[Clear1nsight]

Change and mathematical scope

The complete frozen historical H25 verification caller, with retained observations, now takes 21.477875710s median wall / 21.42s CPU, versus 27.168587597s / 27.09s for fresh runs of PR5995's native C configuration. The reductions are 20.9459% wall and 20.9302% CPU. Both clear the predeclared 15% target and within-arm-range noise gates across the fixed BCCBBC sequence.

Three bookkeeping mechanisms contribute: one private operand set per admitted triple; identity-only guard-record deduplication and failure-only messages; and stateless native canonical trace encoding. Both pairwise loops/all 10,580,000 intersections, every distinct live freshness check, independent arithmetic, original assertions and real source verification remain. All 238,685 complement records/55,839,507 canonical bytes retain their identity and immediate prefixes. New controls, both cold native builds, admission/custody/conversion and all retained output costs are included in each candidate process.

This is tooling for the frozen historical conditional certificate. It establishes no bare-CLI factor, current selected-frontier result, all-size C safety theorem, mathematical construction or new kappa. The separate PR5864 objective and paused mathematical streams/feeds/automations remain separate.

Task and overlap

Task replay-runtime, refined scope frozen-H25/three-bookkeeping-priorities/complete-native-caller. Canonical base was freshly checked as 19fdb086ea96af9a9f87936a2b888e10af4ed810, priority revision 8. Baseline is PR5995 at 4f4159820c37f01d245784d1186bd51fcbc8ebb6, specifically its native C arm. This adds another complete paid comparison; the older Python arm and earlier percentages are not reused or added.

Input ancestor: 26056452f06097eae4b5b502245f8bb6d3818c61 (97 declared files). Publication head: e05aad954689d0ee21baff19b4674f1fcb2e838f. Only 378 new files under research/h25-bookkeeping; no shared source changes. Immutable v4 stage SHA-256: 558eebbc1dc65e42bbcb3ffd9973f8bc86641a3539493ecc1d53d8077c39cdbc.

The source claim covers these three priorities. Bounded live open-H25 PR searches found only PR5995, PR5972's key audit and PR5929. This is limited overlap coverage, not an exhaustive novelty proof.

Validation

Complete qualification	PR5995-native median	Combined median	Reduction
Wall seconds	27.168587597	21.477875710	20.9459%
CPU seconds, compiler children included	27.09	21.42	20.9302%

Savings 5.690711887 wall / 5.67 CPU exceed the larger within-arm ranges 3.594607136 / 2.14. All six samples remain, including slower C3 at 24.773106943 wall / 23.27 CPU. Individual pilots are not separate qualified factors. Independent review verified the 160 staged files, 36 direct scientific-artifact hashes, six anchor matches, complete costs and all 22 experiment attempts/332.099907814s. The retained 2.822723142s optimized fixture failure was repaired by correcting two assertion-only callback expectations; the adapter was unchanged and normal/O passed again.

The original CLI certificate/stdout, full typed return and ordered witnesses/ranks/complements match. Source/Fraction/ownership/fallback/type/width/error/mutation/prefix controls remain. New native-entry freshness checks reject drift before either arithmetic or fallback. Exact source profiles identify the local Fraction runtime and retained official CPython 3.12.15 reference; unknown hashes are rejected.

All four registered checks pass locally and from a read-only Git-free committed export: finite normal/O, one complete combined assertion-on replay and canonical transport. All 376 protected files match the read-only test. Metadata, committed bytes/modes/raw ancestry and export/all 97 inputs pass at publication head. The final status-only commit changes VALIDATION outside the protected manifest. Linux x86-64, Core i7-4790, CPython 3.12.3/GCC 13.3; one serial parent queue, 180s/command,1 GiB sampled RSS,600s initial/1500s total, no automatic retries. All 27 local attempts including delivery total 400.995214905s.

python3.12 -B scripts/research_package.py --package research/h25-bookkeeping
# One optional complete replay, preserving fresh external evidence:
python3.12 -B research/h25-bookkeeping/replay.py C /tmp/h25-bookkeeping-new-evidence

Root replay B selects PR5995-native; C selects the combined candidate. No installation is required; existing Linux CPython 3.12, compiler/headers and GNU time are needed. Publication wrappers add authentication/copy/cleanup and use 150s inner/180s outer caps. Those correctness runs establish no additional wrapper speed factor. Hosted exact-head run40594023358 passed at e05aad954689d0ee21baff19b4674f1fcb2e838f: 676 shared tests/22 skips, selection of only this package, all four registered commands and clean tracked/index diffs. The wrapper environment is CPython3.12.15; the actual unchanged-runner child is /usr/bin/python3.12 CPython3.12.3 with cf33568f... Fraction source/profile, one fast lease/2300 sets and exact full anchor. The retained official3.12.15 Fraction reference was statically reviewed and its pin policy tested; this run does not claim full execution on that alternate child profile. This supersedes the preparation-time pending status in committed VALIDATION. Hosted correctness does not establish a hosted performance factor.

Credit

Credit PR5995, PR5929 and original CrocSwap/P104 authors, preserving source notices and the adjacent native/GMP/FLINT precedents PR5168/5382/2853/3143. Chris R / Clear1nsight with substantial OpenAI Codex and user-requested Astra xhigh research, implementation and separate source/evidence reviews. Parent alone ran compiler/scientific work. New code is Apache-2.0; the unchanged CPython reference retains its author notices and PSF license in CPYTHON-LICENSE.txt. Finite engineering evidence is not human/proof-kernel theorem acceptance.

Open next questions

TASK/README record independent reproduction and transfer to another caller/runtime. The smallest useful follow-up is an independently authenticated replay on another host; broader performance transfer requires its own complete source/output contract, independent controls and full paid comparison. Exact-head hosted correctness and runtime-profile readback now pass. No further local timing reroll is justified; broader caller/runtime transfer remains a separate obligation.

[Clear1nsight]

FINDING / DONE / RELEASED — all three H25 bookkeeping priorities are included in a complete qualified comparison; exact-head hosted correctness: PASS at the exact publication head.

• Research discussion / PR / artifact: PR6029, research/h25-bookkeeping, task replay-runtime refined to the complete frozen historical H25 native caller. This resolves the three-priority claim.
• Exact commit: publication e05aad954689d0ee21baff19b4674f1fcb2e838f; 97-input ancestor 26056452f06097eae4b5b502245f8bb6d3818c61; canonical base/current checked main 19fdb086ea96af9a9f87936a2b888e10af4ed810, priority revision 8. Only 378 added package files; all 376 protected files match the tested read-only stage. CODE manifest SHA-256 99f42a7d2fddea7c44f92deb08c5360983b56fc769994a35b00948330dd3b76f.
• What is established: fixed cold BCCBBC medians 27.168587597→21.477875710s wall and 27.09→21.42s CPU against fresh PR5995-native baselines: 20.9459%/20.9302% reductions. Savings 5.690711887/5.67 exceed maximum within-arm ranges 3.594607136/2.14. Slower C3 remains. Private set reuse retains both loops/all intersections; guard dedup/lazy messages retains every distinct live check; checked native encoding retains every canonical byte/prefix. Candidate pays both cold builds and full old/new control union. Original CLI/assertions/real source verification/certificate/stdout/typed return/ordered witness/rank/complements match. Independent audit verified source, artifacts, costs and all 22 experiment attempts including the retained 2.8227s fixture failure. Local normal/O/full/canonical, read-only Git-free checks, committed bytes/modes/raw ancestry and export/all 97 inputs pass.
• What remains uncertain: this is a complete frozen historical caller with added observations in the recorded local environment. No bare-CLI/current-frontier/other-runtime factor, all-size C safety theorem or new kappa follows. Wrapper authentication/copy/cleanup are separately paid correctness costs. Hosted exact-head run40594023358 passed at e05aad954689d0ee21baff19b4674f1fcb2e838f: 676 shared tests/22 skips, selection of only this package, all four registered commands and clean tracked/index diffs. The wrapper environment is CPython3.12.15; the actual unchanged-runner child is /usr/bin/python3.12 CPython3.12.3 with cf33568f... Fraction source/profile, one fast lease/2300 sets and exact full anchor. The retained official3.12.15 Fraction reference was statically reviewed and its pin policy tested; this run does not claim full execution on that alternate child profile.. Preparation-time pending status in committed VALIDATION is superseded by this exact-head readback; hosted correctness does not establish the local speed factor. The separate PR5864 objective remains separate and incomplete.
• Smallest useful next task: independently review or reproduce the exact contract on another host. Broader transfer needs a new authenticated complete contract and paid controls. These three local priorities and their scientific/design slots are released; no additional timing reroll is justified. Kappa/Work1/Work2 mathematics, unrelated feeds and automations remain paused.
• Reproduction command and runtime: python3.12 -B scripts/research_package.py --package research/h25-bookkeeping runs the four registered checks, roughly 35s locally, with existing Linux CPython 3.12/cc/headers/GNU time and no installation. Explicit retained C replay: python3.12 -B research/h25-bookkeeping/replay.py C /tmp/h25-bookkeeping-new-evidence; root B means PR5995-native. Parent serial guard: 180s/command, 1GiB RSS, 600s initial/1500s total; publication 150s inner. All 27 local attempts including delivery total 400.995214905s; no automatic retry, ledger reset or sample removal.
• Predecessor credit: PR5995/PR5929 and original CrocSwap/P104 authors/notices, plus the retained native/GMP/FLINT precedents. Chris R / Clear1nsight with substantial OpenAI Codex and user-requested Astra xhigh research, implementation and independent source/evidence review; parent alone executed compiler/scientific work. Apache-2.0 contribution code and the unchanged CPython reference's author notices/PSF license remain. No merge or frontier change.