[amaybaum]

Requested by the owner (amaybaum, Claude Code session chat, 2026-11-04): "Ok let's update the AGENTS.md file as well, to prevent this from happening again."

It also carries §A.46, moved here from #1101 at the owner's request: "Why do we need 2 separate open PR's for AGENTS.md file changes?", then "Yes please do" (amaybaum, session chat, 2026-11-04). #1101 is closed without merging, and its branch is kept at c97de3dd.

Why. In PR-AUDIT-1 (#1088), evidence landings were reviewed for byte identity, CI and unchanged verdicts. Their references to branch-only content were accepted as "provenance only" without tracing what each one supported. Four of the five landings named branch-only commits: #1089, #1094, #1095 and #1096. #1093 named none. MR-1's Proposition B2 on main (measurement-record-formation/preregistration.md, "proved on paper from #912's proof") rests on a proof held only on #912's kept branch. The landing records treated keeping that branch, and assigning the gap to an Issue, as enough. §A.42 item 2 already forbids this. The step that was missing was a trace of each reference to its claim.

Change. One file, AGENTS.md, in ten commits:

• e95deb1c adds §A.42 item 8;
• 6e381b33 applies the first review;
• 1f66557d applies the re-review;
• 16fad497 reflows one paragraph;
• 0ad4bd0f adds §A.46;
• 25c7d8c0 applies §A.46's review;
• 8fae35fa applies §A.46's delta review (N1, N2);
• be616ffc revises §A.46 to the owner's fix-or-remediate adjustments;
• 5a994209 applies that revision's review;
• e10a9ba2 rewords item 2's last sentence (delta review nit 1).

0ad4bd0f and 25c7d8c0 are #1101's 3a8920a8 and c97de3dd, cherry-picked; only the first commit's message changed, to name this PR. At 25c7d8c0 the §A.46 text equals #1101's head; 8fae35fa then applies N1 and N2.

What the change adds:

• §A.42 item 8, "Trace every off-main reference." Before a record lands, each reference it makes to content not on main is traced to its claim and classified. That includes a commit, branch, PR, Issue, comment, CI run or path, named by id or by number. The classes are:

  • provenance: the claim does not use the referenced content;
  • support on main: the content the claim uses, including the proof of a result proved elsewhere, is on main and supports the claim as it stands there, with any difference from the cited version stated;
  • support off main: a gap.

  The item also states:

  • "provenance only" is a per-reference finding, never a group label;
  • "proved from", "inherits", "uses" or "preserves" another record's result is support;
  • a result, control or verdict that depends on commit order or ancestry uses those commits, and copying their files to main does not make them reachable;
  • a gap closes no later than the record lands, by landing the content or, for a mutable record, by narrowing the claim;
  • keeping the branch or assigning the gap to an Issue does not close it;
  • an unclosable gap lands only by a recorded owner decision; a record on main quotes it with author and time (item 7) and names the unsupported claim, and it is then an exception to item 2 for that claim alone.

• Enforcement line:

  • reviewers check items 1–6 and 8, beside byte identity, CI and verdicts;
  • item 8 applies to every later landing, frozen records included, and changes no frozen record's text or verdict;
  • a gap in an earlier landing is closed by landing the content or by correcting the citing record as item 6 allows.

• Lessons register: one entry recording the cause.

§A.46, "Research records on main and established claims" (from #1101, then revised). Owner decisions (amaybaum, session chat, 2026-11-04). First, the preservation-first decision:

Suspend the proposed removal operation. … Keep existing research records on main. … Classify verification status explicitly. Distinguish independently verified results, partially verified results, exploratory evidence, and known-invalid findings. … Amend AGENTS.md accordingly.

Then the owner raised a concern:

I am concerned that Rule 2 will lead to more issues where main will contain unverified data. We should aim to fix the unverified data, or remove it if not possible to fix.

And then made these adjustments:

• New established claims must be verifiable from main. This should be a firm requirement.
• New research evidence may also land when it is needed for reproducibility or to support existing records, even if the broader scientific programme remains unfinished. Its status and limitations must be explicit.
• Existing unverified claims must have a remediation plan, an accountable Issue and a clear status. Labelling alone is not sufficient to resolve a verification gap.
• Removal must remain a last resort, not an automatic consequence of failed verification. Where possible, narrow or withdraw the unsupported conclusion while preserving the underlying research evidence.
• Frozen records must retain their existing protection. Corrections or withdrawals follow the applicable reviewed-round procedure.

What §A.46 says at the head:

• Opening: main is the authoritative research record. An incomplete programme may have records on main; an incomplete verification of a specific claim is a gap to be fixed.
• 1. Established claims: only when the claim can be verified from main and meets the standard for its kind. A claim landed under an owner exception (§A.42 item 8) is not established.
• 2. What lands: evidence may land when an established claim needs it, or for reproducibility or to support existing records, with its provenance, assumptions, status and limitations stated. Work that meets none of these conditions stays on its branch, tracked by its guiding pull request or Issue (§A.43). For the owner to confirm: this sentence comes from the coordinator's proposal, which the owner approved. A reviewer suggested deleting it as an addition; it is kept, reworded to match §A.43.
• 3. Not established: incomplete, negative and inconclusive results that do not meet item 1's standard. For the owner to confirm (F6): this departs from the literal "negative … results must not be represented as established", so that a verified negative result can be established (§A.29).
• 4. Unverified claims: a claim whose verification is incomplete has a stated status, an accountable Issue and a remediation plan; a label alone resolves nothing. Its record distinguishes the four kinds of result. For a frozen record, this goes in a mutable record on main.
• 5. Remediation: meet item 1's standard, for example by landing evidence, re-running a computation or recording an independent review on main. Otherwise narrow or withdraw the conclusion and keep the evidence. A frozen record is corrected by the applicable reviewed round.
• 6. Removal: a last resort, not an automatic consequence of failed verification. It needs a justification, a dependency audit and separate owner authorization, and a frozen record is removed only by a reviewed round. Manuscript claims follow §A.27 and §A.30.

Reviews of §A.46:

• First review: CHANGES REQUIRED; F1–F9 applied.
• Delta review: CONFIRMED WITH NITS; N1 and N2 applied.
• Combined review at 8fae35fa: CONFIRMED WITH NITS. Its nit N1 (an owner-exception claim is not established) is applied in item 1.
• Review of the revision at be616ffc: CHANGES REQUIRED, findings 1–8. 5a994209 applies all of them except the deletion in finding 6, which is kept as above.
• Delta review at 5a994209: CONFIRMED WITH NITS, nothing blocking. Nit 1 (reword item 2's last sentence) is applied in e10a9ba2.

Not in this PR. The read-only compliance check of the nine PRs merged during #1088 is complete, and its findings are with the owner for a decision on remedies. That covers #1016, #1085, #1086, #1087, #1089, #1093, #1094, #1095 and #1096. No record, manuscript or tool changes here. No checker is added: the trace requires reading each claim, so it is a review obligation, like items 1–6.

Checks.

• No tool pins the changed wording. The only AGENTS.md text a tool checks is §A.35, in edge_rigidity_probe.py.
• Local checks pass: voice, claims, duplicate, artifact placement, staleness, toolchain, CI-gate presence and legacy records.
• The commit messages and this description contain no closing keywords.

Review of §A.42 item 8.

• The first review, at e95deb1c, returned CHANGES REQUIRED with three blocking wording fixes (B1–B3) and six nits.
• 6e381b33 applies B1–B3, N1–N4 and N6. N5, an optional cross-reference, is not applied.
• It also adds that a claim about commit order or ancestry uses those commits, and that copying their files to main does not make them reachable. The MR-1 compliance check found this case.
• The commit message of e95deb1c repeats the "five" overstatement and is left unchanged.
• The re-review, at 6e381b33, returned CHANGES REQUIRED with one blocking fix (R1) and one nit (R2):
  • R1: the ancestry sentence also caught notes that only record where content came from. It now covers a result, control or verdict that depends on commits.
  • R2: item 7 places a quote "in that record", which a frozen record cannot carry. The owner-decision quote now goes in a record on main.
• 1f66557d applies R1 and R2 verbatim.
• The delta review, at 1f66557d, confirmed the wording:
  • Its one blocking finding was a 101-character line.
  • 16fad497 reflows that paragraph with no word changed; whitespace-collapsed text is identical to 1f66557d.
  • Its description nit is applied here.
  • The optional "named by id or by number" nit is not applied.

Head e10a9ba2: CI, then the owner's merge decision. The only change since the confirmed review is the reworded sentence that review proposed.

Gates.

• Independent review: confirmed at 5a994209, plus the reviewer's own wording at e10a9ba2.
• Green CI.
• Owner approval naming the exact head SHA.
• Merge commit only.

🤖 Generated with Claude Code

https://claude.ai/code/session_013T6pfKuv5mi3MXtKjH2tZb