# FRONTIER SWEEP - 2026-09-14 (the 2026-09-14 sweep)

**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 section 10 (the second and third passes of the same day) added at adoption. Citations are named at artifact level here; the 7.06 record (RPAS-26 amendment 2026-09-14) carries those read at the artifact. Served and anchored so that the priors it names are printed upon discovery (RPAS-26 7.04).

Scope: what was built or first exercised since the 2026-09-07 sweep - the instrument shell (every tool page a workpaper with a re-performance receipt), the first pre-registered confirmatory read of an LLM forecasting arm against a base-rate control (Finding 11), DRIFT-26's known-answer vectors for an evaluation standard, the blind-jury clerk pass with its disposition sheet and the cold seat's measured reliability, the same-day-window seal disclosure, and the keyed/keyless determination doctrine as it was applied at scale (939 rows in one pass). Web swept 2026-09-14. Standing law from every prior sweep holds: every mechanism is occupied at the component level; the desk's mark is the operational record and the narrow, documented vector. Citations below are named at artifact level and want paragraph-level verification before any enters 7.06.

## 1. Audit-trailed, hash-frozen LLM forecasting evaluation with disclosed corrections
- Occupied, and close kin: arXiv 2608.20304 (Aug 2026), "Calibration-Induced Degeneracy in LLM Financial Forecasting: An Audit-Trailed Case Study on Next-Day Market Risk" - a prespecified, hash-frozen design (dates, endpoints, fitting rule, retries, bootstrap, code hashes, spending cap fixed before any outcome was opened), a coverage gate that failed on the first pass, and a procedural correction disclosed "as a substantive procedural correction, not invisible preprocessing"; the paper itself notes its hashes "establish file identity, not an independent timestamp", so the design is "internally prespecified and hash-frozen rather than preregistered". That last sentence is the exact boundary the desk crosses with OpenTimestamps and a host whose clock it does not control.
- What survives as the desk's own: none of the discipline is novel; the desk's mark is that the registration, the estimator's hash, the checks and the read are on a public, externally anchored record, and that the null (Finding 11) was printed under the registration's own preconditions. Cite as the nearest occupant of "audit-trailed LLM forecasting evaluation".

## 2. Temporal leakage and the same-day window
- Occupied, directly: arXiv 2602.00758 (Feb 2026), "Temporal Leakage in Search-Engine Date-Filtered Web Retrieval: A Retrospective Forecasting Case Study" - measures leakage through date-filtered retrieval into LLM forecasts, with an LLM judge for leakage scored against two annotators (0.85 QWK) and a Brier comparison by leakage level. Paleka et al. 2025 (arXiv 2506.00723) remains the class citation.
- Consequence for the desk: the searched juror (seat A) and the unattested arms both retrieve; the desk's foreclosure gate and the KEV-before-seal disclosure of 2026-09-08 (four arm rows priced an addition released three hours before seal) are instances of the leakage classes this literature names. 7.06 cites the class; the sitting record already prints the instance.

## 3. LLM-as-judge reliability and the cold seat's kappa
- Occupied: arXiv 2606.19544 (Jun 2026), "Reliability without Validity: A Systematic, Large-Scale Evaluation of LLM-as-a-Judge Models Across Agreement, Consistency, and Bias" - 21 models, three benchmarks, 541,000 judgments; the finding that raw agreement overstates discriminative ability and chance-corrected measures (Cohen, Krippendorff) are required. arXiv 2512.01786 ("Who Judges the Judge? LLM Jury-on-Demand") and JudgeBench (arXiv 2410.12784) occupy the multi-judge construction.
- Consequence for the desk: the blind jury prints kappa on every sitting (+0.018 on 2026-09-08, the cold seat abstaining by rule on 98 of 101), which is the chance-corrected figure this literature asks for; the clerk pass (load-bearing citations verified at primaries before adoption) is the human-validation loop the guides describe. The structural open stands as recorded: the protocol cannot detect confabulation that lands on the correct verdict. Nothing here is the desk's; the desk's mark is the printed kappa and the printed abstention rule.

## 4. Known-answer vectors for an evaluation standard (DRIFT-26 Chapter 13)
- Occupied, three ways: (a) the Image Biomarker Standardisation Initiative (IBSI) - digital phantoms with benchmark values so that independent radiomics implementations agree on computed features, the mature form of the idea; (b) arXiv 2607.00297 (EPC, 2026) - "a reference test suite ... verifies that independent implementations produce identical gamma and JSD values on fixed input sequences ... passing the conformance suite qualifies an implementation as EPC-v1.0-compatible", a conformance suite for a statistical protocol, published weeks before DRIFT-26 rev0; (c) US patent 12645555, "Systems and methods for deterministic conformance suites, validator implementation certification, and verifiable certificate registries for replay-verifiable conformance receipts" - deterministic test vectors with expected validator outputs (PASS/FAIL/HOLD), a conformance certificate committing to suite digest and validator digest. The patent occupies the certification-registry layer, not the vector idea itself, which is decades old (NIST CAVP known-answer tests).
- Consequence: DRIFT-26 13.02's sentence that reproducing the vectors lets an implementer compute the standard "without trusting the reference build or its author" overstates (the assistant's rework note of 2026-09-07 stands); the vectors pin the reference's behaviour. Chapter 13 should cite IBSI and EPC as the form, and the standard should say what the 09-07 note says: the one non-circular result is 9.01a. The patent is cited as a boundary the desk does not enter: the desk issues no certificates and keeps no registry.

## 5. Prediction-market comparators (H5)
- Occupied: Kalshibench (arXiv 2512.16030, Dec 2025), epistemic calibration of LLMs scored against prediction-market prices; HINDCAST (arXiv 2607.14051) as recorded in the 2026-09-07 sweep. The H5 design memo (STANDARDS_DRAFTS_2026-09-14.md, section 6) cites both and builds neither.

## 6. The workpaper shell
- Occupied at the level of principle: audit documentation standards (AU-C 230; PCAOB AS 1215) require that an experienced practitioner can re-perform the work from the documentation; reproducible-research practice (computational notebooks, replication packages) requires the same of a computation. Not located: a public web instrument that prints, on every tool page, a re-performance footer ending in a copyable receipt of the served bytes. One sweep; not a claim of novelty; the shell is a presentation of two old disciplines, and the standard says so.

## 7. Not located in this sweep, stated as absence
- A published pre-registered confirmatory read of an LLM 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: 2608.20304 (section 1). Absence in one sweep is not novelty; the H2 null is a result about one local model under one rubric and claims nothing beyond that.

## 8. What this sweep changes
- 7.06 (operator reworks): three occupancies to print - 2608.20304 (audit-trailed LLM forecasting evaluation), 2602.00758 (temporal leakage), 2606.19544 (judge reliability, chance-corrected) - and two for the standards' own chapters rather than 7.06: IBSI and EPC into DRIFT-26 Chapter 13, US 12645555 as a boundary note.
- DRIFT-26 rev.1: Chapter 13 gains the citations above and the 13.02 correction.
- Registration: nothing; Section 15 (drafted 2026-09-14) already carries the H6 reference result.
- Nothing here changes a law, a floor or a sealed row.

## 9. Not-thievery, as procedure (unchanged from 2026-09-07)
Cite the prior that beat you by artifact, date and boundary quote; claim nothing wider than the standard's own narrow clause; never "first", "novel" or "invented" on a public surface; the priority record is the dated commits plus the receipts and that is exactly enough; cite the artifact, not the author; do not contact the prior's author. Printing priors upon discovery (7.04) is the whole defense.

## 10. Second and third passes, 2026-09-14 (added at adoption)
- Second pass, nine searches, every hit read at the artifact; printed as RPAS-26 amendment 2026-09-14 items H-K and the absences L-O. Occupied and new since sections 1-7: the cryptographic completeness forms (zkLedger, NSDI 2018; Certificate Transparency with Privacy, arXiv 1703.02209; US 10,027,473; Praesidia 2025-12); an earlier artifact for the commit-anchor-disclose architecture by the same author as section 4's proposal (Sunde and Wright, JRFM 16(11):478, 2023); version drift measured (Wiese, PLoS ONE 2026-02-02, preregistered fixed prompt bank; arXiv 2604.27405; arXiv 2606.29719; arXiv 2601.02370; Chen, Zaharia and Zou); pre-registered LLM evaluations with printed deviations (arXiv 2607.04576, a registered reliability gate tripped and disclosed; arXiv 2607.07184; arXiv 2607.14399; arXiv 2606.27687); per-question deducibility (2602.00758's 0-4 leakage rubric; arXiv 2608.02985; WorldReasoner, arXiv 2606.11816). Section 1's two unread pointers were read at the abstract the same day: Zheng, SSRN 7017500 (Forward-QPOP, a hash-chained pre-registration ledger with a public timestamp anchor and a registered gate failure printed as a result, posted 2026-07-21) and Khan, SSRN 6932243 (mr-audit, one hash-chained record per prediction; a plan the paper calls pre-specified, not timestamp-enforced).
- Third pass, two searches on the two corners most at risk, the committed denominator and the version pair. Adjacent only: IACR ePrint 2026/994 (proof-carrying output with selective disclosure of proofs until audit; no count); Metaculus FutureEval and ForecastBench plot scores against release dates, not paired elicitations. Nothing located against L or M.
- What this changes: the four constructs of NOTES_2026-09-14.md carry their nearest occupants in their Occupied paragraphs as of adoption; DRIFT-26 14.01 names Wiese, 2604.27405 and the input-drift boundary; DECC-26 9.04 is due one sentence naming the 2023 artifact and the 2026-09-14 sweep at the operator's next edit. No law, floor or sealed row changes.