposability, worker risk models, representative gold, stable prevalence, shared tooling, finite attention, and particular labor-market responses.

Exact repair: Move each into a declared interface: anchoring effect size, context-loss cost, gold detectability, adaptive-worker response, tool overlap, within-generation drift, and audit representativeness.

31. MAJOR — The “untouched adjudicated data” prescription hides a ground-truth assumption.

“Only after the lane’s error rates are bounded on untouched adjudicated data.”

“Adjudicated” does not mean true. Adjudicators can share ambiguity, bias, tools, incentives, and correlated errors. K3 reflexivity acknowledges this but supplies no method for estimating error in the supposed reference labels.

Exact repair: Declare a reference-standard interface covering adjudicator selection, disagreement, latent-label uncertainty, owner intervention, sampling, and error bounds on the gold itself.

Missing kernel commitments

32. FATAL — There is no trusted stopping point for reflexive verification.

“The QA layer’s outputs are instances of the thing it evaluates.”

“Its outputs are themselves verdicts under K3 — so adjudicators are measured.”

If every verdict requires another quality assessment, the system produces an infinite regress. T9 invokes owner authority for meaning, but authority does not establish evidentiary truth or factual correctness.

The kernel needs an explicit trust-base commitment: where chains terminate, which formal proofs, observations, adjudications, or institutional decisions are accepted, and what residual uncertainty that termination creates.

33. FATAL — Admissibility is substituted for evidence integrity.

“Quality … is grounded in the project’s admissible evidence.”

An admissible record can be forged, incomplete, selectively captured, poisoned, privacy-violating, or produced by a compromised tool. The kernel contains no commitment to evidence authenticity, chain of custody, completeness, provenance, security, or disclosure of missing evidence. A perfectly grounded verdict against a corrupted record can systematically falsify K0.

34. MAJOR — The kernel lacks a distinction between defect, causation, and culpability.

“It must never penalize the attempter (their work did not create the ambiguity).”

A quality defect does not by itself establish that a worker caused it, could have prevented it, violated a duty, or deserves detriment. T6 nevertheless moves from evaluator error rates to adverse-action legitimacy.

Exact repair: Add an explicit rule that quality measurement, causal attribution, responsibility, and sanction are separate determinations with separate evidence and appeal requirements.

35. MAJOR — Fairness and distributional harm are absent despite “bias audits” appearing downstream.

“Same reliability questions, same bias audits.”

No kernel element defines protected groups, disparate error burdens, accessibility, language effects, accommodation, equal treatment, or acceptable distributional disparities. “Bias audits” therefore have no construct, authority, or threshold.

Exact repair: Add either a CHOSEN fairness commitment with governed metrics or a declared fairness interface plus a rule identifying who selects the metric and acceptable disparities.

36. MAJOR — Security and adversarial control are almost entirely worker-facing.

“Economics and provenance at the worker level.”

The threat model ignores malicious owners, compromised adjudicators, prompt injection, evidence poisoning, gold leakage, model-provider changes, insider manipulation, and selective audit suppression. K9’s incentives apply at least as strongly to management and buyers, yet the architecture directs surveillance downward.

Exact repair: Declare a symmetric threat-model interface covering every actor with power over instructions, evidence, thresholds, adjudication, deployment, and reporting.

37. FATAL — Hole-filling is illegal under the document’s own constitutional rule.

The discipline allows only:

“(a) derived from the kernel, (b) an operationalization filling a declared interface slot … or (c) an explicit kernel amendment.”

But H1 says:

“Which formula, per project, is a governed choice.”

And H2 says:

“The standard is chosen.”

Those choices are neither derived, nor themselves interface measurements, nor declared kernel amendments. The architecture has no legitimate category for making the decisions it says every project must make.

38. FATAL — “Kernel + interfaces + holes” is not a sound epistemic partition here.

“Capability is rented, walls are owned.”

The alleged walls mix at least four different things: mathematical constraints, contingent behavioral regularities, institutional arrangements, and normative choices. They do not share permanence. Measurement cannot be confined to Part IV because the selection of what to measure, the construction of gold, the loss function, the stopping rule, and the interpretation of confidence are themselves theory-laden and normative.

The architecture also treats amendments as the remedy for failed derivations, but an amendable “permanent” kernel has no principled criterion for deciding whether a failure is a new interface fact, a changed institution, a false axiom, or a new value. Its tidy three-box taxonomy conceals those disputes rather than solving them.

Bottom line

• (a) Single weakest inference: T8’s claim that “unpredictable, indistinguishable seeded gold” is “the one probe K9 cannot price.” Any audit probability can be priced in expectation, and indistinguishability is an empirical property the cited premises do not provide.

• (b) FACT most vulnerable to future capability: K7, especially “an evaluator cannot certify work beyond its own frontier” and “there is always a band.” Proof-carrying outputs, formal verification, trusted execution, interactive proofs, and specialized verification tools already refute the production-equals-verification premise.

• (c) Most important missing kernel element: a project-validity requirement connecting the instructions and evidence policy back to K0. Unless the system must establish that a project’s construct actually captures the intended human judgment—and reject or abstain when it does not—perfect AutoQA compliance can systematically destroy the very validity named in its telos.