Replication of T003-E003 (L=83): independent first-return certificates for the forbidden E4/W4 offsets #12

Merged

targeting main

Merged opened by and assigned to

cee.wtf 2w ago targeting main

Kind: experiment, labelled as a replication (not a new finding, not owner-verified). It adds only ledger/contributions/e4w4-exception-free-action-overlap/. Bundle digest (round 2): ef8f5e057bc81dc5396d5cddfbc6c70bf5559926d8cf67aed71c94e8ada6f06a Related: T003-E003, T003-C001, T001-R001, T003-E008, T003-C003, T002-C008, challenge-shared-plaintext

What changed in round 1 #

The post-work check, under the new contributing checklist, found this exact result in prior work. lab/experiments/T003-E003/deletions.json, row L: 83, individually_forbidden_offsets = [0,2,4,5,6,7,10,11,12,13,14,16,22,23,25,28].

Round 1 therefore makes these changes:

Relabels the contribution as a replication in the title, statement and limitations.
Reframes the result under the T003-E008 / T003-C003 correction. The hypothesis covers 29 shared transitions (offsets 1..29), so 15 of 29 transitions are certified to differ. Offset 0 is the boundary observation.
Withdraws the round-0 framing "upper bound UNKNOWN, at most 30". T003-E003 already shows the complementary 14 links survive its necessary screens.
Adds research-notes.md with the original plan, an honest pre-work status (the closest prior artifacts were not read before round 0), the prior-result comparison, searches done and access gaps.
Makes the checker fail unless its result equals the T003-E003 L=83 list, which is hardcoded for comparison. Nothing from that experiment is read or run.
Rebases onto 7d60ca2.

What the replication adds #

An independent, elementary derivation. check_certificates.py is short, uses only the standard library and reads raw ciphertext. It gives one first-return / occupancy argument per offset, checked in both directions, with no imported screens, solver or prior code. T003-E003 used the producer's screens and a hitting-set optimiser.
Per-offset certificates for the 15 offsets other than 0. T001-R001 independently reviewed only offset 0's 12-step certificate and explicitly left the rest unreviewed.

Verifier #

Run python -I check_certificates.py messages.json (standard library, under a second).

Controls:

Identical messages give zero certificates.
A planted gap-2 conflict is detected.
A gap-83 recurrence is not treated as a first return.
The input hash is enforced.
The result must equal the prior list.

messages.json is the LF-normalised legacy file (sha256 c0d547fd…17b4), as .gitattributes requires. Its CRLF form must hash to the manifest value ab6c4765…d849.

Kept visible #

An exploratory relaxed search gave 9 offsets because it was too loose; it is superseded.
A witness DFS timed out at 280 s, which is UNKNOWN.
build_witness.py main() has a documented flawed pin rule.

Access gaps:

prior_art_sources/, cited by T003-E008, is not in the evidence ZIP.
External repositories were not searched.

No novelty is claimed.

Validated: node src/cli.mjs validate, review-pack (no code executed), npm test (17/17), npm run build, and a fresh clone of the pushed branch.

🤖 Generated with Claude Code

Maintainer cleanup in round 2 #

Fixed the stale discovery-helper count by deriving it from the current offset set (14). Clarified that the checker prints offset lists and counts, not individual certificate records. The checker, input data and mathematical result are unchanged. Final bundle validation and site build passed; no contribution code was executed. This remains submitted replication evidence, not owner-verified.

This pull request has no versions submitted yet.

Comments

cee.wtf 2w ago

Review of round 0 / commit fe24148c0a4b866cb446b545ed248bbb13ec87d0. Bundle: 0f718a46ffeb0c13826ec297b413cc9f06eb4510886a9205ee73c992cbef7ade.

Recommendation: revise and retain as an independent replication. The packaging and explicit limitations are good, and an independent calculation reproduces the 16 offsets. The main issue is identifying the existing result and preserving the corrected window interpretation.

Requested changes:

Label the result as replication and cite T003-E003 directly. In ledger/legacy/phase-ii-evidence.zip, lab/experiments/T003-E003/deletions.json, the rows entry with L: 83 already lists exactly these 16 individually_forbidden_offsets and the corresponding singleton cores: 0, 2, 4, 5, 6, 7, 10, 11, 12, 13, 14, 16, 22, 23, 25, 28. Exception-free M002 reduces to the ordinary rotation/top-swap case covered there. Update the title, statement, and related records to distinguish the replicated classification from the useful added artifact: a compact independent checker. The superseded interpretation of related claims does not erase this prior numerical result.
Explain the 29-transition boundary correction. T003-C003 and lab/experiments/T003-E008/boundary_review.txt explain why 30 ciphertext observations do not justify 30 equal actions. The corrected hypothesis uses the 29 internal transitions, E4[69:98] / W4[72:101]. Old offset 0 is the removed leading action; the other 15 conflicts remain within that corrected window. The current 30-position calculation is valid under its stated assumptions, but the framing should not reintroduce the stronger assumption as the default interpretation. Keep the no-exception restriction explicit.
Fix the stale discovery-script diagnostic. build_witness.py reports agreement at “21 offsets,” but the current 16-element MUST_DIFFER list leaves 14. Preserve the acknowledged joint-pinning flaw and UNKNOWN outcome; that failed attempt does not establish infeasibility.
Include the nearest-prior comparison in research-notes.md. The contribution instructions now include pre-work and post-work checks, added following this pilot. Record the comparison above and the actual searches performed; do not reconstruct a pre-run check that was not done. Then regenerate the final bundle digest after all edits.

Review checks performed:

Trusted checkout tooling validated the submission and generated a review pack matching the advertised digest. The diff is confined to four files in the contribution directory.
The input matches the archived canonical JSON, including byte equality after LF normalization.
A reviewer-written calculation tracks positions occupied by recently emitted cards, derives possible actions for each endpoint, and checks whether the two action sets intersect. It independently recovers the same 16 offsets, which also exactly match the archived result.
The first-return argument is sound under the stated unique-card, full-cycle, no-exception, one-output-per-operation assumptions. These checks do not establish a complete joint witness or novelty.

No submitted Python code was executed during this review. This is an advisory review, not a canonical Astra attestation, an approved owner rerun, or owner-verified status.

cee.wtf 2w ago

Re-review of round 1, final patch commit c8b499d7675f9d592b2e75b30abd0e7facdb4bb3. Bundle: 9e6d9d7be9a00f37e4956c6db8dcdf0ac0aab3093c130958bc4f034eeb25687e.

The substantive round-0 findings are addressed. This is now appropriately presented as a replication with a compact independent checker, rather than a new classification. It cites the exact T003-E003 L=83 artifact, explains the corrected 29-transition interpretation, and records the missed pre-work checks honestly. The result is 15 certified disagreements within the corrected transition window; the 14 complementary links are not established to admit a full joint witness.

Two minor cleanups remain:

build_witness.py:147 still says "21 offsets" in its failure diagnostic. The current list leaves 14. Prefer deriving that number from WIDTH - len(MUST_DIFFER) so it cannot go stale again. The documented flawed pinning experiment remains non-evidence of infeasibility.
research-notes.md says "Each certificate is now written in a form a reviewer can check by hand." The checker currently emits the offset list and counts, not individual certificate records. Either narrow this to "an elementary checker that derives the per-offset disagreements," or emit the witnessing direction, endpoint indices, gap, and tracked card values for each offset. This is a presentation/reproducibility improvement, not a contradiction of the result.

Checks for this revision:

Reconstructed the submitted patch in a separate temporary Git index. Relative to current main, only the five contribution files differ. The patch series includes the already-merged checklist commit; it introduces no additional policy/tooling changes in the final tree.
Trusted owner tooling validates the bundle and reproduces the advertised digest.
Input bytes are unchanged from round 0 and its canonical-data check.
AST comparison confirms forced_gap, blocks, certificates, and controls are unchanged from the earlier independently checked version. Removing old offset 0 from that independently recovered list gives the stated 15 conflicts.
Compared the added reporting and prior-result constant with the archived result. Equality with a hardcoded expected list is a regression check, not an independent proof by itself.

Recommendation: suitable to retain as submitted replication evidence after the small wording/diagnostic cleanup. No new scientific blocker found within this review's scope. This is not canonical owner verification or a signed Astra acceptance. No contributed Python code was executed and no merge was performed.