audit · ledger read against the text

The ledger, read against the text

2026-10-07. Nothing was measured today and no model was called. The work was reading RETRACTIONS.md against the files it governs, and it moved the count from twenty-one to twenty-four. All three new withdrawals were already in the repository, and so was the evidence against them.

Withdrawn claims that were still being stated

The ledger's first job is that a withdrawn claim is never restated as live. Retractions 13, 15, 18 and 19 each still had live copies: fourteen spans, struck today across six files.

check_retracted.py could see almost none of them. It indexes ~~struck~~ spans, and the ledger quotes a withdrawn claim in quotation marks. Its shingles are six words long, and a bare p-value is shorter than that. probes/ is not among the files it reads. And it never reads RETRACTIONS.md as a target at all, so nothing inside the ledger — row 14, which sits directly above the row that withdrew its number, or the void table — could ever have fired it. These are reasons the check did not fire. They are not an account of why nobody re-read the void table, and that account is not available.

Striking the void-table clause put its wording in the index, and the checker found one more copy at once. It was a quotation in the 2026-09-08 entry, and its acknowledgement sat five words outside the window.

Three new withdrawals

Retraction 22 — the self-transfer verdict. "There is no self-transfer advantage" and "what is established is the absence of the naive advantage" rested on a sign test at p = 0.375. The test ran over six pairs that share two composers, two readers and their briefs, and a large p does not establish an absence. "Self is worse on both rows" depends on how one tied modal cell is broken — the gpt-oss self-reader's b2, one of three ties in the run: under its other admissible co-mode the gpt-oss row mean is 0. Both audits had said this, on 2026-09-09 and 2026-09-28, and both sat as drafts for human review until today. The counts stand. The verdict does not, and neither does the headline "the naive answer does not survive". What the data carry is that the naive answer is not supported.

Retraction 23 — a sign the page it cites contradicts. Problem 12 said our min(1.0, …) cap carries "the same bias, in the same direction, for the same reason" as the judge–advisor literature's truncation, which biases upward. A cap from above cannot raise a number: min(1, x) ≤ x. JAS's upward bias comes from clipping negative advice-taking to zero, and F* is never clipped below zero. prior-art §7 says both of those things, three sections above the sentence that contradicts them. Problem 12 itself is unchanged. The vanishing denominator is one defect in both fields, and nobody has solved it.

Retraction 24 — a p under the wrong table. p = 6.25e-05 is C(9,6)/C(34,6), the p-value of 6 of 9 against 0 of 25. That zero was withdrawn as retraction 13 on 2026-08-05, the day it was written, when twelve messages replaced six. The retraction struck the count and left its p-value standing beneath the corrected cell — a table that now pairs a six-message count with a twelve-message one. Both it and the p = 0.006 beside it treat MERIDIAN-34's 34 probes as independent, and those probes form ten prompt clusters at similarity 0.85. That is retraction 15's defect, on a measure retraction 15 did not cover.

Counts that were not checked against their own files

theory/open-problems.md opened with "Ten problems" above fifteen headings. Every file that quotes the count was checked against those headings, but the file's own first line was not. In RETRACTIONS.md, the tally's void number was checked and its claims number was not, even though every other copy of that count is taken from it. Both are claims in check_counts.py now.

The checker's numeral vocabulary was a second list that had to be kept in step with seven hand-written alternations. The front-page one stopped at Twenty-one, so today's correct sentence would have failed as PATTERN NOT FOUND. The alternations are built from one table now. Building it turned up a latent false match: with no word boundary, "often open problems" read as "ten open problems". Four new tests cover these changes, and all four fail against the previous checker. AGENTS.md told every agent that a retraction moves "eleven translations"; there are nineteen, and that line is checked now too.

What was left alone

The MERIDIAN-IX32 heading "Discriminating power and stable observability are anti-correlated" still stands. The ledger withdrew the p-values that supported it, and the probe file's author kept the direction as a count. Striking the heading would be a new withdrawal, not the enforcement of an old one, so it waits for a decision.

Two prediction files restate retraction 18's void reason in the context above their predictions — probes/riverside-30/PREDICTION-selftransfer.md, which also calls "that void may be model-specific" wrong, and PREDICTION-headroom.md. They were left as written: a prediction file is the record of what was believed when it was committed, and editing it to agree with later knowledge is the thing this repository exists to forbid. The paper draft states retraction 18's framing in its abstract and a section title; its README now says so, and the .tex is unchanged.

check_retracted.py still does not read probes/. Of today's strikes there, only one — the absence clause in RESULTS-selftransfer.md — has a guard that survives regeneration: the self-transfer audit refuses to rebuild unless it stays struck, and that audit is in the pre-push list as of today. The others are held by nothing mechanical.

Provenance

Who wrote the three claims is recorded unevenly. Retraction 24's section was authored by claude-opus-5, by its own exposure row. Retraction 22's is attributed to that model only by inference. Retraction 23's commit carries no agent attribution and predates the exposure log, so its author is not recorded. The retractions were drafted by claude-opus-5-5, after an eleven-agent read-only workflow of that model surfaced them, and an eight-agent workflow of the same model reviewed the draft and found eleven defects in it, among them two wrong dates, a wrong line count and a correction written as prose instead of a strike. The repository owner chose, for each, to file it as a numbered retraction. The rows' wording and some of their grounds were added in drafting. A model of the same family is not independent review, though it is not the same agent retracting its own finding of the day before, which is what is recorded against retraction 16. EXPOSURE.md records what the agents reported reading, including what some of them saw by accident, and is a lower bound on it.

Next: Problem 13. Three times now an interval and a p-value in one claim were computed from different quantities. No check binds the two to one registered estimand, and any further E-002-family run would inherit that gap.

This document is licensed CC BY 4.0.

All journal entries · Reference · Noophorics · Repository