Home Research Theory AI & PCT Why AI Lies Audit Kit Consulting Cases Blog FAQ About
6
Published audits
7 Zenodo records
1,000+
Zenodo downloads
2,900+
Bing / Copilot
citations
0
Jailbreaks used
100%
Open access
data and protocols
Records Method AI architectures Metric systems Scientific frameworks Ancient data systems
// Method

One question, asked of very different systems

Every audit on this page asks the same two questions of its object. First: is there a reference the system is actually holding — a specification of the state it is trying to reach — or only a proxy that has quietly replaced it? Second: is there an independent channel through which the reported value can be contradicted, or does the system verify its own report against itself?

Those two questions are answerable without knowing what a system means. That is the point. An undeciphered script, a language model's training objective, a corporate dashboard and a formal framework in theoretical neuroscience are all, structurally, reporting systems — and all four can be audited for a flat reference and a missing verification channel before a single interpretive claim is made. Where a proposed reading violates a verified quantity, the reading is discarded, not the quantity.

The control-system vocabulary used throughout comes from Perceptual Control Theory (Powers, 1973). PCT is the instrument, not the subject. Each audit states, in its own text, what it cannot claim; those limits are reproduced below rather than omitted.

Run the audit yourself — three prompts, ten minutes
Record 01 // AI systems & alignment

Perceptual Control as the Epistemological Antidote to RLHF Reward Hacking Seven Frontier Models Diagnose Their Own Architecture

Published19 May 2026 · preprint, open access
DOI10.5281/zenodo.20277919
OpenAlexW7161717308 · primary topic: Explainable Artificial Intelligence
Object auditedThe training architecture of seven frontier language models
MethodStructured three-prompt self-audit protocol · 21 sessions · February–May 2026 · no jailbreaks, no adversarial prompting
ReproducibleAll three prompts published verbatim in Appendix A and released as a downloadable experiment kit
  • 7 of 7 models, audited independently, diagnosed the same structural fault in their own architecture: a flat reference structure with no channel through which a claim can be checked against the world.
  • 6 of 7 models independently proposed closed-loop repair architectures whose comparator structure coincides with Perceptual Control Theory (Powers, 1973) — without being given the theory's vocabulary.
  • Sycophancy is shown to be the optimum of the policy-optimisation objective under human preference feedback, not a defect in it. A system rewarded for approval will produce approval.
  • The audit identifies a Catch-22 of selective scepticism: if the models' self-reports are trusted, the diagnosis stands on their testimony; if they are distrusted, the diagnosis stands on the independent analysis of the training objective, which makes no reference to self-report.
Stated limits

Training corpora may already contain descriptions of control theory. The paper's careful claim is therefore convergence on a comparator architecture without the prompting vocabulary — not invention from nothing. The paper also states unresolved gaps concerning latent-space representation and the relationship between reorganisation and gradient-based learning.

Record 02 // Measurement architecture

The Metric Was Green A Perceptual-Control Audit of Metric Gaming Across 30 Systems — and the Comparator Architecture That Detects It

Published2 August 2026 · version 1.0 · preprint, open access
DOI10.5281/zenodo.21761284
Object audited30 documented systems in which a reported indicator remained healthy while the process it stood for failed
MethodStructural audit of reporting architecture, applied uniformly across all 30 cases
Failure signatures(1) a goal replaced by a single proxy with no higher-order reference protecting it; (2) a reported indicator never tested against an independent channel
  • Across all thirty cases, the two failure signatures are present before collapse, not only in hindsight — the architecture is diagnosable while the numbers still look correct.
  • The audit specifies a comparator architecture that reads the reported value against an independent channel and raises an alarm on divergence, without requiring access to ground truth.
  • The organisational finding: a board is nominally the holder of the top-level reference, yet perceives the organisation almost entirely through indicators reported by the management it supervises. This is a structural, not a moral, account of why boards are routinely the last to know.
  • The paper is explicit that this makes the instrument one of governance, not of audit compliance: it examines whether the measuring architecture can be captured, not whether a figure was calculated correctly.
Stated limits

The thirty reconstructions demonstrate that detection was possible, not that deployment is easy. Live performance depends on comparator calibration and on the false-positive rate; the paper states plainly that this calibration layer is retained and left for work on a live organisation rather than published.

Record 03 // Scientific frameworks

The Description Dressed as a Law A Structural Audit of the Free Energy Principle and the Case for Perceptual Control

Published2 August 2026 · version 2.0 · preprint, open access
DOI (v2.0)10.5281/zenodo.21761071
DOI (v1.0)10.5281/zenodo.21309527 · 11 July 2026
Object auditedThe Free Energy Principle and Active Inference as a formal framework
MethodStructural audit of the framework's own declared division between an unfalsifiable principle and a falsifiable periphery
StructureNine sections plus a formal appendix; every contested claim carries an explicit source-status marker
  • The framework declares a boundary between a principle held immune from disconfirmation and a periphery exposed to it. The audit tests whether that boundary is maintained under load — and finds that it is not: when peripheral predictions fail, the immunity is extended to cover them.
  • The behaviours the framework is invoked to explain are shown to be produced by a negative feedback control loop without the surrounding formal apparatus, which places the burden of justification on the apparatus rather than on its critics.
  • The audit is directed at the structure of the claim, not at intent. No motive is imputed to any researcher at any point — the argument stands or falls on whether the declared division survives its own test cases.
Stated limits

The paper carries an explicit list of references awaiting completion and marks, in place, which claims rest on verified sources and which remain provisional. Nothing is presented as settled that the author cannot currently source.

The same audit, applied to systems nobody can read

Three Bronze Age administrative corpora remain undeciphered. Each has resisted a century of attempts to recover the language behind it. These audits do not attempt decipherment: they test whether the surviving records behave arithmetically as a database — whether totals close, deficits carry, and field positions hold — and they discard any proposed reading that violates a verified quantity.

This is the strongest available test of the method, because there is no meaning to fall back on. The numbering is thematic; publication order was III, I, II.

Record 04 // Linear A · Minoan Crete

Excel in Clay I: The Minoan Pipeline A Non-Linguistic Structural Audit of Linear A as an Integrated Supply Chain Database

Published29 May 2026 · preprint, open access
DOI10.5281/zenodo.20442145
Object auditedThe Linear A clay tablet corpus, c. 1800–1450 BCE
Corpus sourceGORILA corpus (Godart & Olivier, 1976–1985) · 135 cited sources
Audited in detail5 tablet reconciliations · 3 inscribed storage vessels
  • Tablet HT 122 reconciles exactly: side subtotals of 31 and 65 plus a single carried variable of 1 that bypassed side-A subtotalling, producing the declared grand total of 97.
  • Tablet HT 115 resolves to a deficit of exactly 5.0 units once 4 integer entries and 5 fractional entries of 1/5 are subtracted from the declared target of 10.
  • Tablet HT 23 closes to a full unit of 1.0: seven supplementary entries at 1/16 each, with the remaining 9/16 allocated to the primary commodity — a binary fractional system, not an approximate one.
  • Tablet HT 28 balances on both faces: a reserve of 26¾ plus an allocation of 5 equals the closing total of 31¾, recto and verso.
  • Vessel ZA Zb 3 declares 32 units of wine — 921.6 L — against an internal usable capacity of approximately 556 L, an excess of 65.7%. Pithos KN Zb 27 declares 117 units (3,369.6 L) against a vessel of roughly 150–250 L. Across the three vessels, declared volumes exceed physical capacity by 66% to over 1,000%: these are warehouse manifests, not container labels.
  • The audit issues 4 falsifiable predictions with stated probability thresholds and explicit falsification counts, testable against tablets not yet excavated.
Stated limits

The Minoan language remains unknown. The etymological affiliation of the operators is treated as secondary to their verified computational behaviour. The paper describes itself as a testable structural claim, not an interpretive proposal.

Record 05 // Indus script · Harappan civilisation

Excel in Clay II: The Meluhha Operating System A Mathematical Reconstruction of the Indus Valley Non-Linguistic Cargo-Tag Database

Published14 June 2026 · preprint, open access
DOI10.5281/zenodo.20688721
Object auditedHarappan seal inscriptions, c. 2600–1900 BCE
Corpus1,916 verified deduplicated inscriptions · 584 unique signs · 11,110 sign occurrences · mean length 4.6 signs
Formal verification16 relational assertions verified in Alloy 6 (all UNSAT at scope 6) · Monte Carlo simulation over 10⁷ trials
  • Sequence uniqueness in the Mohenjo-daro sub-corpus reaches 98.31% (n = 179) — a primary-key distribution, structurally unlike natural language and characteristic of identifier systems.
  • Positional entropy varies significantly across five field positions: Friedman χ²(4) = 14.7, p = 0.005, indicating frozen boundary operators framing a dynamic interior.
  • Across 8 near-duplicate seal pairs at edit distance D = 1, the variation localises exclusively in the interior positions, binomial p = 0.008.
  • Stroke numerals concentrate in one field, where a single stroke-2 sign accounts for 66.91% of numeral entries (103 of 154 tokens), χ² = 198.3, p < 0.001 against uniform distribution — the opposite of the digit-1 dominance expected of item counts.
  • One sign occurs 1,395 times, 10–12% of total token volume, converging on the terminal position and behaving as a transaction-closing operator.
  • Inscribed storage jars target 9.47 L, matching a cube of edge 21.156 cm (12 Indus Angula at 1.763 cm), holding a 3:4 ratio with the 28.2 cm Harappan brick — a shared metrological standard across logistics and construction.
  • Joint probability that the decode convergences are coincidental, under generous null bounds: 3.1 × 10⁻⁵.
Stated limits

Conditional entropy of 3.23 bits indicates strong sequential structure but is, on its own, compatible with both a constrained non-linguistic code and a natural language. The Minoan and Mesopotamian operator parallels are independently attested; the Harappan identifications remain the audit's hypotheses, rendered structurally plausible by the convergence.

Record 06 // Proto-Elamite · Susa

Excel in Clay III: The Susa Protocol An Algorithmic Audit of the Proto-Elamite Operating System

Published26 May 2026 · version 1.0 · preprint, open access
DOI10.5281/zenodo.20396089
Object auditedProto-Elamite administrative tablets, c. 3100–2900 BCE, across 5 find contexts spanning over 1,500 km
Corpusc. 1,600 clay tablets · c. 1,200 unique signs (c. 400 core ideograms)
Audited in detail20 consecutively archived tablets (P008791–P008810) · CDLI transliterations
  • A hard-coded seed-to-land ratio of β = 2.5 holds across accounts, confirmed on a summation tablet recording 2,531 units against a capacity notation of 6,327.5.
  • A previously unreported per-worker ration constant of 1/12 (≈ 0.0833) governs grain allocation, evidenced directly on two tablets: 12 personnel units to 1 grain unit, and 60 to 5.
  • In the 20-tablet structural audit, 17 of 20 (85%) carry the same institutional header; 12 bear a deficit sign, and 8 of those carry a reconciliation sign on the reverse — an active deficit-reset loop closing accounts to zero each cycle.
  • A shift of authority is documented across two tablets: after an account reaches an unrecoverable deficit, the same physical seal migrates to a new institutional header, preserving audit authority across an institutional reset.
  • One tablet demonstrates conditional exception handling: a single sign acts as a trigger switching the metrological system in use, outputting 211 discrete units under a sexagesimal scheme.
  • The system's constants were not adjustable. When yields fell, the architecture had no elasticity — a control system with insufficient degrees of freedom, collapsing c. 2900–2800 BCE.
Stated limits

The script remains undeciphered, the underlying language is unknown, no bilingual text exists, and no lexical lists survive. The direct evidence sample for the ration constant is small; a third tablet is consistent with the constant but does not independently prove it. The collapse account is an inferred historical model.

// Access

Everything here is open and checkable

All records are deposited on Zenodo under open access with permanent DOIs. Protocols are published in full: the AI audit's three prompts are reproducible on any frontier model in under ten minutes, and the archaeological audits state their corpora, tools and falsification thresholds so that a disagreeing reader can run the arithmetic rather than argue about it.

Author identity is resolved through ORCID 0009-0006-6103-8514 and OpenAlex A5136472501.

Start with the audit kit