43. What happens to a claim after it is recorded

Hunt #131, hunts/claim_halflife/. The measurements, the controls and the doors are in hunts/claim_halflife/RESULTS.md. This page is the front door.

Landed on main 2026-10-10. The hunt was written on the branch claude/repo-hunt-content-2smbid, where it was Hunt #122 and this page was numbered 40; on main #122 is dh_minus_landing and 40 is the page on one form per discriminant, so the hunt is #131 and this page is 43. Every number here was re-derived on 2026-10-10 from the history frozen at 2da62eb and is pinned by hunts/claim_halflife/test_claim_halflife.py. Two statements did not reproduce and one has been overtaken; each is marked where it appears.

Grade: measured, and the headline had to be withdrawn, repaired and re-earned on fresh data before it could be stated. This is not mathematics. Nothing here bears on RH (docs/08), and the standing rule in meta/README.md applies: a session with no mathematics and a tidy ledger produced nothing. This page is about the ledger, and the mathematics is elsewhere in the same session.

1. The question

This laboratory publishes its working record, corrections included. That is a policy, and a policy is a claim about behaviour. Nobody had measured the behaviour.

The narrow, answerable version: once a claim is written into the case log, what happens to it? How often is it revised, how fast, and how much of the record has anything mechanical standing behind it. There are 95 entries in that log.

Three reasons it is worth asking, in descending order of how uncomfortable they are. Every HANDOFF.md record in this tree carries a field called Now caught by:, so whether that field is usually filled is a fact and it is checkable. A claim nobody revisited looks identical, from outside, to a claim that withstood scrutiny, so the denominator has to separate them. And a test that names a hunt is presence, not power, which is a distinction the tree's own playbook insists on and nobody had measured.

Not reproduced, 2026-10-10. At the corpus commit HANDOFF.md's header describes the format ("what now catches the problem"), and 3 of its 24 records carry a Now caught by line. The doctrine is stated; the field is mostly absent.

2. What the study nearly got wrong before it started

The container this ran in had a shallow clone. 262 commits reaching back to 2026-08-13, against 1084 reaching back to 2026-08-01. Nothing announced it.

Reconstructed at several depths:

A git-history study run on the default checkout would have reported 30.5% and called it the answer.

This tree has met the hazard before, in August, and its record of that incident ends with "the better fix, prepared and NOT landed": fetch-depth: 0 in the two CI workflows that still lack it. Still not landed. The first thing this study measured was the reason its own subject exists.

Landed since, 2026-10-09. #287 gave both workflows fetch-depth: 0. True when written.

3. What holds

The guard surface is narrow. One recorded claim in five, 19 of 95, is named by any test at all.

Where it exists, it mostly holds. Of the 25 pairs of hunt and test that can be scored by planting faults, 19 defend at least one of the hunt's numbers against a ten percent mutation, and 8 of those are byte-pinned, which is the strongest guard available. Six read none.

That second number is a correction of this hunt's own first attempt. The first ladder mutated only the first artifact each test names, and reported 9 of 24. Twenty of twenty-six tests name more than one artifact, up to thirty-seven, so that figure was measuring the instrument rather than the tree.

Almost nothing is guarded at the moment it is corrected. Of 73 revisions the classifier called strict, one touched tests/ in the same commit.

This survives the classifier's problems below, because it is lopsided enough that classifier error cannot rescue it. Suppose only 40% of the 73 are real, which is the precision the blind audit measured before repair. That is still 29 real corrections and at most one installed guard. Missing revisions does not help either: it means there were more of them.

Correction is fast. Median latency from a hunt's first commit to a revision of its own claim files: one day. Mean 1.8. Longest 13.

4. The number that had to be earned twice

The first version of this hunt reported that 21% of case-log entries were later revised, at three strictness levels that agreed to within five points. Three agreeing levels look like robustness. They are three variations of one author's idea of what a revision is.

Thirty rows, half classified positive and half negative, went to a separate agent with the answer key moved out of the repository, told to judge each commit from its own diff and nothing else.

It agreed on 16 of 30. Precision 0.40, recall 0.55, on a balanced sample.

That is close enough to chance that the rate cannot be reported, and it is reported here as withdrawn rather than quietly rescaled. The audit's row notes named two structural defects rather than borderline judgement calls:

• The hunt's opening was counted as a revision of itself. The first commit was taken to be the first one touching any of five hard-coded filenames, so a hunt that landed a mission statement first and its results second had that second commit scored as a later revision.
• A withdrawal in a differently-named file was invisible. One row adds a file headed "WITHDRAWN / FALSIFIED" and the classifier scored it clean, because the filename was not on the list.

Both are defects on their own terms, so they were repaired rather than tuned around. Re-scored against the same 30 rows the repaired classifier reaches 21 of 30, precision 0.58. That number is fitted, not measured, because those rows are what found the defects.

So a second sample of thirty was drawn from rows the first audit never saw, and sent to a second blind audit with its key again outside the repository.

The repair holds on rows it was not tuned on. That is what makes a rate reportable: of 414 later commits the classifier flags 73, 17.6%, and scaling by the measured precision over recall gives 15.3%. Fifteen of the ninety-five entries carry at least one. Both numbers rest on a classifier that still disputes about a quarter of rows with a careful reader, so they are worth one significant figure and no more.

The original classifier is kept in the tree exactly as it was. A classifier that scored near chance is evidence about how this kind of measurement fails.

5. The instrument broke the thing it was measuring

The planted-fault ladder mutates artifacts in the working tree and restores them. Its first version restored with git checkout --. A concurrent git command in the same tree made one restore fail with exit 128, and a mutated artifact was left behind until someone noticed.

The operating playbook this laboratory keeps already contains the rule: never background a job that does git checkout in the shared tree. It is now obeyed rather than quoted. The ladder holds the original bytes in memory and asserts that the restore round-trips.

Three things in this session went that way. A rule was written down, in a comment or a playbook or a handoff note, and then the same failure happened anyway to somebody who had not read that particular file. The measurement in section 3, one guard installed per seventy-three corrections, is the same shape counted rather than anecdoted.

6. What this does not measure

Git records what a commit did. It does not record what made someone look.

The attribution column here can say that a commit replaced a number and that it did or did not add a test. It cannot say that a test went red, that a sibling session found it, or that a person read the page and stopped. That has to be written down at the time or it is gone, which is exactly why meta/interventions.jsonl has a caught_by field. Nineteen of the seventy-three revisions in this corpus would be worth one line each in that ledger, and none of them has one.

Not reproduced, 2026-10-10. Nothing in the hunt derives nineteen; the ledger at the corpus commit names none of the seventy-three by hash, which is the point that stands.

That is the door, and it costs a habit rather than a system.