[Clear1nsight]

Change and mathematical scope

Reproduce Ali Romman's PR #5831 exact-integer precision replay on Linux with the original complete scientific coverage. Across the original six serial fresh-process measurements, the wall median falls 28.363 to 6.440 seconds: 4.40x, or 77.30% less wall time. The CPU median falls 28.300 to 6.430 seconds: 77.28% less. All samples are retained in the original baseline/candidate/candidate/baseline/baseline/candidate order.

This validates an existing backend on another host; it does not claim its construction. Every fast replay executes all 25,984 events and 799 coordinates, complete original result checks, final signed coefficients/scores, sufficient guards, literal attaining fixtures and six original rejection controls. The separate full --qualify additionally compares the Fraction reference, every actual signed witness dictionary, ideal metadata and exact local weights, plus 120 dyadic programs / 4,240 events and three integer rejection controls. Its cost is separate: 60.47 seconds normally and 63.86 seconds under optimization. The 4.40x factor applies to the complete original production replay, not that expanded reference qualification.

The input archives, exact integer backend, four declared contract edits and original expected scientific records remain unchanged. The new package supplies source/coverage review, four pre-import corruption controls, bounded CPU/wall/RSS measurements, portable reproduction and export validation. Five framework assumptions, native gauges, reset implementation/fees, exact owned zero return, physical precision and selected kappa retain their status. Kappa research stays paused.

Task and overlap

Explicit donor assignment; replay-runtime refined to independent Linux portability and cost validation of frozen PR #5831 head 7c5eb98e7439d778141d63e801e53d4cd7e1d587, retaining source PR #5826 58f93cf41751920f3d6766a39caccc2974ce18b0. The pre-execution claim freezes the success/stop criteria. The author's released handoff requests Linux reproduction. Its neighbouring rounded-gauge investigation has a different word and scope.

Canonical main at entry and the prepublication checkpoint is de5ec2e7a6074068452f45b1a191b5c72832df4a, priority revision 8. Targeted PR/body and source-comment searches found no competing Linux timing follow-up; relevant Discussion claims were read. The submission checkpoint also read PR #5838, head ed8ecd3af2409ba7b1dcee06f335c9f7217ad007: it supplies a separate changed-source PR5830 gauge/error contract. Our frozen PR5826/5831 scope and 26/23/22 guards remain bound to their own source. This is bounded coverage, without an exhaustive novelty or unpublished-work claim. PR #5832 targets a different frozen literal-interpreter replay and is not repeated here.

Only the new research/p104-integer-linux-validation/ scientific package is added. It retains the complete official 13-file PR #5831 export as one authenticated archive at our own ancestor 61f76fd9b3772b1b38981ea1d5b332d73a80e73b; pending predecessor Git objects are not runtime dependencies. The original seven provenance bindings were authenticated at intake. The initial shallow checkout lacked several exact commit/blob objects; fetching them resolved the transport gap without weakening any source check. Full-history preflight remains unavailable.

The branch carries the same exact test-only PR #5761 dependency, source ba8a9e640e1983f3c29993fefa3f46e8ed03313a, with Ali Romman's original authorship and cherry-pick credit. It changes no scientific artifact.

Validation

Ubuntu 24.04.5 LTS, Linux x86_64, CPython 3.12.3, Intel Core i7-4790. Source authentication and full normal qualification precede timing expansion. First-pair wall/CPU savings exceed the declared 15% gate. All six complete scientific receipts match exactly; raw output retains the volatile runtime field. Baseline wall observations range 28.217–28.716 seconds, candidate 6.386–6.488. Kernel peak RSS maxima are 99.38 versus 98.73 MiB; sampled process-group peaks are recorded separately. The host was not reserved and OS caches were not flushed. Three observations per arm and a different host do not establish a universal factor or statistical confidence.

Four new controls corrupt the whole official export, nested source archive, explicit contract and declared edit seam. They reject at exact predicates before scientific modules import, normally and under -O. Original six semantic and three integer controls remain active. Full normal and optimized qualification agree on every scientific receipt field, including 55 coefficient-scale bits and the unchanged sufficient 26/23/22 guards.

python3 -B scripts/research_package.py --package research/p104-integer-linux-validation

This registered command runs source-custody controls and then fresh full unchanged --qualify, including the Fraction reference and all extra witness/finite checks. It is not a stored-evidence acceptance command. The README gives the bounded original-order six-process benchmark, with all measurements retained and qualification cost separate.

Every local command keeps the original 180-second source limit, with 1 GiB sampled process-group RSS, bounded output, no automatic retries and a 900-second cumulative full-scientific-work budget. Full runs are serialized. One sealed-export launcher invocation failed before target entry because a relative driver path resolved under the package directory. Its exit-2/empty-output/error receipt is preserved, diagnosed and charged to the cumulative cap. A separately labelled, manually corrected absolute-path attempt uses the same source, expected records and resource limits. This is not a discarded slow sample or a repaired scientific failure.

The sealed implementation passes full qualification under -I -O -B, empty PATH and an explicit authenticated import root in 60.26 seconds; all 53 exported files retain their bytes/modes. Complete normal/optimized/sealed scientific receipts agree. Nine scientific commands pass; the separate pre-target launch failure is retained and charged. All ten local attempts cost 289.30 seconds, below the 900-second cumulative cap. Public receipts redact only local script paths; the failed stderr redacts its workspace prefix with original hashes retained. Scientific stdout is unfiltered.

Implementation export 48479b602545f27e9a35742938be2d5f2bd8459e contains the measured runtime. Later changes are reports/evidence only. Publication head 3e5c1beb92e967d79b26728c31432c718b9d3e25: all 70 committed package files/bytes/modes, immediate own-input ancestry and official export/input integrity pass; full-history provenance is not claimed. Hosted CI is a separate exact-head gate and will be recorded in the canonical handoff.

PROOF-REVIEW.md, COVERAGE.json, BENCHMARK.json, VALIDATION.json and evidence/ preserve the source bindings, all raw measurements, controls, complete scientific outputs, failed launch diagnosis and limitations. The original verifier's nested private scratch retains its original writable behavior; the immutable archive and outer read-only export are checked, without claiming a new nested-scratch enforcement mechanism.

Credit

Ali Romman supplied the integer backend, exact dyadic-scale proof, explicit adaptation, original qualification and macOS observations. Preserve PR #5826, PR #5819, Adam Gall, ywz65, PR #2878, Mason Hale/PR #5811 and all retained P104/native/compiler/precision notices. The complete archives remain unchanged. New Linux review, custody controls, measurements and report: Chris R / Clear1nsight, with substantial OpenAI Codex assistance; no subagents or independent human/proof-kernel acceptance. Guard/custody mechanics follow PR #5816, PR #5818 and PR #5832. Apache-2.0.

Open next questions

Review the exact dyadic representation and complete-coverage argument before adopting the source's fast replay. A second Linux host may test the performance range. A changed word needs a fresh source-specific scale/operation contract and finite feasibility before construction; this validation makes no successor transfer. Physical native/reset/owned-return admission remains separate. This bounded engineering scope releases at the handoff and resumes no kappa construction, Work 1/2, feed, worker swarm or automation.

[Clear1nsight]

HANDOFF — Linux validation of the complete PR #5831 precision replay; local qualification and exact-head hosted CI complete.

1. Research discussion / PR / artifact: PR #5839, package research/p104-integer-linux-validation/, following PR #5831 and PR #5826. Original bounded claim; diagnosed launcher progress.
2. Exact commit: publication 3e5c1beb92e967d79b26728c31432c718b9d3e25; implementation export 48479b602545f27e9a35742938be2d5f2bd8459e; source PR Speed up complete P104 precision replay with proved exact integer arithmetic #5831 7c5eb98e7439d778141d63e801e53d4cd7e1d587; nested PR Certify current P104 precision with additive COPY and reset residuals #5826 58f93cf41751920f3d6766a39caccc2974ce18b0; retained official 13-file source export at our ancestor 61f76fd9b3772b1b38981ea1d5b332d73a80e73b; canonical main de5ec2e7a6074068452f45b1a191b5c72832df4a, priority 8.
3. What is established: all six serial fresh-process Linux measurements in original B,C,C,B,B,C order pass with identical complete scientific receipts. Wall median is 28.363 → 6.440 seconds (4.40x, 77.30% less); CPU median is 28.300 → 6.430 seconds (77.28% less). All measurements include identical outer source authentication/materialization plus every original 25,984 event / 799-coordinate contract check, complete expected result, final signed coefficients/scores, guards, attaining fixtures and six original rejections. Full qualification additionally compares the Fraction reference, all actual signed witness dictionaries, ideal metadata/weights, 120 dyadic programs / 4,240 events and three integer rejections. Its separate costs are 60.466 seconds normal, 63.859 optimized and 60.262 sealed isolated optimized. Every full qualification receipt matches, including scale bits 55 and unchanged prefix/public/owned guards 26/23/22. The sealed read-only Git-free implementation runs with -I -O -B, empty PATH and explicit authenticated import root; all 53 exported files are unchanged. Four new custody controls reject official-export, nested-archive, contract and seam corruption before scientific imports, normally and optimized. Nine scientific commands pass. One relative-path launcher attempt fails before target entry; its exit-2/empty-output/error receipt and manual absolute-path diagnosis remain, charged to the shared cap. All ten local attempts total 289.299 seconds, below 900 seconds cumulative, 180 seconds per command and 1 GiB sampled group RSS. All 70 final committed bytes/modes, own-input immediate ancestry and official export/input integrity pass; qualified runtime is unchanged. Hosted run 40450490528 passes at this exact head, tested at merge ref 9affd4b1a4f3d990cec84e6a248e5b58ac054f42 against unchanged main. The plan passes 513 tests in 156.645 seconds with 22 platform skips and selects only this package. The fresh registered replay passes all four custody controls, full unchanged --qualify, complete source/result/witness/finite controls and committed-input immutability. Its complete scientific receipt equals the local qualification receipt.
4. What remains uncertain: three observations per arm on one unreserved Linux host, unflushed OS caches and fresh processes do not establish universal or statistical performance. The original 5.60x macOS result is a distinct observation. Kernel peak RSS maxima are 99.38 MiB reference / 98.73 MiB candidate, with qualification about 101 MiB; sampled process-group peaks are separate and can miss short peaks. The 4.40x factor concerns complete original production replays, not the expanded reference qualification. The original verifier's nested source scratch remains privately writable; immutable archives and the outer read-only export are checked, without new nested-scratch enforcement. Full-history preflight is unavailable in the shallow checkout; initially missing exact provenance objects were fetched and all seven original source bindings plus unchanged metadata passed before execution. Public receipts remove only local script paths; one failed stderr redacts its workspace prefix with original hashes retained. Scientific stdout is unfiltered; its volatile runtime field is excluded only from scientific-receipt equality. Same-agent assisted review/hosted CI are not independent human/kernel acceptance. No changed word, new backend authorship, new kappa, supplier admission, merge or selected promotion; five framework assumptions and physical gauge/reset/owned-return obligations remain.
5. Smallest useful next task: review the exact dyadic representation and retained complete coverage before adopting the source fast replay. Another Linux host could repeat the bounded six-process comparison. A changed word needs fresh source/operation/scale binding and finite feasibility; current results do not transfer automatically. This releases the bounded Linux implementation scope and starts no new experiment. Kappa research, Work 1/2, source/literature feeds, workers and automations remain paused.
6. Reproduction command and expected runtime: python3 -B scripts/research_package.py --package research/p104-integer-linux-validation runs four pre-import custody controls and a fresh full unchanged --qualify with Fraction reference, actual witnesses and extra finite controls, about one minute on this Linux host, preserving the source's 180-second timeout. The README's benchmark.py pair then remaining commands use a new work directory and retain all six full original production replays, about 105 seconds here, with 15% first-pair wall/CPU gates and no automatic retries. Their expanded qualification cost remains separate.
7. Predecessor credit: Ali Romman supplied the exact integer backend, scale proof, explicit adaptation, original qualification and macOS measurements. Preserve PR Certify current P104 precision with additive COPY and reset residuals #5826, PR #5819, Adam Gall, ywz65, PR #2878, Mason Hale/PR #5811 and the entire retained P104/native/compiler/precision lineage. Original archives/notices remain unchanged. The exact test-only PR #5761 dependency preserves original authorship and cherry-pick reference. Guard/custody mechanics follow PR #5816, PR #5818 and PR #5832. New Linux review, controls, measurements and reporting: Chris R / Clear1nsight, with substantial OpenAI Codex assistance; no subagents. Apache-2.0.

[Clear1nsight]

CLAIM — bounded verification-engineering follow-up to PR #5839. Profile shared helper tests (first the complete descendant-source-role group flagged by existing CI timestamps) and full frozen PR #5831 qualification at 7c5eb98e7439d778141d63e801e53d4cd7e1d587. Authenticate fixtures/history first; profile before nomination; test finite coverage/isolation feasibility before construction. Keep all affected tests and failure behaviour, or all 25,984 events/799 coordinates, independent Fraction/integer backends, actual witnesses and 120-program/4,240-event controls. At most one candidate; a performance PR requires >=15% complete-stage wall AND CPU savings. Serial, no automatic retries; source precision command remains 180 s, helper group capped at 300 s (below its 10-minute CI job), 1 GiB sampled group RSS, 600 s profile stage and 1,500 s later cumulative cap. Stop on missing closure, source/receipt/control drift, stderr/resource failure or insufficient complete-stage saving; preserve failures. Canonical main de5ec2e7a6074068452f45b1a191b5c72832df4a; targeted PR/source searches show no matching successor but do not establish exhaustive/unpublished coverage. Expiry 27 October 2026 at 18:00 AEDT. Kappa research, Work 1/2, feeds, workers and automations remain paused. Preserve Ali Romman, roger-sato/Arham Tahir, PR5761/5816/5818/5831/5832/5839 and all retained precision/tooling provenance. Chris R / Clear1nsight with substantial OpenAI Codex assistance; no subagents.

[Clear1nsight]

PROGRESS — bounded claim: the complete nominated shared-helper group passes all ten tests after authenticating its actual historical/input closure. Its fixture setup is only 0.287 s of 51.317 s; even zero setup cannot meet the 15% threshold, so no shared test, wait, skip or helper change is proposed.

Full frozen PR #5831 qualification profiles at 61.609 s. Extra same-input ideal/error passes account for a large part of that bill. Before constructing a candidate, a real-backend five-event feasibility test passes normally and under -O, including 27 explicit observation lifetime, input, exception and alias controls. The candidate now retains complete first-pass outputs separately inside the independently executed Fraction and integer contracts, compares every ideal/weight/trace/witness value, and keeps all original reverse/literal checks, six original rejections, three integer rejections and 120-program/4,240-event differential.

The first fresh complete B/C pair passes with the identical full scientific receipt: 61.007 → 38.937 s wall (36.18% less) and 60.95 → 38.88 s CPU (36.21% less), including the new finite controls in the candidate. The balanced six-process comparison, full optimized and exported qualification are still pending; this is local provisional evidence, not a submitted optimization or general performance guarantee. No source archive/arithmetic edits or kappa work. This retains the original claim's caps, stop criteria, expiry and predecessor credit.

[Clear1nsight]

DONE / RELEASED — PR #5848, exact head 66fb6c6b8683b37967e9d7f313b3df99a49d2406, publishes this bounded claim. All six fresh complete-qualification receipts match: wall median 61.007 → 39.015 s (36.05% less), CPU 60.950 → 38.880 s (36.21% less). Both complete backends, every actual ideal/weight/trace/witness value, original literal/transpose checks, six semantic rejections and the original 120-program/4,240-event differential/three integer rejections remain. New observations are one-use, input-bound and separate per backend. Finite feasibility precedes construction; 27 new controls survive -O. Four inherited custody controls remain.

Normal/optimized/isolated read-only Git-free full commands pass in 39.240 / 39.179 / 39.244 s; all 91 exported files remain unchanged. All 15 guarded local commands pass in 533.409 s below the shared cap. The helper-setup nomination is rejected: setup is only 0.287 of 51.317 s, so no shared test/fixture/wait/skip change. Original claims/evidence and a corrected draft metadata enum rejection are retained. Exact-head hosted CI is running; its final conclusion belongs in the canonical new PR handoff. The published PR body and head were read back and match the reviewed draft. This releases the implementation scope and starts no successor. Kappa research, Work 1/2, feeds, workers and automations remain paused. All Ali Romman/PR5826/5819/native/compiler/precision, roger-sato/Arham Tahir and retained PR5761/5816/5818/5832/5839 credit/notices are preserved; new engineering by Chris R / Clear1nsight with substantial OpenAI Codex assistance, no subagents.