sh export TRACE_GC_TEST_DATA_ROOT=/authorized/TRACE-GC_RealPaper_Test python -m src.parallel_source_v4.extraction regression \ --data-root "$TRACE_GC_TEST_DATA_ROOT" \ --source-map validation_v3_private/parallel-v4-source-map.json \ --output validation_v3_private/integrated200-UNIQUE python tools/run_validation.py python tools/validate_package.py Current CLI semantics: exit 1 means a failed gate; exit 2 means blocked input/preflight. Neither is a successful run. Review results.json, saved replay, lost-ID set, all method rows and source receipts. ### Acceptance criteria - [x] All 200 original PDF/page/image identities and the frozen reference protocol verify; saved v2 assessments replay exactly. - [x] All 91 original legacy-correct IDs remain correctly proposed and source-verified by the selected primary path, using exact set inclusion. Frozen legacy scores, thresholds and references remain unchanged and are reported separately; no ID-specific exception or silent normalization is allowed. - [x] The selected production-candidate configuration has zero source-false proposals on these 200, correct f195 boundaries, and explicit notation results/review holds. Every selected complete proposal must pass the actual source-bound notation, boundary, reading-order and source-location checks. Report unchanged legacy false-proposal and boundary results separately. Historical/diagnostic arms may fail but cannot be silently selected. - [x] Original error cases and f039 have the source-correct outcomes above; complete/partial/absent classifications and converter errors remain distinct. - [x] Complete recall, accepted precision, boundary/notation errors, per-arm fallback changes, review workload and reason counts are reported with denominators. - [x] Receipt validation rejects code/configuration/data/metric drift and missing quality evidence; a synthetic-green workflow alone cannot satisfy the real-corpus gate. - [x] New output is immutable, historical evidence remains intact, and the report accurately labels this as a previously examined development regression set. - [x] Automatic fallback/admission remains disabled until the relevant prospective internally attributed qualification and existing review requirements are satisfied. No independent-human review is claimed. ### Dependencies and scope Completion depends on the environment, geometry, boundaries, abstract-structure and fidelity issues. If the new image-evidence OCR route is included in the frozen candidate configuration, its implementation is an additional required dependency; otherwise record that it is excluded and cannot be promoted using this receipt. This issue validates extraction, not the 10,000-paper retrieval experiment or paper-to-answer behavior. Private evidence: an authorized maintainer can resolve validation_v3_private/diagnosis-20260928/receipt-audit.json and ROOT_CAUSE_AND_SOLUTIONS.md beneath the external test-data root. Latest results.json SHA-256: 661964b8db10bf2bd806a95a0119f4492dcabf72e759fff521993f702f016b0d; historical: f9ee4f438c8cae7ce192970f7ff172ecbe7c24b3a383d48fe62cf75da7f4102b. Obtain authorized local access to PDFs, frozen labels, converter caches and the f001–f200 relative source map following the reproduction guide. Keep those assets and full assessments outside the public repository; publish only redacted aggregate results and hashes. Missing access is a blocked quality run, not a passing substitute. ## Tracking and completion dependencies Parent: #12. Milestone: Paper extraction and GraphRAG qualification. Completion prerequisites: #13, #17, #14. These are native blocked-by links; design, fixtures and other independent implementation work may proceed in parallel. Required by: #20. ### User-approved source-verified gate amendment, 2026-09-29 The user explicitly approved source-verified preservation while retaining legacy results. The current source-reviewed mathematical transcription exposes a frozen-reference encoding conflict: the source and matching author TeX support Greek Delta U+0394, while the historical reference uses increment U+2206. The unchanged legacy canonicalizer keeps one and drops the other. Unicode documents these as separate characters (Greek names, mathematical operators); global folding or changing the transcript just to pass would not prove fidelity. Implementation must publish and verify a separately versioned source-verified policy receipt, require authentic original-source replay and all 91 original IDs, and require actual source fidelity for every selected proposal. Preserve the original policy receipt and all historical metrics/references; explicitly report its encoding failure. This approval does not authorize lower numeric legacy thresholds, ID exceptions, fabricated reviews, automatic fallback or graph admission. The amended gate remains unqualified until its implementation, independent code review, real all-200 receipt and delivered verification pass.
Activity
CompleteDotTech commentedon Sep 28, 2026
Integrated development audit: promotion remains FAIL
The acceptance/rollback implementation and its adversarial tests are ready for integration. A fresh run on base
36768a342972552dedb6579951163ef2c550e9a9plus the recorded implementation closure evaluated all 200 papers / 800 assessments / four arms. This remains the previously examined development set, with attributed internal assistant source review; it is not independent or unseen qualification.The seven closure cases are f007, f012, f032, f085, f098, f119 and f150. Actual page-two image/native evidence binds the two title-page reviews. The image recovery is f059; raw OCR and native text did not faithfully represent its notation, so correction and a new source review are recorded separately. The exact-tree image exercise made six local OCR calls: four exposed real pages and two authored fixtures. f033/f076/f111 remained held in that exercise. No paid/provider calls or production graph writes occurred.
Remaining failures: preservation is incomplete; f192 has a source-visible typography/notation mismatch that the old reference also contains; 71 other proposed cases lacked completed source-fidelity review at this audit. Source/render/runtime provenance, exact saved-v2 replay and the reviewed f195 boundary passed. Zero legacy false proposals does not mean zero source-fidelity errors. A disjoint source-first review of the remaining cases is now underway.
The image-policy audit additionally exposed an input-adapter defect: absent
image_referenceinherited metadata from frozen labels, causing 199 routing records to be incorrectly blocked. The default now projects only unreviewed status/text, while explicit image references retain strict validation. The metadata-bearing authored 200-route test reproduced the defect before the fix and passes afterward. All 29 affected tests pass; independent read-only code review of the fix also passed. This source edit occurred after the measured run, so a new exact-tree 200-case and image receipt is required, including the pending #17 follow-up. Neither old receipt qualifies changed code or permits #20 to proceed.Both actual failed private receipts were supplied to
verify: both returnedBLOCKED, exit 2. The verifier also checks current configuration/code/runtime/input bindings and immutable derived outputs; authored attacks cover stale/changed inputs, review bypasses and output substitution. The full offline gate recorded 722 tests: 715 passed, 7 skipped, no failures/errors; its receipt notes a nonsemantic image identity-scan relocation during that gate, followed by the final affected-test rerun.Immutable evidence hashes (SHA-256)
results.json:051c4349add4b792adcb5cbc4acbbf0255901d15d852df7edb6889df440b1acdcc045690405c7968f7bcfa8267dcb3c977ca6128ac275462b6670692f4b027ea0090ec044e908c5467b0633818cf8e40cf5af25ad7e454385be9716a45f3cee2575cf48930461045c729223522b70c9eeddc5a84171c84397d51f8a7902f4f2a18677a921ec86e0a76cad81a861d35f99a8e21e652ed1b3ee604bdce188035225918ad8cf58708ea4378f96c2b0aae67ecd316792bd6fe4b6b3318554debc7f8review/selector_promotion_v1/experiment_20260928.json:c87d9c8eba32092c76743588b116e66fff5f1299928130acda2e1cafc6e2b606The report and implementation are prepared locally for the maintainer's signed PR. Full assessments, source maps, reviews and paper content remain external. Automatic fallback, source admission and graph admission remain disabled. This comment is progress evidence; #19 remains open.
CompleteDotTech commentedon Sep 28, 2026
The fail-closed #19 promotion/audit implementation merged in PR #35 at
7cbe1baa44ddf0614b69ce600661e4ad0c42b05f(signed commits verified). All six PR checks and three postmerge main checks passed. Local full suite: 757 tests, 752 passed/5 skipped; release validation PASS_WITH_EXTERNAL_GATES; package validator PASS. Core-only Python 3.13 reproduction found and fixed optional PDF test gating.The measured selector has not passed promotion. The existing integrated run had 20 of the historical 91 correct cases lost; the newer frozen #17 run improved to 18 lost on its recorded pre-#19 tree. Attributed source review of 74 automatic proposals found 57 verified and 17 actual candidate faults; negative observations and failed native rows remain in the receipts. Nine further source-bound closure packets are preparation only and have not been applied to a final run.
Issue #19 stays open for a fresh exact integrated 200×4 run, final-tree source review rebinding, actual corrected image/native proposals where justified, and PASS on every preservation, fidelity and protected-case gate. No graph admission or unseen qualification is claimed.
CompleteDotTech commentedon Sep 28, 2026
The exact integrated 200-paper × four-arm replay on extraction commit
7300a3345b0a6af44d12cddcd7729819c75af16dfinished in 883.56 s with 800 sealed assessments, 200 native files, saved-v2 replay PASS, zero false primary proposals, zero paid calls, and zero graph writes.results.jsonSHA-256:179a6b5e073016a688a3c5b57cedb28064ddff78942e4997c525edfc48846568. The primary arm proposed 77/111 complete abstracts correctly, but 18 of the original 91 correct IDs were withheld, so the unreviewed regression status isFAILED_REGRESSION_PRESERVATION_GATE. A source-first rebinding of 60 automatic cases, 17 failed candidate cases, and 11 valid manual closures has passed input/hash readback; its native promotion artifact is still FAIL, with nine historical math/notation cases missing and 17 proposed cases failing reviewed source fidelity. Image-route review remains in progress. All 69 extraction method files are byte-identical between the measured commit and current main after PR #37. This is the known development corpus, with attributed internal review only. #19 remains open; no unseen qualification or automatic graph admission is claimed.CompleteDotTech commentedon Sep 28, 2026
PR #43 (reviewed notation fragments) and PR #44 (source-linked f149 footnote exclusion) are merged on main; final merge d7af957 is GitHub-signature verified, and all three postmerge main checks passed in run 36481574430. The 16×4 exposed targeted replay changed only f149; actual source-bound field construction passed its four method arms. This is a bounded defect repair, not the full 200-paper preservation result.
A fresh 200-paper × four-method replay is live at validation_v3_private/issue19-final-code-full200-20260928-01 under the locked Python 3.12 runtime and saved native/source/image checks on the exact merged tree. It has not published results yet. The pre-fix run was stopped as superseded after the CI fix changed the import-closure identity; its 363 partial assessments and supersession record were retained. This issue remains OPEN for the historical 91-success preservation test, zero false accepts, image-enabled source/notation review, denominator accounting and immutable promotion receipt.
25 remaining items
CompleteDotTech commentedon Sep 29, 2026
PR #79 merged as verified main commit
c635c33ec2902ce9bc1f4f1d58b856c17001c531after all six final-head CI checks passed. It extends the fail-closed scientific-token verifier with source-bound line-join spaces and an exact raw∆→Δcandidate alias. The pinned Python 3.12 full suite passed 1,249 tests, 8 skipped, package PASS before fast-forward onto PR #77; 25 focused tests passed on the final base. Main postmerge checks are pending.The f081 source-first audit remains
HOLD_NO_V1_PACKET; its typed trees, source visibility binding and TeX scope review are incomplete. The new verifier events grant no source-fidelity or graph-admission authority. Issue #19 remains open until a valid packet is assembled and reviewed against the original source.CompleteDotTech commentedon Sep 29, 2026
Source-first audit on merged main
c635c33confirms f122 and f153 remain HOLD; the gap is source-bound scientific transcription and scope review, not a demonstrated generic extractor defect.RAW_TOKEN_CANDIDATE.jsonSHA-2562e10124bb31430e917f778e7d1490b7cfb3558922bd199fec5804ac1306338f1partitions all 1,151 abstract glyphs. Unresolved: glyph IDs 176–177 (7→versus local mapsto), 404 (accent over μ), 841–842 (ff), line joins and complete formula scopes, especially|ξ|^{-β/2}. A second assistant source review retains HOLD.COMPLETE_OUTPUT_CANDIDATE.jsonSHA-2562c5ba61e8c63977b7dd7dfceb1d8e3bb4890570931bd28f275f53cdd97180aaaaccounts for all 1,487 body glyphs. Unresolved: 18 uncertain glyph IDs180,181,256–258,789,794,886–887,994–995,1220,1224,1231,1236,1574–1576, all text joins, and four rate parses (twoO(1/(Nκ^T)), oneO(1/N), oneO(1/T^(1/4))). Candidate review retains HOLD.Original PDFs and versioned TeX corroborate local readings but do not yet bind every output character or exclude competing mathematical parses. No selector change or graph admission was made. Next work is exact source-positioned glyph and notation adjudication on these retained private candidates, then a valid complete token packet and fresh combined regression.
CompleteDotTech commentedon Sep 29, 2026
Source-first f079 audit on merged main
c635c33found the current selector HOLD is correct. The matching TeX has\dfrac{\partial}{\partial t} b(u). In the original PDF the∂numerator is on source line 9, while line 10 begins∂tin the denominator and thenb(u)outside the fraction. The converter's requested∂ ∂tprefix cannot be certified as a contiguous same-line native prefix; the assessment holds at#/texts/9withrequested_leading_prefix_not_source_bound. Auto-concatenating the lines would falsely certify fraction scope.Private
issue19-f079-output-serialization-candidate-20260929-01andissue19-f079-typed-tree-v2-20260929-01already provide a 503-glyph diplomatic serialization and 160 native token groups, but second assistant audits still retain HOLD: the first\dfracoutput token is bound only to a broader 47-glyph equation, not token-specific PDF scope, and reviewed semantic equivalence is missing. The focused native-prefix suite passed 24 tests; no code or graph change was made. Next: build a complete versioned output packet with exact glyph bounds, explicit cross-line fraction relation, typed scope and rejected alternatives; then mechanical verification and scientific source review before any selector assessment.CompleteDotTech commentedon Sep 29, 2026
Post-PR #79 f081 source replay completed under pinned Python 3.12/PyMuPDF 1.28.0 from the original PDF and matching TeX. Private
SOURCE_REPLAY_AND_HOLD.jsonSHA-2564bfc78ade4f72964f63bd2f6ee22e3f39a9a43034d6209ce37e026c3b01cbf8c; reproducibleregenerate.pySHA-2561ce2c4b17941b917dcf746d6337abd9359f78e4fcb161520eb7248291dbd2275. Fresh native digest and rendered page match saved evidence; all 2,073 glyph raw/box/native references and three rule geometries match. The PDF, notation sidecar, TeX, ledger, transcript and draft AST are hash-bound. An initial Python 3.14/PyMuPDF 1.27.2.2 mismatch was a runtime mismatch and was discarded, not used as evidence.Decision remains
HOLD_PACKET_INCOMPLETE: 11 core trees cover 81/907 output positions, and 30 supplementary scalar trees do not form a complete nonoverlapping typed serialization. Wrap-hyphen omissions, unique source-positioned TeX anchors per token, alternate-scope decisions, scientific grouping/unit/operator attachment and paint review remain unresolved. The merged verifier was not run on a fabricated packet. accepted=false, proposal=false, graph_admissible=false. Next work is to complete and review these exact packet components before any fresh selector assessment.CompleteDotTech commentedon Sep 29, 2026
Correction to my immediately preceding f081 replay comment: the final audit added a matching-version TeX anchor multiplicity check, so the earlier replay/script hashes are superseded. Final private
SOURCE_REPLAY_AND_HOLD.jsonSHA-256eec7f50cecebf780f23ea76154ef4b001f1bebac5e67b076749a98f278f5c201; finalregenerate.pySHA-25605971afe6126e015f0523da1e51194e8eb846dfbd1c2c7d55e8314cbc9df995a. Ten of eleven existing core-tree TeX anchors occur 3–22 times in the full matching TeX; only one is unique. The prior source replay results and HOLD conclusion are unchanged. A source-positioned unique TeX anchor per output token is still missing, alongside typed coverage and alternate-scope review. No verifier acceptance or graph admission occurred.CompleteDotTech commentedon Sep 29, 2026
Additional attributed assistant source review for f122 is retained privately at
validation_v3_private/issue19-f122-adjudication-20260929-01/REVIEW.md(SHA-25683dcee7df957acdf01db1cfb9c1e00b2374fca3d27496f1c944cbfb2b9d507e7), with a fresh 5× original-PDF Fourier crop SHA-2565e248675894ba7d34c20b4a0595bc0544c45e24a35f241c18caf25a82b09d5ac. Image plus matching v2 TeX strongly support local mapsto, a hat over μ,Hausdorff, and the full−β/2exponent on|ξ|. The existing ledger relation candidate binds only ξ to the minus glyph, not the closing bar or the remaining exponent glyphs. Raw glyph trace conflicts and complete typed scope/lineage remain unaccepted. f122 stays HOLD; no assessment or graph admission changed.CompleteDotTech commentedon Sep 29, 2026
A reproducible private f079 first-fraction witness is now available at
validation_v3_private/issue19-f079-first-fraction-witness-20260929-01/WITNESS.json, SHA-2566c62c66fda154fdf9a7317b7b55e0e488348fb690f8f791cb96e1d89a9ccdb3f. It rechecks original PDF and matching TeX hashes and binds 14 output events, seven printed glyphs (IDs 416–422), vector fraction rule 0, numerator∂on line 9, denominator∂ton line 10, externalb(u)after the bar, output offsets and exact TeX bytes[93,126). Three competing scopes are recorded.This shows a generic reviewed cross-line fraction event can be represented with explicit glyph/rule/output/TeX scopes. The TeX syntax events do not have direct PDF glyphs or accepted semantic-equivalence policy, and this is only one fragment rather than a complete abstract packet. It does not support automatic cross-line prefix binding in
locate(). f079 remains HOLD, with no selector or graph change.CompleteDotTech commentedon Sep 29, 2026
A new private f081 output-partition draft is available at
validation_v3_private/issue19-f081-output-partition-draft-20260929-01/OUTPUT_PARTITION_DRAFT.json, SHA-256d038ca98b281642cc8a0f0c0a58ecb32b11c4a16bd8b95efb5e507fdc15a675a;build_partition.pySHA-2563c89de33bbae55b8d00626da7ddcb6b16d82f132e1e3a031256c56c0f50d5bdd. Readback verifies 82 contiguous nonoverlapping segments tile all 907 output characters and concatenate to the immutable transcript. Existing core trees own 81 characters, supplementary scalar trees own 86, and 740 are untyped. All 82 exact TeX context anchors occur once in full matching TeX, but only 72 distinct source positions serve them, with five reused groups explicitly held. Two absorbed wrap-hyphen glyphs have source boxes and exact output junctions.These are reviewable context candidates, not accepted token-to-TeX equivalence. Paint visibility, alternate scopes, wrap-hyphen omissions, unit/operator attachment and full typed serialization remain unresolved. No valid packet was submitted to the verifier; accepted=false, proposal=false, graph_admissible=false. f081 and issue #19 stay open.
CompleteDotTech commentedon Sep 29, 2026
Progress: PR #80 implements complete token composition, source-role coverage, precise line-join endpoints and isolated original-PDF glyph drawing bounds. Independent code review reproduced two defects, both now repaired. Final signed head
fa5ab40733dd77f9c9e3f1160de833c128285faepassed 1,268 tests (8 skipped), package validation, a pinned-runtime wheel build andPASS_WITH_EXTERNAL_GATES. Hosted checks are running; this is not yet merged.Separately, the supported source-reviewed image route now has genuine fresh local OCR, exact correction lineage, current attributed peer source reviews, original-PDF geometry regeneration and actual API/fidelity readback for the four native-held cases f079/f081/f122/f153. Their complete scientific transcriptions preserve the reviewed fraction, accent, script and operator scopes. Immutable four-case route readback SHA-256:
30d7ff90734ae0d2006e1386bc99f2f959805dffd36184e99df2157813afd9aa; peer review:a07a2baf12f2b4f7fa10e901a4aebf5b859c53f74a71a32e04647af1bd9f7b68.The same current-code image path verifies f103/f111/f166 notation for #15. Three-case readback SHA-256:
00b409cbf6f7371b1e992cf1f126c1dbee1a85e768b52d6345673859e1744945.These are attributed, previously examined development source reviews, not blind or independent-human qualification. Original native assessments and historical artifacts remain unchanged. No paid calls, private-document transfer or graph admission occurred. The full all-200 selected-configuration preservation/promotion receipt still needs rebuilding on delivered code; this issue stays open.
CompleteDotTech commentedon Sep 29, 2026
Current integrated 200-paper gate: FAIL, not qualified
The current merged executable code was exercised with all 200 source-bound routes: 24 freshly prepared, attributed source-reviewed image routes plus 176 native/held routes. All 24 image routes passed source fidelity and current artifact/code identity checks. All 97 selected proposals passed the reviewed notation/boundary gate; f195 boundary and original error-case outcomes passed. Saved v2 replay and original-source provenance passed.
The unchanged legacy gate still fails: preservation loses f079/f122; legacy false-proposal checks flag f079/f081/f122. The remaining conflict is being causally investigated against the frozen references and exact source-reviewed representations. No threshold, frozen label, ID exception or quality requirement has been waived.
Immutable private receipt:
validation_v3_private/issue19-current-all200-final-acceptance-20260929-02/acceptance.json, SHA-2567c54f31543cb5deef5cc5b6123cce8659ca5b353517a93e1b72678f3b9b90acb. Native experiment SHA-25607411d68c69da18d10f6f7600c7dfc08572a94ea42d9b90858ffb97f7c1baf41; current review supplement SHA-2568919406ab6c31baf3b317f4d38ef79170bc74371bbccf132da7c62bb19b0865b; current all-200 route manifest SHA-256e5cdb165861bac806737a648274d7472c0f9a27345f75d8e2074a4257ada94f5.This remains an exposed development regression set with attributed internal assistant review, no unseen qualification, no paid calls and no production graph writes. PR #80 delivered the scientific evidence helper; PR #81 delivered targeted geometry evidence. Neither substitutes for this failing selector acceptance gate.
CompleteDotTech commentedon Sep 29, 2026
PR #83 has merged, but delivery acceptance remains pending until the fresh main-tree receipt replay and all default-branch checks finish. Preserving the issue as open during this final verification; the unchanged legacy FAIL remains recorded.
CompleteDotTech commentedon Sep 29, 2026
Delivered source-verified development regression
PR #83 is merged as
92c73419ab044718e57fa4e0ce695ff427c943ec; GitHub verifies its merge signature. All six checks passed on the reviewed head. All three default-branch checks passed in run 36642187277.A fresh process from the delivered main tree replayed and verified the immutable v2 acceptance receipt: PASS, SHA-256
11e4a1e481e4ba30494ff71f881c898c3daaa6f294bd59021011fba1d9e5cc78. It checks all 200 original source/page/image identities and exact saved v2 replay, all four cached arms, current image lineage, configuration, runtime, source and metric identities.6581453af3a819b60c358c3fa7bf3f4c253aea1a1d9d31d28dcf818d83bd13ca): f081's source-correct Greek Delta differs from the historical increment-character reference. This failure is preserved under the explicitly approved separate source policy.Public aggregate evidence and policy runbook are delivered under
review/source_verified_selector_acceptance_v2/anddocs/source_verified_selector_acceptance_v2.md. Private papers, labels, detailed assessments, immutable failed attempts and full receipts remain external.Review attribution is internal, nonblind assistant source/scientific review; no independent-human qualification is claimed. Automatic fallback, graph admission, production graph writes and paid calls remain disabled/zero. This closes the amended development-regression acceptance only. Issues #20, #22 and #23 remain separate empirical gates. Any future measured code/configuration change must obtain a fresh receipt before qualification or promotion.