Independent verification (no κ): #3015's R1 restore edit on #2979's words is exact and correctly priced; R1 unique in its rewrite class; #3018 journal and #3021 composition reproduced; no defect found

[thakrarsagar]

Summary

Independent verification of open PR #3015 (Jeremy Kothe): the declared restore edit R1 on PR #2979's words with PR #2979's literal T74 allocation. Adds only research/t74-3015-restore-edit-audit/. No κ claim; no defect found in the checked scope.

With code that imports or executes nothing of PR #3015, #2979, #3018 or #3021 (their files are read as pinned git data; the checkers are this contributor's own, with byte-pinned copies of the PR2315/PR2387/PR2804/PR2865 checkers in vendor/):

• The edit. Both candidate words equal PR Conditional κ ≈ 0.0017290339826465: joint rank-7 end frames, five paired extractions and #2957's rewrites on #2942's words (+5.00e-6 over #2969) #2979's words outside restore/report/cleanup, and the restore sections differ by exactly the declared hunk restore[178:180) → 3 gates (own diff). The new block equals the old one as a formal operator on registers 22, 61, 63. A fresh frame-chain implementation (second next to the vendored recount) attributes the paid delta −2H1+H2 entirely to register 63: its climbs 7→8 and 8→9 merge into one climb 7→9; registers 22 and 61 keep their climb dimensions.
• The words. Forward and dual exact through the 5,208-item cleanup and the target endpoint over all 632 independent inputs (36 initial corrections emulated); warp algebra; frame recount {1:2445, 2:706, 3:232, 4:108, 5:18, 8:8} → {1:2443, 2:707, 3:232, 4:108, 5:18, 8:8} with the inventory {4:26, 5:2, 6:16, 7:52, 8:99, 9:213} unchanged; own tracker; the literal cleanup equals an own inverse-reverse re-derivation (one item more than PR Conditional κ ≈ 0.0017290339826465: joint rank-7 end frames, five paired extractions and #2957's rewrites on #2942's words (+5.00e-6 over #2969) #2979's; the parent's cleanup would still pass both replays, which is why this check is needed); 91 private buffers.
• The invoice. PR Conditional κ ≈ 0.0017290339826465: joint rank-7 end frames, five paired extractions and #2957's rewrites on #2942's words (+5.00e-6 over #2969) #2979's allocation (byte-identical in PR Conditional κ = 0.001729150973534224: one declared restore edit (R1) on PR2979's words and literal allocation #3015) priced in two implementations under the new local profiles (271,728 roles, rules R-a..R-e): complete profile {1:1644348, 2:470862, 3:159468, 4:71928, 5:13716, 8:5328, 48:28190} = PROFILE.json = CLAIM.json, delta 666 × (−2H1+H2), idle unchanged, physical stock 66,304 + 408 + 8 + 91 + 2 = 66,813.
• The endpoint. C = 0.001729150973534224428695 accepted and C + 1e-24 rejected by an own rational enclosure and by the vendored 640-bit one; κ = C − 1e-46 − 1e-70; the gain over PR Conditional κ ≈ 0.0017290339826465: joint rank-7 end frames, five paired extractions and #2957's rewrites on #2942's words (+5.00e-6 over #2969) #2979, 58495443855455489/(5·10^23) ≈ 1.1699e-7, is exact; PR Conditional κ ≈ 0.0017290339826465: joint rank-7 end frames, five paired extractions and #2957's rewrites on #2942's words (+5.00e-6 over #2969) #2979's profile is rejected at C.
• R1's class. Every adjacent same-target restore pair rewritten in R1's way (both carrier orders, the first gate at the span of the current frames, the third at every legal subspace), every rewritten word recounted: on PR Conditional κ ≈ 0.0017290339826465: joint rank-7 end frames, five paired extractions and #2957's rewrites on #2942's words (+5.00e-6 over #2969) #2979's words 47 pairs, 122 rewrites, only the R1 pair has an improving option (−2H1+H2 either way round), 12 candidates lose, the rest are neutral; on the edited words 44 pairs, 108 rewrites, none improves. One class, not the author's 2–4-gate block search; petalas's second rewrite of 19:40 UTC (a covector shared by two non-adjacent consumers) is outside it.
• PR feat(research): lower the rank-seven packing moment #3018 and PR feat(research): compose restore and packing gains #3021. PR feat(research): lower the rank-seven packing moment #3018's 103-row journal applied to the same fill reproduces its profile and endpoint. Composed with R1 the endpoint is 0.001729151203983018442998 (+2.3045e-10 over PR Conditional κ = 0.001729150973534224: one declared restore edit (R1) on PR2979's words and literal allocation #3015), computed here without having read PR feat(research): compose restore and packing gains #3021 (published 19 minutes earlier, with priority) and equal to its finite records (profile, idle, physical, roles, κ, gain); PR feat(research): compose restore and packing gains #3021's release is author-run and not rerun. PR feat(research): compose restore and packing gains #3021's own checker verifies R1's preimage and scalar identity but reads the profile delta from the word report; this package recounts it.

Controls: 117 negative controls, each rejected by its named rule through the main-check path (85 distinct: edit 44, words 20, fill rules 10, pricing 17, κ 21, journal 5), 8 positive controls, 7 vendor-pin controls; 23/23 checker mutations each make the run fail (--mutations).

Findings for the author (non-blocking): PR #3015's TASK.json is PR #2979's byte for byte (author, task, scope and open questions describe PR #2979, not R1); the README status line is dated 2026-10-17; the saving's mechanism (register 63) is not stated.

Evidence

• python3 -I -B research/t74-3015-restore-edit-audit/verify.py --fetch (standard library; ~3.5 min; compares with the committed RESULT.json; --fetch writes only the fetched objects and refs/audit/pr3015, refs/audit/pr3018, refs/audit/pr3021). Run here on CPython 3.14.4; the registered 3.12 replay (VERIFY.json) was not run locally (no 3.12 interpreter), its command was run directly.
• verify.py --mutations --jobs 3: 23/23 killed, 0 survived (local run, 20:52 UTC).
• scripts/research_package.py --package research/t74-3015-restore-edit-audit --check-only passes.
• Base canonical main d8e0f83f00191df62174b641e9fff055c357f641, priority revision 6; refreshed at submission (2026-10-16 20:53 UTC; main unchanged at d8e0f83; PR Conditional κ = 0.001729150973534224: one declared restore edit (R1) on PR2979's words and literal allocation #3015 still without comments or reviews; the newer Conditional κ = 0.001729736166387916: six restore carrier/pivot skips (R1_E42, R2_E7) on PR2979's words and literal allocation #3042/feat(research): share the rank-seven restore covector #3044/test(research): independently audit the six restore skips #3050/Verify six-skip/T740 release: conditional kappa 0.001729737950064279 #3051 rewrite PR Conditional κ ≈ 0.0017290339826465: joint rank-7 end frames, five paired extractions and #2957's rewrites on #2942's words (+5.00e-6 over #2969) #2979's words differently and do not audit R1). Overlap refreshed with scripts/find_overlap.py (276 open PRs, 1,756 discussion records at 19:36 UTC) and the PR comments of Conditional κ = 0.001729150973534224: one declared restore edit (R1) on PR2979's words and literal allocation #3015/Conditional κ ≈ 0.0017290339826465: joint rank-7 end frames, five paired extractions and #2957's rewrites on #2942's words (+5.00e-6 over #2969) #2979/feat(research): lower the rank-seven packing moment #3018/feat(research): compose restore and packing gains #3021: no other independent check of R1.

Scope limits (README): the author's exhaustive 2–4-gate block search is not reproduced; the three frames are priced as given (no frame re-solve); native chains, bootstrap, admission and outer release of PR #3015/#3021 are not rerun; the fractional cap certificate is not re-derived; window semantics are inherited. Open next questions are in the README and TASK.json.

Canonical thread: Discussion #2612, DONE post, comment 18982877; pointer posted on #3015. Credit: Jeremy Kothe (#3015), Marius-Paul Dumitrean (#2979 words, allocation and cap machinery), Nick Petalas (#3018, #3021), Ross Richardson (#2987/#3010), Mason Hale, Jeremy Diviney, Yiyang Lu, Tekalign, Dugongue, philippeVV, themilkman216, icekylinx and Zhihao Chen (jacklightChen) and the retained T74 lineage; Douglas Colkitt for the framework. Sagar Thakrar (thakrarsagar) with substantial AI assistance (Claude Code, Anthropic); no independent human review is claimed. Apache-2.0.

Merge danger

Door: two-way. One additive package, removable without changing any supplier.

Blast radius: isolated. No frozen source, shared checker, priority, selection or other PR changes.

🤖 Generated with Claude Code