Synthesis

Recursive audit closure across domains

A synthesis separating recursive audit closure from the conditions that turn it into erosion, across five domains.

1 Introduction

In 2008 the category was AAA, and the bill arrived through the banking system. In AI evaluation the category is state-of-the-art, and it helps select what will be deployed. Six notes have asked, each from its own side, whether those two sentences describe one structure. This seventh paper — the series’ 2026g installment — says they do, and says where the oneness stops.

A series incurs a debt its installments cannot individually pay. The first note named three conditions under which evaluation categories lose transparency (Ortulanu, 2026a). The second specified the capacity relation and reserved the problem’s non-recognition (Ortulanu, 2026b §9). The third isolated the use channel (Ortulanu, 2026c §2); the fourth derived its structural inverses (Ortulanu, 2026d §3); the fifth gave the dynamics a clock (Ortulanu, 2026e); the sixth carried the structure beyond AI (Ortulanu, 2026f §2–§6). The pieces now name one object: recursive audit closure.

The synthesis separates closure from erosion. Closure says that judgment has entered the process it judges; erosion says that the closed loop rewards a learnable proxy faster than it preserves the task property. The first is a topology, the second a directional claim. The instances and mechanisms remain those developed in the preceding notes, but their assembly fixes the level at which the central claim holds. Scope conditions precede the cases, and disconfirmers travel with the claim.

2 The Invariant

Strip each instance to what it cannot do without, and the same four-element configuration remains. Call it recursive audit closure.

A categorical benchmark: an evaluation whose judgments are finite in actual use — by output, as with ratings and pass-fail verdicts, or by institutional conversion, as with thresholded scores and quartiles (Ortulanu, 2026c §2); a scalar turns categorical the moment institutions consume thresholds rather than quantities. Institutional reliance: third parties act on the judgments without independent inquiry, which is what makes the evaluation an audit rather than an opinion (Ortulanu, 2026b §6). A recursive use channel: the judgments enter the production, selection, or reward of the population to be judged next — the element the third note isolated and the fifth weighted (Ortulanu, 2026c §2; 2026e §2). And a capacity-bounded auditor: an evaluator whose verification capacity is bounded by the lineage, reference, or tempo available to it, whether or not that bound has yet become binding relative to the population (Ortulanu, 2026b §2; 2026d §3).

The four elements close the loop. They do not fix its direction. A perfect-alignment case makes the distinction concrete: when an exact oracle checks the intended task property, selection for correct answers can improve the population while the evaluator’s performance on a fixed comparison panel holds. Correctness within that task and coverage of a broader claim remain separate questions (Ortulanu, 2026d §5). Erosion requires three further conditions. Proxy misalignment gives selection a direction away from the task property. Differential learnability lets the proxy move faster than the residual task signal. A verification deficit prevents the auditor from recovering the distinction at the tempo of the population. The fourth closure element supplies the latent capacity constraint; the verification deficit names that constraint becoming binding at population tempo. Under that triad, the population reshapes toward the categories; category names remain while their referents narrow; scores and agreement hold or improve while current-population discrimination falls. Repeated evaluation of a fixed comparison panel helps locate changes in the judge; external task outcomes distinguish genuine improvement from proxy migration. A narrowing score spread alone can reflect successful selection. The fifth note’s discrimination measure is defined over the current population, so the panel and task outcomes supply the additional checks this distinction requires (Ortulanu, 2026e §2).

The first note’s failure conditions are closure seen from the category’s side; the second note’s capacity asymmetry is the verification deficit seen from the auditor’s side; the third note’s contamination is the population loop in operation; the fourth note’s conditions interrupt the loop; the fifth note’s regimes are its tempos. The synthesis adds the missing separation: closure is the topology, erosion the conditional trajectory.

The point-by-point negation deserves its grid. The duality the second note drew between the first note’s failure conditions and its own capacity dimensions (Ortulanu, 2026b §8) extends, through the fourth note’s derivation, to the design conditions (Ortulanu, 2026d §3); the three triples are one structure in three vocabularies.

Failure condition (2026a §2)	Capacity dimension (2026b §2)	Design condition (2026d §3)
Self-reference	Self-reference depth	Lineage — criterion built from disjoint sources
Anchor drift	Semantic reach	Anchor — a non-migratable reference with verification access
Proxy displacement	Bandwidth	Rate — observation at auditee tempo

Read across, each row is one constraint seen from the auditee, auditor, and designer. The third row is the hinge: proxy displacement is what bandwidth deficit looks like from the system’s side — the running question migrates into a standing category when observation cannot keep pace (Ortulanu, 2026b §8). Read down, the columns describe failure, capacity, and resistance. The erosion triad sits between the columns as mechanism, not as a row-wise translation of the grid. Misalignment is a relation between rewarded signal and task property; differential learnability is a property of population response; verification deficit is a relation between evidence and auditor. Each can cross more than one row. In the fifth note’s bookkeeping, β weights the use channel, h and L index anchor and lineage exposure, and c(Pₜ, B) denotes the contraction pressure the benchmark’s categories exert on the current population (Ortulanu, 2026e §3). The erosion triad separates the directional assumptions behind that schematic loss term. Verification deficit is the auditor-side failure to recover the task distinction at the required rate. It can be investigated by pairing the fifth note’s judge floor δⱼ with the fourth note’s rate condition (Ortulanu, 2026d §3; 2026e §4). The two decompositions therefore cut the same mechanism at different resolutions. An inhabitable anchor makes proxy migration possible, rate determines whether the residual is observed before it moves, and lineage determines whether the judge carries an independent correction. Forcing the three mechanism terms into one row each would purchase symmetry by discarding these interactions.

Campbell’s law states that indicators under decision pressure corrupt; performative prediction states that deployed predictions shift their target distributions (Campbell, 1976; Perdomo et al., 2020; Ortulanu, 2026c §3). Goodhart taxonomies separate the routes by which optimization breaks a proxy (Manheim & Garrabrant, 2018). Recursive audit closure specifies a different object: the categorical judgment enters the production of the next judged population, and the auditor’s capacity determines whether the movement remains visible. The erosion triad specifies the sign. Finance supplies the price-layer homologue: risk measures feed balance-sheet decisions, the decisions move prices, and the moved prices return as measured risk; agreement can record coordination on a shared signal rather than accuracy (Adrian & Shin, 2010; Morris & Shin, 2002). MacKenzie’s performativity account established that financial models can become engines of the markets they describe (MacKenzie, 2006). The present framework adds the audit layer: which judgments become institutional gates, what population they produce, and which external evidence can still discriminate the result. The category can remain intact while its power to separate the task property moves.

Once measurement feeds what it measures, it becomes part of the production function. Its validity is no longer a static property of the instrument but a dynamical property of the loop.

3 The Instances, Aligned

The series’ instances differ in century, substrate, and stakes. Aligned against the configuration, they read as one table.

The alignment in summary

Instance	Category in use	Use channel	Markers in the record	Recognized event
Structured finance, 2003–2008 (2026a §3; 2026b §6)	AAA, by output	Ratings as design constraint for new instruments	Reach inversion across asset classes; divergence settled externally	July 2007 downgrade wave
AI evaluation, 2024–2026 (2026c §6)	Pass/rank; saturated-vs-informative, by use	Judgments as training, selection, and release signal	Saturation and judge biases; scope-contraction test still open	No case-wide adjudication in the assembled record
Citation indexing (2026f §2–§5)	Quartiles; high-impact, by use	Editorial optimization; venue selection; institutional reward	Journal–article discrimination weakens; rank–retraction divergence; causality open	DORA — use reform; annual suppressions — channel sanctions
Standardized testing (2026f §6)	Proficiency bands, by use	Instruction reshaped to tested form	Internal gains against flatter audit tests; channel and exposure bundled	Atlanta investigation, 2011
Peer review (2026f §6)	Accept/reject/revise, by output	Verdicts gating careers; channel intruded upon	Boundary case: intent-dominant	Retraction waves from 2014 onward

Read forward, the table says the closure travels: decades, substrates, and incentive cultures vary, yet judgment re-enters production in the same place. The citation record contains a longitudinal weakening of journal–article discrimination, but its cause is open; the remaining chronology is consistent with scope contraction preceding external divergence only at coarse resolution, below what the fifth note’s order requires (Ortulanu, 2026e §6; 2026f §4). Read backward, the boundary cases locate the mechanism. Peer review remains at the edge because intent dominates its channel. The Zcash Orchard episode sits outside the table because the historical-exploitation question runs through observability rather than recursion: the June 2026 disclosure states that the old pool’s privacy properties prevent a cryptographic proof that no counterfeiting occurred before remediation (Wilcox, McGee, & Hornby, 2026). The fourth note reads the same episode for its lineage lesson: review constituted differently from the resident ecology carried different blind spots (Ortulanu, 2026d §5). That lesson is compatible with exclusion here. Discovering the defect and establishing its prior exploitation are different audit tasks. The episode supplies no categorical judgment recursively shaping the next population; the historical question tests observability without closure. Boundaries give the invariant a shape.

Read a third time, for the differences, and the table carries its most expensive information. The columns were chosen for what aligns; the rows also refuse to align, and the refusal is not noise. What the categories answer to differs by domain — deployment behavior, and in the verifiable slice an oracle, for AI evaluation; learning, behind the proficiency bands; solvency, behind the ratings — and that difference fixes whether the fourth note’s anchor condition can be met at all. Where the referent can stand outside the population — a machine-checked proof, a measured outcome, an audit test — the conditions are satisfiable in principle and the design problem is real work. Where market prices enter marked-to-market balance sheets, the actions being assessed can move the reference itself (Adrian & Shin, 2010). Later payments, defaults, and recoveries still supply evidence about contractual outcomes; the design problem is obtaining a relevant independent check at the decision’s horizon. AI evaluation sits at neither pole: its verifiable subdomains have exits that its preference-shaped subdomains lack. The alignment, read at full strength, is one configuration recurring at different depths with different exits — and the difference in exits, not the sameness of the loop, is what a synthesis is for.

The alignment’s value to each instance is mutual. The 2008 case lends the arc its settled forensics — the one row where the external bill arrived and itemized everything. The AI case lends direct reuse of judgments in training and selection. The citation case lends a documented channel and a measured sign reversal. Testing lends the cleanest scissors. Each instance triangulates the others’ reading, which is what a structural claim, short of a fitted model, has for support — stated as such in the first note and unchanged since (Ortulanu, 2026a §8).

4 Where Audit Holds

A synthesis owed its frame’s negative space. The configuration requires four elements; most measurement, most of the time, lacks at least one, and this erosion mechanism is absent there — bounded auditability is the second note’s positive case (Ortulanu, 2026b §5).

The channel can be absent: an evaluation whose verdicts nothing consumes opens no population loop (Ortulanu, 2026e §3). The rewarded proxy can remain aligned with the task property, turning recursive selection into improvement. The anchor can be uninhabitable: where judgments answer to machine-checked proof, a game’s outcome, or a physical measurement, the population can overfit the anchor’s surroundings but cannot move the reference itself, provided the auditor can verify the claimed result (Ortulanu, 2026d §3). The judge can hold evidence the population cannot reproduce. Each condition closes a different door.

The framework’s practical content is an exposure map. Closure is read through channel weight, institutional reliance, and shared substrate. Direction is read through proxy alignment, relative learnability, and verification access. Resistance is read through lineage, anchor, and rate. The three layers prevent a coarse verdict of “recursive” from doing the work of diagnosis.

Cross-domain use also requires a fixed core for each marker. Marker 1 requires loss of discrimination at the relied-upon frontier together with rising cost of restoring an independent distinction; saturation and inflation are possible phenotypes, not substitutes for those components. Marker 2 requires agreement or administrative consistency to hold while operative scope contracts, while rerun stability fails to improve; a scope trace alone is a partial mapping. Marker 3 requires divergence from a task-external consequence, with its causal status stated. The expressions can vary. The measurements cannot disappear in translation.

5 The Frame Question

The second note reserved a question that does not concern any single instance: why the dominant framing of frontier-safety work positions capacity asymmetry outside the space of recognized problems (Ortulanu, 2026b §9). With the arc assembled, the question generalizes, because the main instances show the same reception structure. The skewness result was published in a flagship venue of the science it indicted and the quartile thickened anyway. The accountability literature documented score inflation for decades while the proficiency bands hardened. The ratings critique was older than the crisis that vindicated it. Recognition could come from within the practice and still fail to change its gates. Seglen and DORA make the distinction visible: recognition of the problem and institutional uptake are different events. The first note’s historical asymmetry therefore concerns uptake as well as discovery (Ortulanu, 2026a §7; 2026f §5).

The first note located the sources of the asymmetry for one field: problems are sorted by the frame into what its substrate can address, careers form inside the sort, and the signal-to-noise structure of self-referential evaluation keeps the residual invisible (Ortulanu, 2026a §7). The generalization the table licenses is structural: a field’s evaluation infrastructure is maintained by the population it evaluates, so critiques of the infrastructure arrive as critiques from the evaluated — discountable by construction, payable only in the currency the infrastructure itself denominates. The citation case makes it concrete: an argument against the impact factor gains standing by where it is published. The configuration defends itself with its own category system, without any actor defending it.

What has historically broken the loop is named in the rows. An external bill, where the world supplies one — 2008’s defaults, Atlanta’s investigators. A less inhabitable anchor adjacent enough to expose the gap — the audit test beside the state test. The framework’s expectation for AI evaluation follows: recognition through argument alone is the channel its own analysis rates weakest; recognition through measured divergence — the scissors, run where an oracle exists — is the channel it rates strongest and can specify. The series’ empirical companion is that specification, and the frame question is the reason it exists.

6 The Reflexive Position, Completed

The series has carried a reflexive clause since the second note: any review of these arguments is itself an audit relation, subject to the conditions the arguments name (Ortulanu, 2026b §8). At the arc’s end the clause comes due in full, and it cuts in three directions.

It locates the series. These notes address evaluation infrastructures from outside the practices that maintain them. That position supplies distance and withholds standing at once: the first note’s historical asymmetry applied to its own argument (Ortulanu, 2026a §7).

It disciplines the series. The markers’ joint pattern, conditional order of assembly, alignment counterexample, and resistant-domain survey supply tests of the mechanism. If the vocabulary is absorbed while the tests go unrun, the name displaces the property. If the elements are assigned only after outcomes are known, scope conditions become a protective belt. The fifth note requires assignment before outcomes are known (Ortulanu, 2026e §7). Applied to the alignment distinction here, that means specifying its estimator and threshold in advance, classifying cases blind to outcomes, and testing any later hidden-proxy claim on held-out evidence. The terms travel with their measurements or they should not travel.

The same distinction governs review. Critique from financial regulation, measurement theory, and the sociology of knowledge brings evidence and categories the arc did not produce (Ortulanu, 2026c §8). Internal coherence can be checked from within the vocabulary. External validity requires a reference the vocabulary did not select.

7 Limitations of the Arc

The parts leave four limits on the synthesis.

The evidence base has two forensic-depth instances, one documented-channel instance with a longitudinal trace, and two shallow contrasts. Structured finance and AI evaluation carry forensic detail. Citation indexing documents the channel and a weakening journal–article relation, while causal closure remains absent. Testing and peer review are compressed; performance review is absent. Five rows align; five rows are not a population, and the alignment is structural correspondence, not statistical demonstration — the inference’s character has not changed since the first note flagged it (Ortulanu, 2026a §8).

The fifth note’s regimes remain qualitative, and channel weight is mostly private in the AI case (Ortulanu, 2026e §8). Its factors are not separately identifiable from a single trajectory. The design conditions are derived but not proven minimal, and their joint satisfiability at frontier scale remains open (Ortulanu, 2026d §6).

The evidence is mainly retrospective. The framework’s prospective record begins with the fifth note’s predictions and the sixth note’s resistant-domain test. Those tests now carry the forward claim.

The two-layer distinction adds its own empirical burden. Closure can be observed without erosion, and erosion can be misattributed if judge drift is not separated from population selection. Any application must identify both the closed channel and the source of direction.

8 What Remains

The arc closes by converting its weight into tests. Four were accumulated; none is discharged here.

The Combination Problem. Whether any composite design satisfies the lineage, anchor, and rate conditions together at frontier scale — the fourth note’s open center, on which the possibility of contamination-resistant AI evaluation turns (Ortulanu, 2026d §6, §8).

The Channel Measurement. Whether the weight of benchmark judgments in production updates can be estimated from public observables, or made disclosable — the fifth note’s largest obstacle, and the quantity every prediction is stated over (Ortulanu, 2026e §9).

The Resistant-Domain Survey. Whether categorical evaluation systems that have not eroded under heavy use are the ones the conditions predict — the two-sided test the sixth note specified and left unrun (Ortulanu, 2026f §8).

The Execution. Whether the scissors opens where an oracle exists — internal agreement holding while external validity falls, under quality convergence and judge kinship. The companion track holds the design; the fifth note’s predictions hold the refutation conditions (Ortulanu, 2026e §7).

The arc now makes one claim at the level it can bear. Recursive audit closure turns judgment into production; proxy misalignment, differential learnability, and verification deficit determine whether that closure erodes the task signal. Under the fifth note’s stated erosion assumptions, the markers either follow the predicted sequence or they do not. The series began with a failure that does not announce itself. It ends with the topology that permits the failure, the conditions that give it direction, and the observations that can cut the loop open.

References

Adrian, T., & Shin, H. S. (2010). Liquidity and leverage. Journal of Financial Intermediation, 19(3), 418–437.

Campbell, D. T. (1976). Assessing the impact of planned social change. Occasional Paper Series No. 8. Dartmouth College, Public Affairs Center.

MacKenzie, D. (2006). An engine, not a camera: How financial models shape markets. MIT Press.

Manheim, D., & Garrabrant, S. (2018). Categorizing variants of Goodhart’s law. arXiv:1803.04585.

Morris, S., & Shin, H. S. (2002). Social value of public information. American Economic Review, 92(5), 1521–1534.

Ortulanu, G. (2026a). A category-level failure mode not captured by distribution metrics. osservadore. doi:10.5281/zenodo.19627563.

Ortulanu, G. (2026b). The audit asymmetry problem: Capacity conditions for substrate-independent audit. osservadore. doi:10.5281/zenodo.20092930.

Ortulanu, G. (2026c). The reflexive contamination of categorical benchmarks. osservadore. doi:10.5281/zenodo.21252512.

Ortulanu, G. (2026d). Substrate independence as a structural requirement. osservadore. doi:10.5281/zenodo.21802089.

Ortulanu, G. (2026e). Recursive evaluation collapse. osservadore. doi:10.5281/zenodo.22743710.

Ortulanu, G. (2026f). Categorical integrity in non-AI domains. osservadore. doi:10.5281/zenodo.23856581.

Perdomo, J. C., Zrnic, T., Mendler-Dünner, C., & Hardt, M. (2020). Performative prediction. Proceedings of the 37th International Conference on Machine Learning, 7599–7609. arXiv:2002.06673.

Wilcox, Z., McGee, J., & Hornby, T. (2026, June 4). The Orchard counterfeiting vulnerability—and next steps. Zcash Community Forum.