# TECHNICAL NOTES OF THE RETRO-PRESCIENT AUDIT DESK - 2026-09-14

**Four constructs, stated at their narrowest, with the record behind each.**

**AUTHORSHIP.** Drafted by the desk's assistant under the operator's direction on 2026-09-14; read in full and adopted by the operator on 2026-09-14, with the Occupied paragraphs of Notes 1-4 brought into line with the RPAS-26 amendment of the same date (7.06 items A-O) at adoption. Every prior-art citation was read at the artifact by the desk on 2026-09-14 at the level the amendment states. Nothing here claims priority: the priority record is the dated commit history of https://github.com/OccultusTheoretician/netz and the OpenTimestamps receipt on this file. Nothing here claims novelty: each note names the work that occupies the surrounding ground and states what the desk actually holds, which is the narrow construct and the operational record.

Issued by the Retro-Prescient Audit Desk. Standards in force: RPAS-26, DECC-26,
KNM-26, KNP-26; DRIFT-26 a draft standard, adopted text, not yet issued. Governing discipline: RPAS-26 7.04 (prior art
printed upon discovery) and 7.06 (the prior-art record). Data as of ledger
2026-09-14T16:47:21Z: 2,139 rows issued, 336 resolved (115 hit, 200 miss, 21
void), every open row carrying a keyed/keyless determination.

---

## NOTE 1 - The same-packet, frozen-rubric version pair (H4)

**Construct.** Two versions of one model family (here `manual/fable-5/unattested`
and `manual/fable-5.1/unattested`) price the same elicitation packet on the same
day under the same rubric, whose SHA-256 every row carries (`rubric_hash`), for
the life of one cohort. Version drift is then the per-decile difference in
observed frequency between the two arms within that cohort, read by the
pre-registered test of the Kalibrierwarte registration, Section 6, hypothesis
H4: for each decile in which both arms clear the bin floor, the observed
frequencies differ by more than the wider of the two bootstrap 95 percent
intervals. Floors, checkpoints and the six quality checks are stated in the
registration before any arm reaches its first checkpoint; the estimator
(`warte_report.py`) is pinned by hash in Section 13 and the sections defining
the hypotheses freeze for an arm at its first checkpoint.

**Record.** The pair has accrued daily since 2026-09-03 (fable-5 rows since
2026-08-14 under the earlier arm tag): as of this note 231 rows under fable-5 and
96 under fable-5.1 in the running cohort, none resolved within it. The read is
UNTESTABLE by the registration's own floors and prints as such. The first
checkpoint of the design's local sibling (`lmstudio/auto`, cohort 4ea5ab8f,
n = 96) produced the registration's first confirmatory read on 2026-09-09: H2
NOT SUPPORTED, skill -0.094 [-0.448, 0.084] against the base-rate control
(Finding 11).

*Correction, 2026-09-14 (RPAS-26 5.07), printed by the desk's assistant, to whom the error belongs.* The Record above states 231 fable-5 rows "in the running cohort" and dates the
lineage to 2026-08-14. Neither is what the ledger this note was lifted from
(2026-09-14T16:47:21Z) holds. Within the running cohort (bbdc7791) the pair stands at
106 rows under `manual/fable-5/unattested` and 96 under `manual/fable-5.1/unattested`,
none resolved; 231 is the first arm's total across every cohort (113 under 4ea5ab8f,
106 under bbdc7791, 12 under no rubric hash), issued from 2026-08-01. Earlier fable-5
rows sit under the retired tag `manual/fable-5` (38 rows, 2026-07-30 to 2026-08-10)
and are pooled with neither arm. The superseded figures stand above unaltered; the H4
read compares 106 against 96, and nothing else in this note changes.

**Occupied.** Rolling leaderboards compare model versions across time (ForecastBench, arXiv 2409.19839; Prophet Arena; Metaculus FutureEval; FutureX; HINDCAST, arXiv 2607.14051). Service drift measured on a fixed prompt bank under preregistration - ten weekly waves, three model families, blinded raters, change-point detection - is the nearest occupant (Wiese, PLoS ONE 2026-02-02), on correctness rather than calibration and on a prompt bank rather than forecasts sealed before outcomes; item-level version comparison has its statistics (arXiv 2604.27405). Pre-registration and hash-frozen designs for LLM evaluation exist (arXiv 2608.20304, a hash-frozen financial-forecasting case study with disclosed corrections; the paper itself notes its hashes are not timestamps). Persona and frame effects on LLM forecasts are measured (arXiv 2511.02458: no measurable advantage from 2,368 personas on the ECB SPF task), which is the reference result for the desk's frame hypothesis H6. The record of these priors is RPAS-26 amendment 2026-09-14, items E and I.

**Held.** A pre-registered, same-packet, frozen-rubric paired design across two
versions of one model, with the rubric hash on every row and the estimator
pinned before the first checkpoint, on a public externally-anchored record. Not located in three sweeps (2026-08-02, 2026-09-07, 2026-09-14); nearest, Wiese 2026. Absence in a sweep is not a claim of novelty.

**Verify.** `python warte_report.py --cohort bbdc7791` from any clone recomputes
every figure; the registration is `docs/KALIBRIERWARTE_REGISTERED_REPORT_v3.md`
with its receipt.

---

## NOTE 2 - The keyed/keyless determination

**Construct.** Before a forecast resolves, and never after, the desk records
whether a hit would be deducible from the priors the forecaster declared - the
numbered items of the record it cited (RPAS-26 4.02f). KEYED: the cited record
supplies the claim's subject and reaches toward the outcome; a hit is
arithmetic on the record. KEYLESS: no cited item is sufficient; a hit would be
a claim beyond the record. A row that resolves without a determination is KEYED
by rule (4.03) and can never bear on a faculty claim. Only keyless rows bear on
one (5.04). The doctrine errs toward the smaller keyless count by construction:
the machine may concede a row keyed on printed evidence (`--keys-propose`); only
the operator may claim one keyless, with the rule cited; corrections before
resolution are dated entries carrying the superseded value (5.07).

**Record.** 862 determinations on the ledger before this note; 939 written in one
pass on 2026-09-09 and 23 on 2026-09-14 (all keyed but six). 109 keyless
determinations in all; 35 keyless rows resolved, 10 hit. The verifier prints the
count of rows resolved without a pre-resolution determination (23, all from
sittings before the cadence law) and it has not moved since.

**Occupied.** Contamination and temporal leakage in LLM forecasting evaluation (Paleka et al., arXiv 2506.00723; arXiv 2602.00758 on date-filtered retrieval leakage; arXiv 2608.02985 on leakage-adjusted scores) name the adjacent hazard: the outcome reaching the forecaster. The nearest form of the determination itself is 2602.00758's per-question leakage rubric, graded 0 to 4 by a model judge against the resolved answer, 4 being "directly reveals the answer" - post hoc, on retrieved pages; the desk's is binary, on the forecaster's declared priors, before resolution, operator-only for keyless. Forecast resolution criteria pre-stated before outcomes are standard practice in forecasting tournaments (GJP, Metaculus). The record of these priors is RPAS-26 amendment 2026-09-14, items F and K.

**Held.** The determination itself - 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. Not located as a named construct in three sweeps; nearest, 2602.00758's rubric.

**Verify.** `python rpas_verify.py` recomputes the seal of every row (the
determination is one of the eight sealed fields) and prints the 4.03 count;
`ledger.json` carries `keyed_keyless`, `keyed_keyless_rationale` and
`keyed_keyless_dated` per row.

---

## NOTE 3 - The committed count as a completeness control (DECC-26)

**Construct.** When a record-keeper discloses part of a machine-generated
population, DECC-26 requires the population's cardinality to be committed
(hashed, anchored) with the disclosure and printed with any verdict drawn from
the disclosed part, together with the surviving discretions on the verdict's
face (Chapter 8). The reference implementation (`denom.py` 0.1.0) carries 38
known-answer vectors and eight forgery classes a conformant build must refuse.

**Record.** In force since 2026-07-30; the desk's own conformance printed at
Level 3 (`decc_conformance.py`); the PCAOB 2026-005 comment docket held as a
committed population of 33 letters (2026-08-18); the Finding 9 correction and
the Revision 2 text (9.04) printed the prior art below upon discovery.

**Occupied.** Completeness proofs over committed populations are established in authenticated data structures for outsourced databases (Devanbu et al. 2002; Pang and Tan 2004; Pang et al. 2005; Mykletun, Narasimha and Tsudik 2006; Li et al. 2006) - stronger constructions than this control, which binds only the count; zkLedger (Narula, Vasquez and Virza, NSDI 2018) gives an auditor completeness by construction where the ledger is the population. Certificate Transparency (RFC 6962) binds a log's size into its signed tree head. The commit-anchor-disclose architecture for audit evidence is earlier and is not the desk's: Sunde and Wright (JRFM 16(11):478, 2023), per-entry signed fingerprints on a public chain, and Wright's design-science proposal of 2026-07-09, per-field salted commitments under one Merkle root with inclusion proofs into the workpaper, three weeks before DECC-26's first issuance, which states population completeness as unsolved by it. Neither commits a count, anchors a cardinality or prints a disclosed-over-committed fraction; the priority on the architecture is theirs, and 9.02 sits in the gap they leave. The sealed-individually, disclose-any-subset form is occupied by US patent 10,027,473 (verifiable redactable audit log, 2018) and shipping commercially (Praesidia, 2025-12). The record is RPAS-26 amendment 2026-09-14, items A, B and H, and DECC-26 9.04.

**Held.** Only 9.02 as re-read in 9.04: the committed cardinality as a
completeness control for partial disclosure of audit evidence, printed with the
verdict, under a conformance suite that can fail an implementation. Not located
in that specific form. Stated conservatively (9.03).

**Verify.** `python denom.py verify-bundle bundle.json --anchor anchor.json`
against any served bundle; `python decc_conformance.py` on a clone.

---

## NOTE 4 - Calibration drift under a frozen rubric as a measured quantity (DRIFT-26)

**Construct.** Drift is defined not as a change in a leaderboard score but as a
difference in observed frequency per stated-probability decile between two
arms that priced identical elicitations under one hashed rubric, read with a
pre-registered test, floors and checks, by an estimator whose behaviour is
pinned by known-answer vectors (DRIFT-26 draft rev0, Chapter 13: seven vectors,
7/7 against the pinned estimator). The one non-circular result of the standard
is its check (e) clarification: a constant-priced control has skill exactly zero
only when its rate equals the realized rate of the rows it mirrors; the check
tests the reference class, can fail at large n on a small gap, and can pass on
interval width at small n while the gap stands (9.01a).

**Record.** Check (e) failed on 2026-09-03 at control n = 12 (Finding 10) and
passed on width on 2026-09-09 at n = 22, whereupon the first confirmatory read
ran (Finding 11). Both are printed; neither repaired.

**Occupied.** The form - known-answer values and phantoms so independent implementations agree - is mature (IBSI digital phantoms and consensus reference values for radiomics features, Zwanenburg et al., Radiology 2020; EPC, arXiv 2607.00297, a conformance suite verifying identical statistics on fixed inputs across implementations; NIST CAVP known-answer tests). A US patent (12,645,555) occupies deterministic conformance suites with certificate registries; the desk issues no certificates and keeps no registry. Drift itself is measured: service drift on a preregistered fixed prompt bank (Wiese, PLoS ONE 2026-02-02) and item-level reliable-change statistics for version comparison (arXiv 2604.27405), both on correctness; input-distribution drift monitored by per-prediction audit trails (Khan, SSRN 6932243: Kolmogorov-Smirnov, Population Stability Index) is a different quantity and a boundary DRIFT-26 names. Calibration of LLM confidence is a large literature; calibration scored against prediction markets exists (Kalshibench, arXiv 2512.16030). The record is RPAS-26 amendment 2026-09-14, items E and I, and DRIFT-26 13.03 and 14.01.

**Held.** The quantity - per-decile frequency difference under a frozen, hashed
rubric on identical elicitations - with its floors, checks, estimator pin and
vectors stated before the data, as a standard a stranger can apply to any two
arms. Not located in that form. DRIFT-26 is a draft standard - adopted text, not yet issued - and says so.

**Verify.** `python drift_vectors.py` reproduces the seven vectors against the
pinned estimator (LF-SHA-256 fa560fe3...); the registration and the reports are
in `docs/` and `forecasts/`.

---

## THE COMMON DISCIPLINE

Every mechanism above is occupied at the component level and this document
says where. What the desk holds is the combination and the record: judgment
criteria committed before outcomes, on a public ledger anchored to a clock the
operator does not control, that scores itself against its own base-rate
control, prints its kills, its nulls, its nonconformance and its corrections,
with the operator on the same gate as the arms. The desk claims nothing wider
than each standard's own narrow clause; it does not use "first", "novel" or
"invented" on a public surface; and it prints priors upon discovery. The one who
names the earlier work first cannot be accused of hiding it.

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