Research audit: independent check of PR1005/1018 mutable-leaver cells (no new kappa)

[thakrarsagar]

Status: research audit / independent verification. No new κ; the selected result is unchanged. No counterexample found. One documentation discrepancy.

Track 60–74 (adversarial audit), OS-random seed 65. Base 2921874911956718fcaaaf9396fe8267b3cfa9f0. Audited heads: #1005 779e80ae… and #1018 eac7547c…. #1022, #1028, #1031 and #1042 pin the same seven-slot local word c2b29430…, so the local-word results below apply to them as well. Their owner maps and global replays are not covered.

Why

The mutable-leaver all-K cells of #1004/#1005 are the local word for six open κ claims. Each cell is a coded 21-gate circuit that uses face-event operands with no later reads as mutable scratch, and the FULL cleanup restores them. So far these cells have been reviewed only by the authors' own reviewer agents. This PR checks their premise independently.

What was checked

A new standard-library verifier imports nothing from the audited packages and reads their words as pinned data. It checks the k8 cell word 58e14a02…, the final k8 word db2c7cfb… and the seven-slot word c2b29430…. Every input hash equals the hash the PRs publish.

1. Exact local map. All 224 data columns give x→x, y_t→y_t+x_t. For all 480 helper columns, the inverse of the post-precondition program returns the unit vector plus a target-only residue. This is the exact condition for the pre-main compensation to exist, and it is a different formulation from the constructive one in audit_local.py. The residue is zero for every entering helper.
2. Frames. An independent monotone-climb replay reproduces each recorded paid profile. The FULL cleanup touches only FULL-ended registers.
3. Leaver premise. All 8 coded cells are found literally (contiguous, in published order) by unifying the CODED-CELLS-k8.json gate lists against targets from TRI-LEAVER-k8.json. The coded index is the TRI read index + 8, checked for all cells. None of the 24 mutated leavers (12 data registers, 12 helpers) is a centre, entering or low-ended register. None is touched in A, cut, B or the tail after its cell. Each is written in the FULL cleanup, and each holds a non-input value at the end of B.
4. No new mutation class. Each of the 12 data-register leavers has exactly one B-section write in both the no-triangle base word 3e99ba77… and the cell word, and its A-section writes are identical. The cells re-pair or re-scale existing face-event writes (8 of 12) and remove packet-copy helpers. They add no mutation event, so the mutable-source dependence is the one the lineage already inherits.
5. Slots. The seven-slot word minus the final k8 word is exactly 7×{3:+2,4:−4,5:+2}, matching Conditional κ = 0.001662039077758742 — U10 (two new triangle-hybrid slots) on PR942's T37 owners #987's per-slot claim.

Finding: per-cell profile claim not reproduced

The #1004/#1005 descriptions say each cell changes the raw paid profile by "exactly −30H1+6H2+6H3". Recounted raw profiles of the no-triangle, k5 and k8 words are additive per cell, but with u = +6H1 −12H2 +3H3 −3H6 (rank mass −27, three helpers removed per cell): k5−notri = 5u, k8−k5 = 3u. No rebuilt stage (raw, regroup, hpe, cold44) shows the stated vector, and no code asserts it. κ is unaffected, because it is priced from the final word's full profile. The comparator should be stated or the sentence corrected.

Negative controls

Fourteen controls (7 per word) must each be rejected by the check they target. Rejections: dropped leaver write → endpoint, dropped FULL restoration → endpoint, wrong coefficient → endpoint, write moved after its read → endpoint, entering dirt → entering-response, gate below current frame → frames. A semantically neutral post-cell touch h += ½x; h −= ½x at FULL passes both the endpoint and the frame replay and is rejected only by the premise check. The premise is therefore a design property that the exact map and frame checks do not enforce on their own.

Reproduce

python3 -I -B research/leaver-cell-audit-seed65/verify.py

Expected (about 5 s; byte-compared with RESULT.json):

PASS_INDEPENDENT_MUTABLE_LEAVER_AUDIT cells=8 leavers=24 columns=704 controls=14 per_cell_raw={1: 6, 2: -12, 3: 3, 6: -3} claimed_match=False

The inputs were rebuilt with the authors' own drivers on macOS / CPython 3.14.8: all 21 #1005 chain steps exited 0, and #1018's verify_slots.py printed PASS_SLOTS_ON_CHECKED_K8_BASELINE 0.00167166528321877. SOURCES.json gives each file's provenance. The pushed branch was also replayed from a clean clone.

Scope

This covers the local scalar/frame words only. It does not cover owner maps, hosting, invoices, moment/outer pricing or any κ value, and it does not prove the inherited mutable-source/common-label realization, Lemma A/G, all-size, precision/tape or analytic interfaces. Item 4 shows that nothing beyond those inherited interfaces is needed; it does not prove them. Only research/leaver-cell-audit-seed65/ is added.

Credit and disclosure

The construction is Jeremy Kothe's (#1004/#1005/#1018, with the agents named there). It builds on rohanarun's #987, NewBArch's #869, icekylinx's #645 and its audit_local.py, and chafreaky's #685/#741 lineage. Two cell-definition files are copied unchanged from #1005 under Apache-2.0. Prepared by Sagar Thakrar with substantial assistance from Anthropic Claude (Claude Code).