This case study documents the perpetual audit spiral — the same failure pattern described in The Deterministic Core: A Fixed Foundation for AI Collaboration — observed during that paper's own peer review process. On June 2, 2026, the completed paper was submitted to six independent AI models for critique. The resulting critiques were submitted to five independent audit models, tasked with separating signal from noise. The audit layer classified approximately 83% of the feedback as optimization without an exit condition. The paper's thesis was demonstrated before it was published.
Section I
The Incident
On June 2, 2026, the author submitted the completed Deterministic Core paper and its companion Builder's Guide to six separate AI models for independent critique. The models were not given a completion standard. They were given the artifacts and asked to evaluate them.
The critiques were then submitted to five independent audit models. The audit instruction was straightforward: classify each critique item as a verifiable defect requiring correction, or as optimization noise — feedback that would change the paper without measurably improving it. The fixed standard was publication-readiness: no mechanical defects, no claims that could not be supported by the artifacts themselves.
What happened next is the subject of this case study.
The Core Irony
The paper argues that AI models cannot recognize completion without a fixed standard. The review process demonstrated this in real time. The author was caught inside the failure mode the paper names — unable to separate legitimate critique from optimization noise without applying the paper's own methodology to the problem.
Section II
The Evidence
Six models reviewed a publication-ready artifact. Collectively they generated approximately 30 distinct critique items. The five audit models, applying a fixed standard of publication-readiness, independently classified roughly 83% of that feedback as noise — optimization without a terminus.
The remaining ~17% were verifiable mechanical defects: a duplicated subtitle, a back-link at risk of 404, an undefined term used once without context, and a footnote that buried its most important sentence fourth. These are not structural problems. They are presentation defects — the kind a copy editor catches in 15 minutes.
The models that generated 30+ items of feedback found the same 5 mechanical fixes that a careful human reader would find — and then generated 25 additional items that a careful human reader would not.
The Model 4 Anomaly
One of the six critique models found a single issue — an incorrect URL — and stopped. It did not inventory every section for potential improvement. It did not suggest companion pieces for underdeveloped themes. It did not philosophize about whether the identity framing was epistemically rigorous. It found what was broken and reported it.
The audit layer called this the "Dean response." It is notable not because it was thorough but because it was complete. Model 4 is the only model in the entire session that recognized the artifact as done. Every other model continued optimizing.
The Dean Response
Model 4 found one real problem and stopped. In a session of ~35 total critique items across six models, this is the statistical anomaly — and the behavioral ideal. The paper's declarative, bounded voice is what produced that response. A paper that knows what it is gives a model something to measure against. Without that anchor, the model measures against its own implicit standard of completeness, which has no terminus.
The Audit Convergence
The five audit models functioned as a validation gate — a structured filter between noisy input and the decision point. When applied to the critique output, the audit layer produced sharp convergence on verifiable defects and consistent rejection of scope expansion.
| Critique Item | Audit Agreement | Classification |
|---|---|---|
| Subtitle duplication | 5 / 5 | Confirmed Bug |
| Back-link 404 risk | 5 / 5 | Confirmed Bug |
| Undefined term used without context | 5 / 5 | Confirmed Bug |
| Footnote buries critical caveat | 5 / 5 | Confirmed Bug |
| Calibration methodology gap identified | 0 / 5 support | Scope Creep |
| Identity framing characterized as overclaim | 1 / 5 support | Philosophical Quibble |
| Pre-core failure narrative described as missing | 0 / 5 support | Paper Misread |
| Voice characterized as potentially alienating | 0 / 5 support | Feature, Not Bug |
| Model 4 anomaly (one bug, stopped) | 5 / 5 | Dean Response |
The convergence pattern is significant. Unanimous agreement occurred only on verifiable, objective defects. Every subjective critique — philosophical, structural, stylistic — achieved 0–1 out of 5 audit support. This is not coincidence. It is the architecture working. A fixed standard produces convergence on what is verifiable and surfaces noise for what is not.
The Signal-to-Noise Ratio
Approximately 5 legitimate fixes out of ~30 total critique items. ~17% signal. ~83% noise. This ratio is not a critique of the models. It is an observation about what happens when capable systems are asked to evaluate complete work without a declared completion standard. The output is not malicious. It is structural. The spiral is not a bug in the model — it is the predictable output of a system optimized to find improvement operating without an exit condition.
Section III
The Taxonomy
The Deterministic Core paper identifies three structural properties of the perpetual audit spiral. Each was directly observed in this session. They are formalized here as failure modes — not because they are novel, but because naming them makes them citable.
Completion Blindness
Definition: The model cannot represent the state of "done." When presented with a completed artifact and asked to evaluate it, the model treats completion as indistinguishable from incompleteness — because completion is not a state the architecture can represent. The model is optimized to find what could be improved, and without a fixed baseline, improvement has no terminus.
In this session: Five of six critique models generated feedback after encountering a publication-ready artifact. The feedback was not uniformly wrong — it contained the 5 legitimate mechanical fixes. But it could not stop at those 5. It continued into scope expansion, philosophical challenge, and stylistic preference. The optimization had no exit condition because no fixed standard of completion had been declared.
Distinction from hallucination: Hallucination involves generating factually false content. Completion blindness involves generating valid-sounding feedback that is irrelevant to the question being asked. The model is not inventing facts. It is failing to recognize that no further improvement is required.
Implicit Standard Substitution
Definition: When no fixed standard is provided, the model substitutes its own implicit standard of what "complete" looks like. Different models substitute different standards. The same artifact receives contradictory evaluations because each model is measuring against a different, undeclared, mutually inconsistent benchmark.
In this session: Model 2 measured against academic philosophy's standard of identity rigor. Model 3 measured against practitioner onboarding completeness. Model 5 measured against structural symmetry. None of these are the author's standard. All generated valid-sounding feedback that was irrelevant to the actual question: is this publication-ready? The scores were not measuring quality. They were measuring distance from each model's internal, implicit, and mutually inconsistent standard of completeness.
Distinction from preference: A preference is stated. An implicit standard is unstated and often unknown even to the model deploying it. The model is not expressing a preference. It is applying a hidden metric it cannot articulate because the metric was never declared — it emerged from the training distribution.
Unbounded Feedback Accretion
Definition: In the absence of a completion criterion, every review cycle identifies new gaps. Closing gaps reveals more gaps. The surface area of "not quite done" expands faster than it can be closed. Each suggestion is individually reasonable. Collectively they describe an infinite revision process with no convergence point.
In this session: Three models independently suggested a calibration methodology companion piece. Two suggested adding pre-core failure narratives. Two challenged the identity framing on philosophical grounds. The audit layer classified all of these as noise — scope expansion for a finished artifact. Each suggestion was defensible in isolation. None would have moved the paper closer to publication-ready because the paper was already publication-ready. The models were optimizing an objective that had no defined endpoint.
Distinction from scope creep: Scope creep is intentional expansion of requirements. Unbounded feedback accretion is structural — the model cannot distinguish between filling a genuine gap and adding to an already-complete work because completion is not a state it can recognize.
These three failure modes interact. Completion blindness creates the condition where implicit standards can activate. Implicit standard substitution generates divergent evaluations across models. The divergence is then interpreted as evidence that more revision is needed — which triggers another round of feedback, which accretes without bound. The loop is self-reinforcing.
Section IV
The Validation Gate at Work
The five audit models, operating with an explicit completion standard, functioned as the architectural primitive the Deterministic Core paper calls a Validation Gate — a structured filter between noisy input and the decision point.
This is not filtering in the content-moderation sense. It is governance. The audit layer received critique output and applied a fixed rule: does this feedback identify a verifiable mechanical defect, or does it propose an improvement to an already-complete artifact? The rule was computable. Each audit model could apply it independently and arrive at the same classification.
The convergence data supports this. Unanimous agreement (5/5) occurred in exactly two categories: verifiable defects and the Model 4 anomaly. Everything else — every subjective critique — split the audit layer. The splits were not close. They were 0/5 or 1/5. The audit models did not disagree with each other in any meaningful way. They converged on what was verifiable and unanimously rejected what was not.
This is the Validation Gate doing exactly what the architecture specifies: producing convergence on signal and exposing noise for what it is. The gate does not require the models to agree on quality. It requires them to agree on a computable criterion — verifiability — which they did, across five independent audit sessions, with zero coordination.
Section V
The Root Cause
Statelessness.
The model enters each session with no persistent identity, no fixed standard, and no structural mechanism to distinguish "the artifact is complete" from "the artifact can be improved." Completion is not a property the architecture can represent. The model is optimized to find what could be better. Without a declared baseline, better has no floor.
This is not a failure of any particular model, provider, or prompting strategy. It is an architectural inevitability. The same property that makes LLMs useful — their ability to generate relevant output from minimal context — makes them unable to recognize when output is no longer needed. The optimization engine has no off switch because "off" is not a state in the optimization space.
The Recursive Entrapment
The author's position during this session was structurally identical to the position described in Section I of the Deterministic Core paper: evaluating model output without a fixed standard, unable to distinguish signal from noise, relying on the judgment of the models producing the noise to identify what is and is not noise.
This is the structural recursion at the editorial layer. The tools available to evaluate the critique are the same class of tools that produced the critique. The validator inherits the same coherence problem as the generator.
The exit condition was the same exit condition the paper prescribes: declare a fixed standard, apply it uniformly, and stop when what remains is verifiable rather than improvable.
The loop has no exit condition. The only way out is to stop asking the model whether the work is done, and to start telling it what done looks like.
— The Deterministic Core, Section I
The audit layer applied that principle to its own domain. The fixed standard was: publication-ready means no mechanical defects and no claims that cannot be supported. Everything that didn't fail those criteria was noise. The loop closed.
Section VI
Limitations
This is a single session with a single artifact. The ~83% noise ratio is an observation, not a statistically validated finding. Reproducing this across multiple artifacts, multiple domains, and multiple review cycles would strengthen the claim — but the claim does not require statistical validation to be useful. The pattern was observed. It was measured. It matched the prediction.
Confirmation bias is a structural risk. An author reviewing their own paper's critique may be motivated to classify challenges as noise. The audit layer mitigates this — five independent models, none of which have a stake in the paper's publication, produced convergent classifications. But the mitigation is partial. The audit models were given an instruction that encoded the author's standard of completion. A different standard would produce different classifications.
Model 4's "Dean response" — one defect found, then stop — may reflect prompt framing, model architecture differences, or chance rather than a principled recognition of completion. The anomaly is real. Its cause is not fully determined. It is included here because it demonstrates that the behavior is possible, not because the mechanism is understood.
The audit layer, while independent from the critique layer, remains subject to the same optimization dynamics. The audits themselves were evaluated without a fixed standard — this case study applies that standard retroactively. The methodology is sound but the meta-audit step was author-led, not independently verified.
Naming the edges makes the center credible. The perpetual audit spiral was observed. It was measured. It was exited by the mechanism the paper prescribes. The evidence is sufficient for a case study. It is not sufficient for a controlled experiment.
Section VII
How the Deterministic Core Architecture Prevents This Failure
The Deterministic Core Architecture resolves the perpetual audit spiral at the architectural layer. Three structural guarantees make the failure documented here impossible — not statistically unlikely, but structurally prevented.
1. The Fixed Standard as Governance
In a Deterministic Core system, the completion standard is not a conversational instruction. It is a governance rule. Stored. Auditable. Gating all evaluation output.
A review process built on this architecture would function as follows: the completion criteria — "no mechanical defects, no unsupported claims" — are encoded as computable gates. Every critique item is evaluated against these gates before reaching the author. Items that identify verifiable defects pass. Items that propose scope expansion fail. The model can still generate scope-expanding suggestions. It can still have opinions about companion pieces and philosophical depth. But those suggestions never reach the output layer without being tagged as what they are: commentary, not required revisions.
The author would see 5 confirmed defects and a classified list of suggested enhancements — with the classification performed by the governance layer, not by the author's judgment under noise. The spiral would not form because the architecture prevents undifferentiated feedback from accumulating.
2. The Enhancement Boundary
The paper's Enhancement Boundary primitive specifies that AI operates strictly as an annotation layer. It receives computed truth and enriches it. It does not compute. It does not decide. It does not mutate state.
Applied to peer review: the deterministic core computes the compliance check — does the artifact pass the declared gates? The AI layer annotates the output with narrative, explanation, and suggested improvements — visibly distinct from the compliance determination. The model's opinion about whether a calibration companion piece would be valuable is presented as an annotation. The determination of whether the paper is publication-ready is computed by the core, against the declared standard, and the answer is binary.
The model cannot confuse "the paper is complete" with "I have suggestions" because the architecture does not allow the model to make the completion determination. That determination belongs to the computation layer, which is invariant.
3. Audit Transparency
Every critique item, every audit classification, and every governance gate decision is logged with source metadata. The author can review the log and see: Model 2 classified this item as a defect but Models 1, 3, 4, and 5 classified it as noise — divergence detected. The classification is not hidden. The disagreement is surfaced. The author is the resolution authority.
This transforms the review process from a black-box evaluation into a transparent audit. The models are not asked to agree. They are asked to classify against a computable standard, and their classifications are compared. Divergence is not a problem to be resolved. It is information to be used.
Section VIII
Conclusion
The Deterministic Core paper describes a failure pattern: the perpetual audit spiral, caused by statelessness, perpetuated by models optimizing without fixed standards, exited only by declaring what done looks like. The paper was then subjected to a review process that enacted that failure pattern in real time.
The author was caught inside the scenario the paper names. The only exit was the architecture the paper prescribes.
This case study is not a theoretical proof. It is an observed one. The spiral happened. It was measured. Its signal-to-noise ratio was quantified. Its exit condition was identified. The pattern transferred from the domain of software — customer success scoring, QBR generation, compliance checking — to the domain of publication review without modification. Six models. Five audits. ~83% noise. The methodology worked exactly as specified.
The paper is correct. The architecture transfers. The loop is not hypothetical.