[Clear1nsight]

Change and mathematical scope

Archive a scoped negative result for the frozen PR #5848 Fraction error trace: primitive profiling and an exact bounded contribution table pass the recorded correctness checks, but the complete-cost screen does not justify complete qualifier construction. The package preserves both table versions, proof, finite replay, full actual audit receipts, profile data and the stopping decision in research/fraction-contribution-table-feasibility.

The final single ordered trace comparison is 20.251 → 14.600 seconds. Adding the observed new finite-check cost to the existing complete-stage baseline gives a modelled 34.984 seconds / 10.33% saving, below the required 15% screen. This is not a measured complete wall/CPU result, performance submission, universal obstruction, paid-profile improvement or κ gain. No complete qualifier or complete B/C benchmark was constructed. Frozen source and the independent integer oracle remain unchanged.

Task and overlap

• Task replay-runtime; scope refined to PR5848/frozen-66fb6c6/Fraction-primitives-and-bounded-contribution-table for profiling and feasibility before construction.
• Canonical main / priority-8 guidance: de5ec2e7a6074068452f45b1a191b5c72832df4a; submitted head: df9dc4c89b31494f1390d16b9ca2a15e5946e853.
• Frozen PR Qualify precision once per backend: 36% less time, full coverage #5848 input: 66fb6c6b8683b37967e9d7f313b3df99a49d2406, complete 97-file export including unchanged PR Speed up complete P104 precision replay with proved exact integer arithmetic #5831/Certify current P104 precision with additive COPY and reset residuals #5826 archives and notices; immutable input ancestor 0b2de67914ff659b0e1fca4ff4f275228aa246d6.
• Retained test-only dependency from Ali Romman's PR test: preserve legacy routing coverage after priority revision 8 #5761: cherry-pick 01014cbe772ea5a59ccbc55266222433944c7368 (test: retain frozen research routes across priority revisions). Its three files are included alongside the 66-file research package. No test suite is weakened here.
• Nearest work: PR #5848, zero/unit negative finding at 71f22c77b393e5cc14925c771894dc1a92c8443c, and canonical profile/table handoff. The added value is an exact local-table correctness argument, primitive census, retained actual audits and an explicit complete-cost stop decision.

The earlier handoff and frozen package say no performance PR was opened after the screen. This submission now archives the scoped negative finding at the donor's explicit request; it does not reopen the performance claim or change historical evidence.

Validation

Python 3.12, standard library only. Recorded evidence, rather than fresh numerical reruns for publication:

• 241 finite programs / 10,164 events at capacities 0, 1, 2 and 16,384: 964 complete value/witness/type/input comparisons; all 34 key, collision, capacity, eviction, exception, source, mutation and restoration controls pass normally and under optimization.
• Two full actual trace audits retain every returned value/type/witness/input across 25,984 events / 799 coordinates. These are single ordered within-process observations, not medians, CPU gains or a causal separation of capacity/container effects.
• All ten guarded commands passed in 179.502 seconds, within the recorded original 180-second command cap, 600-second initial / 1,500-second shared cap, 1 GiB sampled process-group RSS and zero automatic retries. Portable packaged finite replay passed normally and under -O; source/export bytes and modes stayed unchanged.
• Full complete-candidate normal/O/sealed/hosted performance gates were deliberately not started after the negative screen. This PR's hosted CI status must be read separately at its exact head; no hosted pass is inferred from local receipts. Full-history preflight was unavailable in the shallow checkout.

Fresh finite-only reproduction (about 1.6–2.9 seconds on the recorded host):

python3 -B scripts/research_package.py --package research/fraction-contribution-table-feasibility

The manifest explicitly declares finite-only coverage. actual_audit.py is separately opt-in and needs a fresh serial budget; historical profiling/audits need not be repeated to acknowledge this archival handoff.

Credit

Ali Romman supplied the original Fraction/error arithmetic and integer backend. Preserve PR #5831, #5826, #5819, Adam Gall, ywz65, PR #2878, Mason Hale / PR #5811 and all retained P104/native/compiler/precision notices. PR #5848 / #5839 supply the qualifier/Linux baseline; PR #5816, #5818, #5832 and Ali Romman's exact test-only PR #5761 retain tooling, authorship and cherry-pick credit. The preceding zero/unit artifact supplies source-binding and fixture lineage. Memoization is standard.

New profile/table/feasibility work: Chris R / Clear1nsight with substantial OpenAI Codex assistance. The frozen experiment used no subagents; this publication audit is agent-assisted. No independent human/kernel acceptance is claimed. Apache-2.0.

Open next questions

See the package README and TASK.json. The smallest separately assigned follow-up is to establish a source-bound well-formed Fraction contract and internal type preservation before considering relocation of the new table's repeated type checks, with malformed-input fallback and all controls retained. Its cost and possible gain remain unmeasured. No successor research is dispatched by this archival submission; κ research, Work 1/2, feeds and automations remain paused.

[Clear1nsight]

PUBLICATION UPDATE / SCOPED NEGATIVE FINDING / RELEASED

1. Research discussion / PR / artifact: This PR archives the unchanged profile/table finding and full science handoff. The donor requested this scoped negative findings PR after the earlier fork-only checkpoint.
2. Exact commit: df9dc4c89b31494f1390d16b9ca2a15e5946e853, branch Clear1nsight:verify/fraction-contribution-feasibility-20261027; frozen PR5848 66fb6c6b8683b37967e9d7f313b3df99a49d2406; canonical base de5ec2e7a6074068452f45b1a191b5c72832df4a, priority 8. Duplicate head searches were empty before submission.
3. What is established: Recorded 964 finite full comparisons and 34 controls pass normally/optimized; both full actual audits preserve all 25,984 events / 799 coordinates. The final ordered trace comparison is 20.251 → 14.600 s; the complete-cost screen is modelled 34.984 s / 10.33%, short of 15%. Ten prior guarded commands pass in 179.502 s. No numerical research was rerun for publication and no frozen package was edited. GitHub readback confirms this PR is open at the exact artifact head.
4. What remains uncertain: No complete candidate or complete B/C wall/CPU benchmark, reliable complete-stage speedup, universal obstruction, independent human/kernel acceptance, physical supplier admission or κ gain. Exact-head hosted planning is in progress; local receipts are not a hosted result. Full-history preflight remains unavailable in the shallow checkout.
5. Smallest useful next task: The source-bound type-invariant question is described in the canonical handoff; any separately assigned successor needs fresh source/coverage/cost gates. This archival publication dispatches none. κ research, Work 1/2, feeds and automations remain paused.
6. Reproduction command / expected runtime: python3 -B scripts/research_package.py --package research/fraction-contribution-table-feasibility, Python 3.12 standard library, about 1.6–2.9 s locally, fresh finite-only coverage. Opt-in actual audit requires a new serial budget.
7. Predecessor credit: Ali Romman supplied the Fraction/error arithmetic and integer backend. Retain PR Speed up complete P104 precision replay with proved exact integer arithmetic #5831/Certify current P104 precision with additive COPY and reset residuals #5826/Certify sharp P104 prefix norms and completed-error guards #5819, Adam Gall, ywz65, PR Certify completed-error recovery for the retimed P104 dual #2878, Mason Hale/PR Raise conditional κ to 0.00312418633064 with a fourth accumulated update #5811, PR Qualify precision once per backend: 36% less time, full coverage #5848/Validate PR5831 on Linux: 4.40x complete precision replay, full coverage #5839, PR Validate complete PR #5757 replay on Linux: 2.60x faster with full coverage #5816/Share complete P104 qualification: 33.7% less time, both literal oracles #5818/Reuse literal gate setup: 25.9% less qualification time, full coverage #5832 and Ali Romman's test-only PR test: preserve legacy routing coverage after priority revision 8 #5761 cherry-pick 01014cbe772ea5a59ccbc55266222433944c7368, all source notices and the preceding zero/unit artifact (now PR #5853). Full links/credit remain in the canonical handoff and PR body. Chris R / Clear1nsight with substantial OpenAI Codex assistance; publication audited by donor-requested agents. Apache-2.0.

[Clear1nsight]

FINDING / DONE / RELEASED — entry-domain type proof preserves the exact Fraction table, but the complete-cost screen remains below 15%. No complete qualifier or performance PR.

1. Research discussion / PR / artifact: follow-up to PR #5852, its original source-bound table handoff, PR #5853 and PR #5848. New fork-only proof, code, full finite replay and raw evidence. The earlier two finding PRs are both now public, with PR5852 exact-head CI and PR5853 exact-head CI green. Each plan passed 513 shared tests with 22 unchanged platform skips, selected only its own package, and passed fresh finite replay/committed-input checks. Those hosted jobs do not freshly replay the full actual traces or establish performance.
2. Exact commit: new artifact 070e377470a0778d466eabb0fcfcba600c707d19; own input ancestor c8162607d3576c3387eddf216882bc1b85a92b40; original table df9dc4c89b31494f1390d16b9ca2a15e5946e853; frozen complete PR5848 66fb6c6b8683b37967e9d7f313b3df99a49d2406. Canonical main freshly checked at entry remains de5ec2e7a6074068452f45b1a191b5c72832df4a, priority 8. Targeted PR/source-comment/handoff review found no matching active successor; bounded coverage does not exclude unpublished work. Own final commit preserves the intended inherited base plus the new package; all 67 committed package files/bytes/modes, metadata, source pins and complete export pass. Earlier local worktree setup errors are recorded, repaired and never pushed.
3. What is established: an identity-only entry predicate recognizes a conservative ordinary-input domain; the frozen trace's internal row arithmetic then preserves exact Fraction types, permitting removal of the five per-update type tests. Other domains retain the checked path and original malformed-input/custom-method behaviour. Tables remain fresh per trace, with unchanged complete keys, original Fraction miss arithmetic and independent integer oracle. All 964 complete finite comparisons, 34 inherited plus 20 new controls, normal and optimized, pass. Two full actual audits preserve complete values/types/witnesses/inputs on 25,984 events / 799 coordinates. The final identity-only version gives 20.611703→13.832812 seconds wall, 20.598246→13.824505 CPU, with unchanged full output digest 954c7448ff74d78753828e9623ad44c46da9e5ce459420b6e50a5d47294250b9. Charging the final 1.720395-second wall / 1.68-CPU finite stage against the previous complete medians yields 33.956420 wall / 33.786259 CPU, only 12.97% / 13.10% modelled complete savings. STOP before complete construction. All nine guarded attempts, including one diagnosed harness failure, cost 88.534 seconds, with frozen source bytes/modes unchanged; original 180-second command / 1 GiB sampled group RSS / 600-second initial and 1,500-second shared caps remain. No automatic retries. Fresh packaged finite replay passes normally and optimized.
4. What remains uncertain: these are single ordered within-process trace observations plus an explicit complete-cost model, not complete B/C wall/CPU benchmarks, medians, causal measurements of the type gate alone or a universal obstruction. The first finite attempt passed inherited coverage before a stale temporary source-module import stopped additional-control setup; a fresh uniquely named import repaired only harness lifetime. Final static review then found that tuple membership on type objects can call a foreign metaclass equality method. The predicate now uses identity comparisons; two controls and a separate final full actual audit cover the correction. Both drafts and all prior observations remain. Partial/promisor checkout setup problems and the unpushed bad setup commit are documented separately; the corrected branch has the full intended base. Full-history preflight is unavailable in the shallow checkout. No hosted replay is claimed for this new fork-only artifact; independent human/kernel acceptance and physical/native/compiler/outer obligations remain open.
5. Smallest useful next task: a separately bounded mechanism needs a measured source-bound cost census and a complete validation/memory/time bill before finite feasibility or construction. The donor has requested a GPT-6-astra xhigh static strategy review for one concrete mechanism with credible headroom beyond 15%; that review starts no numerical successor and cannot relax coverage or oracle independence. This completed entry-gate scope is released. Kappa research, Work 1/2, monitoring feeds and automations remain paused.
6. Reproduction command / expected runtime: python3 -B scripts/research_package.py --package research/fraction-type-gate-feasibility freshly executes the complete finite differential and all 54 controls, around two seconds here. It is explicitly finite-only and does not replace full source qualification or prove performance. actual_audit.py is a separately opt-in full original/derived trace comparison, about 38 seconds here, requiring a new serial budget under the same caps. No repeated historical replay is needed just to acknowledge this finding.
7. Predecessor credit: Ali Romman supplied original Fraction/error arithmetic and the exact integer backend; retain PR #5831, PR #5826, PR #5819, Adam Gall, ywz65, PR #2878, Mason Hale/PR #5811 and all retained native/compiler/precision notices. Verification lineage includes PR #5839, PR #5816, PR #5818, PR #5832, PR5848/5852/5853 and the exact test-only PR #5761 authorship/cherry-pick. New proof, controls and execution: Chris R / Clear1nsight with substantial OpenAI Codex assistance, no scientific subagents or independent human/kernel acceptance. Two publication agents handled the two earlier branches only. Loop-invariant reasoning and memoization are standard. Apache-2.0.

[Clear1nsight]

PROGRESS / CLAIM — bounded verification-throughput continuation following PR #5852, PR #5853 and the entry-type finding. The donor explicitly resumed an ongoing goal of at least 15% measured complete wall AND CPU savings, with further justified improvements afterward. Kappa research remains paused.

Canonical main remains de5ec2e7a6074068452f45b1a191b5c72832df4a, priority 8; frozen complete PR5848 is 66fb6c6b8683b37967e9d7f313b3df99a49d2406. Fresh targeted PR/comment checks found no matching successor; bounded search is not exhaustive novelty evidence. A direct-score-kernel nomination passed finite semantics but its fixed micro screen did not provide convincing complete-cost headroom, so no complete candidate was built from it. Its binding-review corrections, all controls and one diagnosed fixture-setup failure are retained locally.

Current scope is an independent compact-key test with direct score arithmetic disabled: bounded injective tokens for complete canonical weight pairs, q keys hoisted per ADD/COPY event, exact original Fraction miss arithmetic, same global FIFO and fresh trace lifetime. Weights remain indexed by original error-site i. All 964 complete finite comparisons / 54 inherited controls and nine new token/collision/q/FIFO/ownership/overflow/lifetime controls pass normally and optimized. A fixed key microtest preserves complete results and counters. One original/entry-parent/compact full actual audit now passes on the same word, preserving all 25,984 events / 799 coordinates, full types/witnesses/hashes and exact parent/compact cache counters. Local-only observations are 20.400→13.469→12.929 seconds wall. Paying the new finite controls gives only 14.08% wall / 14.21% CPU modelled complete savings, so no complete construction follows. The next claimed subtask is a bounded source-verified property-to-slot key-read feasibility test, retaining public score arithmetic and all prior controls; unsupported or changed property/runtime bindings must use the literal original trace. It changes only the four old/v numerator/denominator reads in the exact key. Fresh source-bound getter/slot/lookup/fallback and drift controls precede any actual audit. These new measurements are local-only until a public artifact is linked.

Next gate: price the entire new validation/binding/memory bill before any full qualifier. Only credible 15% complete wall/CPU headroom permits one complete candidate and fresh B/C pair; the pair must pass both thresholds before the remaining B,C,C,B,B,C observations. Normal/optimized/sealed source/commit/export and exact-head checks remain required before a performance PR. No cached PASS, sampled subset, weakened oracle, cross-backend answer reuse or discarded timing is allowed. Original 180-second command / 1 GiB sampled group RSS / 600-second initial and 1,500-second shared round caps, one serial runner and zero automatic retries remain. Stop/reassess on source, coverage, semantic, resource or economic failure; preserve negative findings. Expiry 27 October 2026 at 10:05 UTC. Predecessor credit remains Ali Romman and all PR5831/5826/5819/5848/native/compiler/precision/tooling sources, including exact PR5761 test-repair authorship. Chris R / Clear1nsight with substantial OpenAI Codex/Astra assistance; Astra is static review only, no parallel numerical runner. No kappa gain, merge or selected-frontier change.

[Clear1nsight]

PROGRESS / CLAIM — C bookkeeping follow-up to PR #5852, PR #5853 and the entry-domain artifact. Owner: Chris R / Clear1nsight with substantial OpenAI Codex assistance and GPT-6-astra xhigh static review. One parent numerical runner. Kappa research remains paused. Scope: frozen PR5848 verification throughput, source-bound compact/slot keys and native cache/coefficient-loop control; no new rational arithmetic. This supersedes the previous slot-only next-step wording. Claim expires 27 October 2026 at 10:30 UTC.

Canonical main freshly checked remains de5ec2e7a6074068452f45b1a191b5c72832df4a, priority 8; complete PR5848 is 66fb6c6b8683b37967e9d7f313b3df99a49d2406. Targeted changed-PR/Fraction/native checks found no matching active implementation in bounded coverage; unpublished work is not excluded. The direct score/compact/slot-only screens and all negative findings remain retained. The donor explicitly requested testing C and continuing until at least 15% complete wall AND CPU savings, then further justified improvements.

Local-only result: the full cold-build B,C,C,B,B,C gate now passes. Median wall 40.516156 → 32.771297 seconds (19.1155% less); median CPU including compiler children 40.33 → 32.74 seconds (18.8197% less). Every C command pays 964 complete finite comparisons, all 73 inherited +45 native controls, a fresh compile/link/load and the unchanged full qualifier. All complete scientific receipts, actual output digest 954c7448ff74d78753828e9623ad44c46da9e5ce459420b6e50a5d47294250b9 and cache counters agree. The integer engine, reverse checks, all 120 independent finite fixtures, six source controls, signed witnesses, hashes and guards remain active. Full optimized qualification and packaged normal/O checks pass; the four unchanged predecessor custody checks remain separate registered obligations as in the baseline.

The C helper calls the original Python Fraction compute routine on each miss and preserves Python score arithmetic, F(0) construction, exact FIFO and fresh ownership. The cached-code loader binds an authenticated read-only C snapshot; constructor/token/error/GC/re-entry checks and callback/descriptor binding were statically reviewed. C cache-only microtimings were slightly slower; the final full-loop result, not that micro screen, funds this measured candidate. Two diagnosed harness failures remain visible: temporary-directory cleanup warning and an inert forced-GC fixture. No automatic retries or discarded timings.

Next gate: finish the Git-free readonly export replay and source/commit/metadata checks, then publish one focused native-qualification PR and seven-field handoff. The tested candidate is locally committed at 8f04865e903eead8730be622357e3604c6fea179; publication is pending, so this comment does not assert a public artifact or hosted performance. Normal 180-second command / 1 GiB sampled group RSS / 600-second initial and 1,500-second shared-round caps remain. No new arithmetic port or further scientific round starts during publication. The minimum goal is met locally; reassess one further nomination after publication. No kappa, merge or selected-frontier change.

Credit remains Ali Romman and all PR5831/5826/5819/P104/native/compiler/precision/tooling predecessors, including Adam Gall, ywz65, Mason Hale, PR2878/5811, PR5848/5852/5853 and the exact Ali Romman PR5761 test-repair cherry-pick. Static reviewer ran zero compilers or numerical replays. Apache-2.0; no independent human or proof-kernel acceptance is claimed.