[mihir-s-05]

Change and mathematical scope

Four revision-7 task cards still invoke the frozen complex-local-word-ceiling verifier, which raises KeyError: 'complex_profile' against the selected #4470 schema. This tooling contribution restores their historical positive control without altering the predecessor checker, catalogue, scientific receipts or selected frontier.

The new research/historical-complex-ceiling-replay package authenticates the complete nine-file execution closure, including the exact selected-record blob at the original PR1240 base e3bb23b4c6bae16ba277f9443e11257a0a37780a. It reconciles the original source pins, runs the unchanged checker in a temporary tree and reproduces its exact result. The standing procedure now gives separate current-orientation and historical-replay steps. The retained #1094 checkpoint in today's tree is byte-distinct from the original selected record, so it is not substituted for the frozen pin.

No new construction, scalar saving or κ improvement is claimed. The historical ceiling is not a bound on #4470. The current selected value remains 60403/20000000.

Task and overlap

• Started from current main 7c472563474e60185e6ca101a17622133d047e56; fetched main again before publication and confirmed that head.
• Assigned card: complex-rank-shape-decoder, priority revision 7, chooser nonce 3d916505ccc35c2632298bca8333380c. Scope changed from decoder synthesis to the bounded first-step reproducibility repair; the original card and originating catalogue remain in TASK.json and SOURCES.json.
• Canonical scoped claim; prior mismatch handoff, credited rather than rediscovered.
• Complements Fix current task startup guidance while preserving frozen inputs #4906, which routes generic baseline commands but deliberately preserves named scientific replay commands, and Add a source-bound baseline for the selected productive-helper schema #4463's current-schema orientation. Fix sparse baseline preflight for selected explicit inventories #4631/Fix Windows command overflow in sparse baseline preflight #4936/Load current selected preflight checker from source #5050 cover sparse loading and workflow execution. This adds the authenticated historical execution route, without changing those shared tools.
• Entry search covered 304 open PR bodies and 4,625 Discussion records. Final refresh at 19:13–19:14 UTC on 2026-10-23 covered 306 open PR bodies and 4,638 Discussion records with complete remote pagination. Local note coverage was sparse; directly relevant packages, DONE: source-bound baseline for the selected productive-helper certificate #4458/Handoffs and requests for independent checks #1166 and targeted all-state PR searches were inspected. Arbitrary diffs, omitted unrelated notes and unpublished work were not covered. These searches are not a global novelty claim.

All eight external executable dependencies are pinned to exact Git blobs/SHA256 values at the base. The retained original selected record is separately bound to its historical Git object and checked as the ninth executable input. The package does not import pending PR code.

Validation

At head bbbcc8e583a2c395a8af91db90ecb29cfaca8ddf:

python3 -B scripts/research_package.py --package research/historical-complex-ceiling-replay --check-only
python3 -B scripts/research_package.py --package research/historical-complex-ceiling-replay
python3 -B scripts/check_research_commit.py --package research/historical-complex-ceiling-replay
python3 -B scripts/export_research_package.py export --package research/historical-complex-ceiling-replay --ref HEAD --output /tmp/historical-complex-ceiling-replay.tar
python3 -B scripts/export_research_package.py check /tmp/historical-complex-ceiling-replay.tar
python3 -B scripts/research_package.py --changed-since 7c472563474e60185e6ca101a17622133d047e56

• Metadata/source validation and both registered normal/-O modes passed on Windows CPython 3.12.0.
• Each mode runs the entire original scientific replay and its four original controls, plus 18 missing/corrupt-input rejection cases and three scientific-data rejection cases. The latter exercise unchanged original guards for profile disagreement, an impossible previous endpoint and an altered result receipt after deliberately bypassing the wrapper's earlier hash gate.
• Committed-byte and export checks passed for 14 payload files. The planner inspected the sparse manifest inventory and selected only this new package.
• The committed export also passed both direct modes in a fresh, read-only Linux directory with Git unavailable (PATH=/nonexistent), CPython 3.12.3: 13.861 s normal and 12.608 s optimized. Output matched and all 15 files, including the export receipt, retained their hashes. Export SHA256: eda19caf2353a1dd07a0ff76c502539871269eb4bc0613e66beb3d53b35d3f3f.
• Exact historical output: PASS_COMPLEX_LOCAL_WORD_CEILING a_ceiling 0.00169005410908900006 kappa<= 0.0016872026453254019. The wrapper explicitly reports context=PR1094 current_supplier_bound=false.
• GitHub's public commit comparison confirms the original base is an ancestor of the pinned current main. The optional full local source-history audit was not run in this shallow checkout.
• git diff --check passed; working tree clean. No current selected cold replay, Lean, Docker, new cost-law proof, independent human review or proof-kernel acceptance is claimed. Hosted Registered research packages CI passed at this exact head: onboarding helper tests, dependency planning, both registered replay modes and committed-input preservation. Both published checks concluded SUCCESS.

Credit

Credit Douglas Colkitt and the original credited Claude assistance for PR1240, PR1012's coverage identity, and the #1094/#1042 construction and audit authors retained in the source packages. Credit the existing mismatch handoff, Mihir Srivastava's #4463/#4906, lee-phys's #4631, the retained #4936/#5050 contributors and Dennis's historical-check replay work. Original bytes and Apache-2.0 licensing are preserved.

Prepared by Mihir Srivastava (@mihir-s-05) with substantial OpenAI Codex assistance in inspection, implementation, documentation and verification; no subagents. This is author-checked tooling evidence.

Open next questions

See the package's bounded follow-ups.

1. Derive and source-bind an applicable Conditional kappa = 0.00302015 (+0.6304% over #4447): joint helper boundaries #4470 helper-inventory cost law and complete consumer invoice before screening current suppliers. This repair supplies neither.
2. Migrate the four emitted first-step commands with an exact-path dependency inventory and explicit treatment of frozen catalogue receipts. Contributors who bypass the standing procedure still encounter the legacy command; no catalogue repin is included here.

[jwr404]

Independent review (James Reed, with Claude Code). Head bbbcc8e583a2c395a8af91db90ecb29cfaca8ddf.

Verdict: correct, and a useful repair. Every factual claim I could check independently holds.

• The failure is real. On canonical main 7c472563, python3.12 -B research/complex-local-word-ceiling/verify.py ends with KeyError: 'complex_profile'.
• Exactly four cards are affected. Walking docs/research/agent-task-catalog.json finds research/complex-local-word-ceiling/verify.py in exactly centre-path-sharing, complex-new-helper-inventory, complex-rank-shape-decoder and coverage-law-change, the four named in the procedure text.
• The historical record is byte-exact. historical-selected.json hashes to Git blob ae6a99f28aa468649eb66a361e764d769363bf11. That is the same blob the GitHub contents API reports for certificates/selected-result.json at the PR1240 base e3bb23b4c6bae16ba277f9443e11257a0a37780a. So the vendored record is the original, not today's retained Conditional kappa 0.001672660197076918: rare-pattern T16 exchanges #1094 checkpoint.
• Replay. It passes locally: PASS_HISTORICAL_COMPLEX_CEILING_REPLAY context=PR1094 current_supplier_bound=false inputs=9 input_controls=18 scientific_controls=3 unchanged=true, then PASS_REGISTERED_RESEARCH_PACKAGE. The frozen research/complex-local-word-ceiling/ tree is untouched; the PR adds only the new package and the procedure paragraph.
• The procedure text is appropriately cautious. It says outright that the historical ceiling is not a cap on Conditional kappa = 0.00302015 (+0.6304% over #4447): joint helper boundaries #4470, and it routes the current orientation separately.