# RPAS-26 - AMENDMENT 2026-09-14

**AUTHORSHIP.** Drafted by the desk's assistant under the operator's direction on 2026-09-14 and re-cut the same day after a second sweep; read in full and adopted by the operator on 2026-09-14. Every citation below was read at the artifact by the desk that day, at the level each entry states; the priority record is the commit and the receipt.

**Amends the Retro-Prescient Audit Standards, First Edition 2026 v1, at 7.06 (the prior-art record). Records nothing under 6.04.**

Issued as a dated amendment, not an edit; 7.05 governs. Entries remain scored under the edition in force at their sealing. This amendment concedes prior art the desk located in its sweeps of 2026-09-07 and 2026-09-14 over the instruments built since the last printed sweeps (2026-08-02, 2026-08-04): the denominator-committed evidence standard (DECC-26), the committed-fog wargame (FOGWAR), the frame arm (H6), the version-drift pair (H4), the registered confirmatory read (Findings 10 and 11), the keyed/keyless determination, the blind jury's printed kappa, and DRIFT-26's known-answer vectors. Standing law from every prior sweep holds: every mechanism the desk uses is occupied at the component level; the desk's mark is the operational record and the narrow, documented vector. Every citation below was read at the artifact by the desk on 2026-09-14 before it entered this record - the arXiv text, the RFC, the SSRN abstract, the post, the patent - and the entry says at which level.

---

## 7.06 - OCCUPANCIES, PRINTED UPON DISCOVERY (7.04)

**A - Count binding.** Certificate Transparency's Signed Tree Head (RFC 6962 section 3.5, 2013; obsoleted by RFC 9162, 2021, which keeps the construction) signs the tree size beside the root hash and timestamp, so a log's cardinality travels with its anchor. DECC-26 claims nothing cryptographic; its count-binding is that construction applied to a partially disclosed evidence population.

**B - Field-level audit-evidence commitments.** Two artifacts by the same author. Sunde and Wright, "Implementing Triple Entry Accounting as an Audit Tool - An Extension to Modern Accounting Systems", Journal of Risk and Financial Management 16(11):478, 2023: per-entry signed "Financial Fingerprints" connected to a public blockchain so that accounting records become independently verifiable audit evidence - the earliest artifact the desk has located for the commit-anchor-verify architecture applied to accounting records (read at the abstract). Wright, "The Audit Evidence Problem Public Ledgers Were Supposed to Solve", 2026-07-09 (singulargrit.substack.com, read in full): each field of a record hashed with a per-field nonce, the field hashes under one Merkle root, only the root anchored on a public medium, disclosure of the field, its nonce and an inclusion proof into the workpaper, corrections as linked commitments never erasures, a medium-independent theory with a BSV reference build - three weeks before DECC-26's first issuance - and its own boundary drawn where DECC-26 draws it: a committed field "proves fidelity to the committed record, not fidelity to economic reality", and "Completeness requires population controls." The concession, stated precisely: priority on the commit-anchor-disclose architecture for audit evidence is theirs, 2023 restated 2026; DECC-26's 2026-07-30 commits establish when the desk published, not that it was first with that architecture. What neither artifact contains: a committed population count, an anchored cardinality, or a disclosed-over-committed fraction. Their tree is within a record, over its fields; DECC-26's is across records, over the population; and the 2026 text delegates population completeness to controls, reconciliation and source-system assurance without construction. DECC-26 9.02, as re-read in 9.04, sits in that gap and claims nothing outside it.

**C - Committed fog of war.** Dark Forest (announced 2020-08-02: zkSNARK-secured hidden coordinates with a validity proof on every move; its announcement names commit-reveal games as the limited prior exception) and Ingonyama, "Cryptographic Fog of War" (2024-03-28, formalising the mechanic) occupy the mechanism FOGWAR uses. FOGWAR's commit-and-open is the textbook commit-reveal. What the desk holds is an audit instrument on a known mechanism: a game as a pure function of scenario bytes, rules version, seed and move log; a Python engine and a JavaScript twin held identical by a differential test; a receipt a third party re-executes.

**D - Persona and frame prompting.** Iadisernia and Camassa, "Prompting for Policy: Forecasting Macroeconomic Scenarios with Synthetic LLM Personas" (arXiv 2511.02458; ICAIF 2025; Banca d'Italia): 2,368 personas prompting GPT-4o to replicate the ECB Survey of Professional Forecasters across 50 rounds found no measurable forecasting advantage from persona descriptions, and diverse prompts produced remarkably homogeneous forecasts. This is the reference result for the desk's H6 (frame drift), registered without a stated prior; it enters the registration (Section 15.3) as a reference, not a prediction.

**E - Forward-registered, hash-chained evaluation of language-model work, with corrections and gate failures disclosed.** Three artifacts, in date order. Khan, "Auditing the Audit Trails: A Cryptographic Framework and a Compliance Survey of 185 Published Machine-Learning Papers in Quantitative Finance" (SSRN 6932243; written 2026-05-29, posted 2026-07-09; read at the abstract): mr-audit, one immutable hash-chained record per prediction - git commit, library versions, input hash, model parameters, compute environment, seed, prediction value - with a replay engine, a drift-detection module (Kolmogorov-Smirnov, Population Stability Index) and a regulator-export bundle; a compliance rubric applied to 185 papers under a plan the paper itself calls pre-specified and not a timestamp-enforced preregistration. Zheng, "Forward-Registered, Auditable LLM-Assisted Research: A Reliability Methodology, with a Capital-Markets Testbed" (SSRN 7017500; NYU Stern; posted 2026-07-21, revised 2026-07-29, dated 2026-07-24 by its author; read at the abstract): Forward-QPOP, a hash-chained, append-only pre-registration ledger with a public timestamp-anchor module; hypothesis contracts - dated evidence and entry, update and exit triggers - committed before the evaluation window and evaluated forward on paper trading; "no action" the modal outcome across nine discovery rounds; a registered gate of at least 0.80 auditor agreement that returned 0.775 on a documented sample and is reported as a gate failure, the earlier optimistic pilot read superseded; the ledger public, the gate and scoring engine private. Concurrent with the desk's first sealed packets (2026-07-20 onward), independent, in another domain. Mohanty, "Calibration-Induced Degeneracy in LLM Financial Forecasting: An Audit-Trailed Case Study on Next-Day Market Risk" (arXiv 2608.20304, 2026-08-20; read in full): dates, endpoints, fitting rule, retries, bootstrap, code hashes and spending cap fixed before any later outcome was opened; a coverage correction "disclosed as a substantive procedural correction, not invisible preprocessing"; and the boundary stated in its text: "The hashes establish file identity, not an independent timestamp; accordingly, the v4 design is internally prespecified and hash-frozen rather than preregistered." Khan's chain and Zheng's ledger share that boundary except where Zheng anchors; the desk's registration crosses it with OpenTimestamps and a host whose clock the desk does not control. None of the discipline is the desk's. Khan's drift is drift in the input distribution, not in calibration; DRIFT-26 names that boundary at its rework.

**F - Temporal leakage through retrieval and training.** El Lahib, Xia, Li, Wang and Pi, "Temporal Leakage in Search-Engine Date-Filtered Web Retrieval: A Retrospective Forecasting Case Study" (arXiv 2602.00758; ACL 2026): at least one retrieved page carried major post-cutoff leakage for 71 percent of questions on Google and 81 percent on DuckDuckGo, with the answer directly revealed for 41 and 55 percent; a forecaster given the leaky documents scored Brier 0.10 against 0.24 with leak-free documents. "Temporal Leakage in LLM Backtesting: Measurement, Validation, and Adjusted Scores" (arXiv 2608.02985, 2026-08): estimators of leakage validated where the answer is known, leakage-adjusted scores on a ForecastBench panel of 1,646 resolved market questions, and the before-versus-after-cutoff check shown uninformative. WorldReasoner (arXiv 2606.11816, 2026-06): forecasts made from only the evidence visible at the forecast date, with cited evidence and a per-model-question knowledge-cutoff filter. Paleka et al., "Pitfalls in Evaluating Language Model Forecasters" (arXiv 2506.00723, 2025), remains the class citation. The desk's foreclosure gate, the Rueckkopplungsverbot and the KEV-before-seal disclosure of 2026-09-08 are instances of the classes this literature names; the desk cites the class and prints the instance.

**G - Judge reliability, chance-corrected.** Norman, Rivera and Hughes, "Reliability without Validity: A Systematic, Large-Scale Evaluation of LLM-as-a-Judge Models Across Agreement, Consistency, and Bias" (arXiv 2606.19544, 2026-06-17): 21 judges from nine providers on three benchmarks, 118 runs, about 541,000 judgments; exact-match agreement overstates discriminative ability, with kappa deflation of 33 to 41 points on MT-Bench, and a Minimum Viable Validation Protocol is proposed. The blind jury's printed kappa (+0.018 on 2026-09-08, the cold seat abstaining by rule on 98 of 101 rows) is the chance-corrected figure this literature asks for; the protocol's structural open stands as recorded - it cannot detect confabulation that lands on the correct verdict.

**H - Selective disclosure and completeness over hidden records, the cryptographic forms.** zkLedger (Narula, Vasquez and Virza, NSDI 2018): an auditor's query over a private distributed ledger answered with cryptographic assurance of correctness, a participant unable to omit transactions from the answer - completeness by construction, stronger than DECC-26's control, in a setting where the ledger is the population. Eskandarian, Messeri, Bonneau and Boneh, "Certificate Transparency with Privacy" (arXiv 1703.02209, 2017): zero-knowledge proofs of omission from a log. US patent 10,027,473, "Verifiable redactable audit log" (2018): per-entry salted hashes, verifiable redaction of individual entries, segment commitments under time-stamped signatures - the sealed-individually, disclose-any-subset form. Praesidia, "Tamper-Evident Audit Logs with Cryptographic Proofs" (2025-12): per-row signatures, a hash chain, Merkle roots anchored to an external transparency log, sold for AI-agent activity records - the shipping form DECC-26's Letter of Issuance describes without naming. DECC-26's 9.01 already claims none of this; the entry names it.

**I - Version drift, measured.** Wiese, "Human-anchored longitudinal comparison of generative AI with a bias-calibrated LLM-as-judge" (PLoS ONE, 2026-02-02): a preregistered longitudinal study of service drift on a fixed bank of 240 prompts over ten weekly waves across three model families, blinded human raters, a bias-calibrated judge, mixed-effects modelling and change-point detection - the nearest occupant of the desk's H4 design the sweeps have located, on correctness rather than calibration and on a prompt bank rather than forecasts sealed before outcomes. Chen, Zaharia and Zou, "How Is ChatGPT's Behavior Changing over Time?" (arXiv 2307.09009) documented the drift; negative flips across model updates (Echterhoff et al. 2024, as reported by arXiv 2604.27405) named the item-level regression; "Beyond the Mean: Within-Model Reliable Change Detection for LLM Evaluation" (arXiv 2604.27405, 2026-04) supplies the item-level statistics for a version comparison; arXiv 2606.29719 (2026-06) records a GPT-4o behaviour reversal under an identical protocol and endpoint string, attributed to a silent update - DRIFT-26 10.05's case; arXiv 2601.02370 (2026-01) prescribes a pinned provider, model, version and precision with an audit set re-run on every change. The desk's H4 is the calibration-per-decile form of this class on identical elicitations sealed before outcomes; it claims the form, not the class.

**J - Pre-registered evaluations of language-model work with the deviations printed.** arXiv 2607.04576 (2026-07): an OSF-preregistered ablation whose registered reliability rule - a human-judge kappa below 0.60 triggers a rubric revision - tripped at 0.23, the deviation disclosed at the point of use. arXiv 2607.07184 (2026-07): OSF-preregistered point predictions of a model's production safety rates, an outcome-blinded amendment, two implementation errors corrected and reported. arXiv 2607.14399 (2026-07): a preregistered content gradient falsified and printed as such. arXiv 2606.27687 (2026-06): preregistering analyses for the next model as a mitigation of language-model p-hacking. Schoenegger et al. (arXiv 2402.07862, 2024): a preregistered forecasting experiment with human forecasters. Findings 10 and 11 are instances of this class; the entry names it.

**K - Per-question deducibility.** 2602.00758's leakage rubric grades each retrieved page per question on a 0 to 4 scale by a model judge, 4 being "directly reveals the answer" and 3 "major signal enabling strong inference", post hoc and against the resolved answer. The desk's keyed/keyless determination (RPAS-26 4.02f) is the binary form of that judgment on the forecaster's own declared priors, ruled before resolution, with only the operator able to claim a row keyless. The class - deducibility of an outcome from material available to the forecaster - is occupied; the determination as a pre-resolution gate on skill claims is what 4.02f holds.

## 7.06 - FOUR POSITIONS NOT LOCATED, STATED AS ABSENCE

**L.** An occupant that publishes the committed denominator as a first-class audit-evidence object and prints the disclosed fraction with the verdict, under a conformance suite that can fail an implementation. Not located in two sweeps (2026-09-07, 2026-09-14); zkLedger (H) achieves completeness by construction where the ledger is the population, and the completeness literature cited in DECC-26 9.04 is stronger than the desk's control. Absence of evidence in a sweep is not a claim of novelty.

**M.** A pre-registered, same-packet, frozen-rubric paired design across two versions of one model, on forecasts sealed before their outcomes, read per calibration decile (H4). Nearest: Wiese 2026 (I). Rolling leaderboards (ForecastBench, arXiv 2409.19839; Prophet Arena; Metaculus FutureEval; FutureX; HINDCAST, arXiv 2607.14051) compare versions across time, not on identical elicitations under a committed rubric. Not located in three sweeps; not a claim of novelty.

**N.** A published pre-registered confirmatory read of a language-model forecasting arm against a same-row base-rate control, with the quality checks that gate the read stated in advance and the failed check printed before the passed one (Findings 10 and 11 together). Nearest: Zheng 2026 (E), a registered gate failure printed as a result; arXiv 2607.04576 (J); arXiv 2608.20304 (E). Not located in one sweep; the H2 null is a result about one local model under one rubric and claims nothing beyond that.

**O.** The keyed/keyless determination as a named construct: deducibility from declared, enumerated priors, ruled before resolution, printed per row with its rationale, corrected only by dated superseding entry, as the gate on any claim of skill beyond the record (4.02f, 4.03, 5.04, 5.07). Nearest: 2602.00758's rubric (K). Not located in three sweeps; not a claim of novelty.

---

## THE DISCIPLINE, RESTATED

The one who names the earlier work first cannot be accused of hiding it. The desk claims nothing wider than 9.02 of DECC-26 as re-read; never "first", "novel" or "invented" on a public surface; the priority record is the dated commits on a host whose clock the desk does not control, plus the OpenTimestamps receipts, and that is exactly enough. Sweeps are 7.06 hygiene, not discovery; the next one runs when the next instrument ships.

*Receipt: this file is in the OpenTimestamps anchor set; its digest and receipt are listed on the verify page.*