Repository navigation
Reliable paper extraction, retrieval and real GraphRAG qualification #12
Description
Outcome
Recover trustworthy paper extraction, qualify the retrieval pipeline, and demonstrate a real paper → reviewed evidence → persistent graph → cited answer workflow. The 200-paper first-page regression is one explicit gate within that outcome. A paper without a first-page abstract must not be made to produce one merely to improve coverage.
This epic records the 2026-09-28 source/receipt diagnosis of main d5ca636a48e963619f4c89e2958d0bb1a3e72f80. Reinspect current main and existing local work before implementation. The previously merged fixes are present; they did not pass the corpus preservation gate.
Verified starting point
The frozen set contains 200 first physical pages: 111 complete abstracts, 7 partial, 82 without a first-page abstract. Labels are assistant-reviewed and already examined; they do not provide independent held-out qualification.
| Method | Correct under historical text98 metric | False proposals | Complete abstracts withheld |
|---|---|---|---|
| Frozen structure v2 | 91 | 5 | 16 |
| Parallel structure v4 | 38 | 0 | 73 |
| Parallel GROBID v4 | 33 | 0 | 78 |
| Parallel MinerU v4 | 38 | 1 | 72 |
| Parallel olmOCR v4, latest rerun | 39 | 0 | 72 |
The latest run verifies all 200 original source/page/image identities and replays v2 successfully. It loses the same 55 previously correct primary cases as the historical v4 run. Of the 73 withheld complete abstracts, 28 already contain text matching the reference at the historical threshold. First recorded reasons for the 55 losses: 29 column continuation, 10 missing closing boundary, 6 alignment, 5 missing candidate, 2 style transition, 2 structured subsection ambiguity, 1 terminal notation. One column case also exceeds the request character budget.
Confirmed mechanisms include same-line fragments mistaken for columns; narrow Abstract-heading boxes hiding Introduction boundaries; correct initial paragraphs extended into body text; paragraph-level candidate rules; and structured-abstract subsection errors. Separate in-memory diagnostic interventions recovered 12 boundary-correct cases each and corrected MinerU f039. These interventions were not combined/full-corpus validated and are not a production patch.
Two evidence corrections matter: historical 91 text98 successes contain only 90 boundary-plus-text98 successes (f195 has a trailing footnote marker). The latest olmOCR 40→39 difference is caused by changed native extraction evidence under a different runtime; it does not explain the persistent 55-case primary regression. Normalized alphanumeric agreement is insufficient to prove formula fidelity.
Evidence and reproducibility
- Public 200-paper comparison, earlier failure audit.
- Public v4 integration receipt, experiment report, reproduction instructions.
- Authorized external data layout:
validation_v2/,validation_expanded200/, andvalidation_v3_private/parallel-v4-source-map.jsonbeneathTRACE_GC_TEST_DATA_ROOT; original PDFs, renders, labels and conversion caches remain outside Git. - Latest external result:
validation_v3_private/parallel-v4-regression-rerun-20260927-220439/results.json, SHA-256661964b8db10bf2bd806a95a0119f4492dcabf72e759fff521993f702f016b0d. - Historical result:
validation_v3_private/parallel-v4-regression-03/results.json, SHA-256f9ee4f438c8cae7ce192970f7ff172ecbe7c24b3a383d48fe62cf75da7f4102b. - External diagnosis directory:
validation_v3_private/diagnosis-20260928/, containingROOT_CAUSE_AND_SOLUTIONS.md,summary.json,receipt-audit.json,selector-causes.json,src/selector_probes.pyand their manifest. The environment/metric receipt SHA-256 iscab1d4c2048ce6de1006975087bec29adc5a470f76f97c491564a7f255387e03.
These external artifacts are access prerequisites, not downloadable public attachments. Obtain the authorized data root from its owner and verify the identities before real-corpus work; missing inputs must yield an explicit blocked result. Public fixtures can validate code but cannot replace the corpus. Do not upload private papers, labels, full predictions, keys or personal filesystem paths into issues or PRs. Publish compact permitted aggregate/hash receipts.
Completion criteria
- Runtime, imported code, native evidence and converter identities are frozen and checked before expensive evaluation.
- Geometry, section boundaries and multi-paragraph/structured abstract repairs are delivered with causal positive and adversarial tests.
- A separately reviewed image-backed OCR evidence route handles absent/corrupt native text without invented offsets or automatic publication authority.
- Historical text98 metrics remain reproducible; strict boundaries and scientific notation fidelity are measured separately.
- The complete configured pipeline passes the 200-paper preservation/error gates, with every residual case and regression reported. Source-verified criterion approved in Require a full 200-paper preservation and fidelity receipt before selector promotion #19; the legacy f081 encoding failure remains preserved (current evidence).
- The frozen system receives an unseen, blinded internally reviewed quality evaluation with preregistered risk/coverage criteria and accurate reviewer attribution.
- Source-bound retrieval fields are connected and the real 10,000-paper/60-query trial measures candidate coverage and ranking independently.
- A bounded real-paper pilot demonstrates evidence admission, durable graph publication, retrieval and supported generated answers, including abstention and update/replay behavior.
- Required implementation PRs are reviewed, validated and delivered; retained receipts identify the final relevant source revision. Each child is complete against its acceptance criteria before this epic closes.
Coordination and scope
The native sub-issues and dependency edges define completion prerequisites. Preparatory design and fixture work may run in parallel. Keep one active writer per shared source area/index; use scoped handoffs and preserve unrelated checkout changes.
Related initiative: #4 and its domain-classification/storage children #5–#11. This epic reuses delivered graph contracts and does not duplicate domain taxonomy, partition decisions or deployment adapter work. Domain-aware behavior requires the relevant #4 children if selected for the pilot. A local reference-backed real-data pilot may proceed without production #11; it must be labeled as such, and production qualification still belongs to #11.
The first-page benchmark remains first-page-only. Complete-document ingestion and later-page evidence are separate pilot behavior. Model agreement is not independent truth; exact provenance is not semantic correctness. The pilot must use an authorized real provider/model and actual trained service when relevant. Credentials alone do not grant new paid compute, publication targets or private-data transfer authority; identify genuinely missing inputs while completing independent work.
Linked delivery plan
Completion dependencies are also recorded through GitHub's native blocked-by links. Design and preparatory work may proceed in parallel.
| Issue | Delivery | Completion prerequisites |
|---|---|---|
| #13 | Freeze the extraction runtime and native evidence; validate replay configuration before assessment | None |
| #14 | Add versioned notation and source-boundary fidelity metrics without rewriting historical scores | None |
| #15 | Fix false column transitions by reconstructing logical native lines with source provenance | #13 |
| #16 | Fix abstract section ownership and reading order to exclude body and metadata contamination | #13 |
| #17 | Support complete multi-paragraph and structured abstracts with evidence-based closure | #15, #16 |
| #18 | Add verified image evidence for OCR when native PDF text is absent or corrupt | #13, #14 |
| #19 | Require a full 200-paper preservation and fidelity receipt before selector promotion | #13, #17, #14 |
| #20 | Qualify the frozen extraction configuration on an unseen internally reviewed paper cohort | #19, #18 |
| #21 | Wire source-bound abstract assessments into retrieval field generation | #13 |
| #22 | Rerun the full 10,000-paper retrieval trial with bound fields and candidate-stage metrics | #21 |
| #23 | Implement and validate a real paper-to-graph-to-cited-answer pilot | #20, #22 |
The regression gate may first qualify native/cached arms. Adding the image-evidence route requires a new receipt for the changed configuration before unseen internal qualification. Completion of the local real-paper pilot does not close production adapter #11.
Activity
7 remaining items
CompleteDotTech commentedon Sep 29, 2026
#19's amended development-regression gate is delivered and closed via PR #83, verified merge
92c73419. Fresh main-tree verification and all 6 reviewed-head plus 3 default-branch checks pass. All 91 original IDs remain source-verified; 97/111 complete abstracts selected, zero source-false selections. Frozen legacy result remains FAIL for f081's source/reference encoding conflict; references and scorer are unchanged. Detailed acceptance: #19 (comment).#20, #22 and #23 remain OPEN. The 10,000-paper assessment run stopped after 108 durable rows and 109 sealed conversion attempts. A reviewed recovery branch fixes the source-owned paragraph verifier and prepares immutable, versioned conversion reuse; required full validation is pending before publication. No completed retrieval trial or cited-answer pilot is claimed. Model probes remain held under the current memory/resource guard. A changed measured closure will receive a fresh #19 receipt before downstream qualification.
CompleteDotTech commentedon Sep 30, 2026
Delivery and empirical checkpoint — 2026-09-30
4ab4fb5e6e8ca684478a874b2699fab8e1228063: verified signatures, all six final-head checks and all three main checks passed; local full suite 1,303 tests / zero failures or errors, and 25 post-merge lifecycle tests passed. It bounds future converter sessions and verifies actual Windows worker ownership/exit.Core PaperPilot delivery was separately mapped: all 125 core-library/runtime-schema files are byte-identical between its frozen reviewed checkout and the new merge. This mapping does not rebind the broader extraction/converter gates or invent a new model run. Source reviews are attributed assistant reviews under the user's waiver; no independent-human qualification is claimed.
The epic, prospective qualification, retrieval, and final paper-to-answer issues remain open. Detailed private receipts and rejected attempts remain in the durable execution ledger; public aggregate updates are linked on #53 and #23.
CompleteDotTech commentedon Sep 30, 2026
Verified source phase; empirical completion remains outstanding
The seven-claim, eight-question real-paper pilot has independently reproduced its selected source checks: 8/8 passed, with the native quantitative notation route retained on HOLD and its separately reviewed image route selected. Root executed the exact pinned replay helper; its result was byte-identical to the earlier execution. The logical contents of four SQLite stores and the ACL were unchanged before/after. Reviews remain attributed assistants, with no independent-human qualification claim.
Scoped measurement receipt:
faa1fc639c5252fea375d38cb09ae3034160bdcc00dc9650bfde293ce0afe5ed; root executed review:bbb922d5a2921b356b4ecf11cdd79bc775f10ecb760abcf5b00087161f6ebf0b; peer addendum:537ced4b21352464dc9d9f75fdc78744e326082eb526956b1f24e715bde03544. This measures the source phase only. It does not qualify unseen extraction, semantic assertions, retrieval, generated answers or recovery.A supported public lifecycle service is being implemented on an isolated branch from current main
31268c8b6fb03fa06b811dfecf876052590df2dc. Review identified historical local-import freshness, immutable status-receipt replay, and complete parent-archive checks to repair before final review and merge. The aggregate evaluator is also receiving authenticated raw/import/checkpoint and attributed abstention checks with adversarial disposable API fixtures. Fixtures establish mechanics, not real-model accuracy.The actual pilot still has eight admitted sources, graph version 1, zero assertions and zero executed model calls. The full current-source 200-paper/image gate, unseen qualification, complete 10,000-paper/60-query retrieval trial, real semantic/answer execution and postpublication lifecycle acceptance remain outstanding. No numeric criterion or resource floor has been reduced, and no empirical issue or this epic is being closed.
CompleteDotTech commentedon Sep 30, 2026
Recovery-service delivery checkpoint
PR #88 is merged at
d908283d94d44c22b8427b5ad463493424c5c967. The reviewed/tested source tree matches delivery; all six PR checks and all three default-branch Python checks passed. Local validation recorded 1,307 passing tests and eight explicit optional/external skips, followed by four passing post-merge service tests. All 582 audited source-package members match delivered files.This delivers the recovery service, not empirical qualification. #20, #22 and #23 remain open. The approved #19 source-preservation policy retains all 91 original IDs and unchanged legacy scores, including the reference-encoding failure.
The next full-200 execution uses a hash-verified input mirror and an independently reviewed launch wrapper. No new corpus-quality result is claimed: launch remains subject to the original sustained resource guard. Private recovery review found and is correcting a circular parent/recovery acceptance requirement and unknown-outcome status reconciliation. All nine actual recovery scenarios remain unqualified until authenticated execution evidence is available. Historical failures and receipts are preserved.
CompleteDotTech commentedon Sep 30, 2026
Delivered extraction work and current integrated source gate
The current default branch is
d908283d94d44c22b8427b5ad463493424c5c967(merged PR #88). Default-branch CI run 36683988151 is successful. The delivered runtime/image-evidence repairs and versioned fidelity contracts remain separate from the empirical qualification work.The fresh integrated 200-paper source-verified gate and its separate receipt replay both passed on this exact delivered runtime. Receipt
3b1406c2036f8f9795f5845e910d99aef5688997a627b28d4f454345d8637063binds all 83 method files and 124 image-code files. Results: 97/111 complete abstracts source-verified (73 native, 24 reviewed image), 14 withheld, all 91 original success IDs preserved, and zero source-verified false proposals. All 200 pages and four cached arms remain accounted for. Detailed gate evidence and retained legacy result.The unchanged legacy base audit remains FAIL solely for f081's source-correct Δ versus historical ∆ encoding. The separate source-verified criterion is the owner-approved #19 policy; the legacy scorer, reference and failure were preserved. These are attributed assistant-reviewed, exposed development results, not independent human or unseen qualification.
Remaining acceptance work:
The new conversion supervisor is held during fixes for receipt-field validation, atomic drain requests, owned-child cleanup and the original hard cutoff. No shared workload was terminated, no WSL VM was restarted, and no paid/offhost inference or production graph writes were performed. This epic and its qualification milestone remain open.
CompleteDotTech commentedon Sep 30, 2026
Delivered repair and fresh integrated source gate
PR #92 is merged at
6b62693c338d222b0c415ec075a14afc543a1a53, with verified signing, exact-head CI and all three post-merge Python 3.11/3.12/3.13 CI jobs passing. Twenty-one targeted post-merge tests pass. The delivered tree is identical to the measured tree; all 85 method files and 129 image-related files rehash exactly.The fresh 200-paper run and 24 fresh local OCR/source-review routes now have an integrated source-verified acceptance receipt and successful full replay verification: 97/111 complete abstracts verified, 14 withheld; all original 91 success IDs preserved; 7 partial and 82 absent cases retained. Every arm contains 200 audited cases. No scientifically false or unverified selected proposals; no paid calls or production graph writes. Review is attributed assistant review, with no independent-human qualification claim.
Source-verified receipt SHA-256:
984c97957067655710121760bb20128e204f1fa85a7b58ea87304c57dd7e4eae. Policy SHA-256:010a19d1e8b28b06d1babe62464b441a2a7ac1bf47845fb82345b2c519760065. Legacy base receipt SHA-256:d514233c30eaf4f94c7be8a22f4ccda9fd52090e74f20385fd8cc7fd92a76f5f; it remains FAIL solely for the previously approved f081 Δ/∆ reference-encoding mismatch. Frozen reference bytes and historical scoring remain unchanged.The original full-corpus producer window expired before recovery could start. No successor was launched or completed conversion retried. The reviewed V3 plan preserves the 2,333-row prefix and sealed pending conversion. Issue #91 remains open for actual useful recovery progress; #53/#22 remain open for their full-corpus and retrieval criteria. #20/#23 remain open for unseen qualification and actual graph-answer acceptance.
CompleteDotTech commentedon Oct 1, 2026
Delivery progress
The previously published 200-paper source-verified result remains preserved as historical evidence. Worker/field/reconciliation method changes require a fresh bound gate before current-code qualification. No unseen-paper, full 10,000-paper retrieval or real generated-answer acceptance has been completed by these deliveries. Converter reconciliation and the current-trust/cumulative-budget pilot guard remain under explicit review.
CompleteDotTech commentedon Oct 1, 2026
Reviewed reconciliation boundary delivered
PR #100 merged as
24e1ee062103163e08d2c02669aee3bf4258fecf. The reviewed head8df3b45e879811a30a49e75cfbd54d4373d57c70and hosted merge have verified signatures and identical source trees. All six exact-head hosted checks passed across Python 3.11, 3.12 and 3.13; default-branch run 36802011453 is currently running.Final Windows validation: 1,388 run, 1,380 passed, eight explicit skips, zero failures/errors. Skips are two optional NumPy tests, five Windows symlink privilege tests, and one unavailable external pinned-core integration. Post-merge converter checks: 43 run, 42 passed, one Windows symlink skip, plus legacy package validation PASS. The source package audit and fresh installed-package CLI/import checks also passed.
The delivered API verifies reviewed old-attempt reconciliation and exact distinct-successor/application bindings; it retains historical unknown outcome and disables standalone caller-supplied effect approval. The original failed attempt remains immutable and
UNKNOWN_UNATTESTED_NO_RETRY. The scoped code/package reviews are attributed assistant reviews, not independent human empirical qualification.#94 remains open. The new exposed f097 plan is being directly validated against the delivered public source before root-owned signing, enrollment, live exclusion and any genuine Docling execution. No successor converter has run. #20/#22/#23 remain empirically incomplete.
The method closure now contains 87 pinned files. A fresh full 200-paper cached-arm regression has started against the delivered merge in a new immutable output, with the original pinned replay runtime and fresh native extraction. Previous legacy and source-verified results remain preserved; they are not substituted for this current run.
CompleteDotTech commentedon Oct 1, 2026
Current empirical and delivery checkpoint
24e1ee062103163e08d2c02669aee3bf4258fecf; no completed quality result is claimed. Fresh local OCR was run for all 24 development image cases, and a separate attributed assistant visually reviewed their new source/crop/transcription evidence. The new reviewed handoff index SHA-256 is109c210cc56c811df6e5319d17a15f9a7182d968da2c87cba2598a4a4888b8f8. Its binding to the new native assessments and integrated acceptance is pending.UNKNOWN_UNATTESTED_NO_RETRY. A distinct exposed-page local successor release and launcher are now pinned and separately reviewed, but no new worker has started. The unchanged 20 GiB free-RAM guard is not met.The unseen cohort, full retrieval trial, real cited-answer pilot and actual transition acceptance remain incomplete. Existing historical failures, reservations and expired-run cutoffs are preserved. The epic remains open.
CompleteDotTech commentedon Oct 1, 2026
Current delivered 200-case source acceptance
The fresh native four-arm run completed all 200 cases per arm with zero paid calls and graph writes. Its saved replay passed, but the native preservation gate failed: four previously correct cases were withheld. The native result remains
FAILED_REGRESSION_PRESERVATION_GATE.The separately reviewed current image route selected 24 cases; 73 native cases and those 24 image cases form 97 source-verified proposals out of 111 complete abstracts. Fourteen complete abstracts remain withheld. All original 91 correct cases are included. The base receipt remains
FAILon itszero_false_primary_proposalsgate for f081; the source-verified receipt isPASS, with 97 reviewed-fidelity proposals and no measured scientific false proposals. The actual publicsource_verifycommand exited 0 against the exact source receipt.This is a previously examined development cohort and assistant-attributed source review. It does not establish independent human qualification, graph admission, a 10,000-paper trial, or a new model result.
Current receipt identities
Measured delivered source:
24e1ee062103163e08d2c02669aee3bf4258fecf.Source-verified acceptance SHA-256:
3f77057a0d845b71320003cfc64553cc429b762abfa75c2ec86becf19c507253.Policy SHA-256:
83d59e9738da9857c058fe9513eebd99c2ed421da1457afef03fb87c41e9d072.Unchanged failing legacy base SHA-256:
ad532faa89cf47c811afce0fca7b0cc493fc9050de574c0405021dab257c0157.Fresh native regression SHA-256:
8c82a79fc44465bf08a26ea80873304f5aaf9108fe6a4bdb2cc1cdd9e6cb875c.Scoped assistant peer readback SHA-256:
f7319787375bb90a73c7dff736cc7550fd764acaf3bd245e499a27d5183a6f09.The failing legacy comparison retains the previously approved f081 Δ/∆ reference-encoding exception. Reference bytes and historical scoring remain unchanged.
Issue #91 remains open: the original full-corpus producer window expired, and no recovered producer was started. Issues #20, #22 and #23 retain their separate actual qualification requirements. This update does not extend the execution window or authorize retries of historical unknown outcomes.
CompleteDotTech commentedon Oct 1, 2026
Initial transition authorizations issued and verified; transition not executed
The corrected read-only ownership observation succeeded: all five historical state-file identities matched, four database read handles excluded writers while all three old locks were held, and no historical owner process matched the census. Runtime bytes were checked before and after the observation. This is preparatory evidence; the actual fence must repeat the observation continuously under its own locks.
The root-owned sign-only operation then completed with exit 0. Both Ed25519 authorizations were independently replayed against the enrolled public key and bound to the exact quiescence/fence release fingerprints and frozen capsule. Their validity window ends at
2026-10-01T05:51:25.133605Z. The operation stopped before invoking the owner transition.Evidence SHA-256:
5a37b89976901dd76d1f8c8471f369c0ecc6483ea14b56bc883f690490aec8e0.f0e4551240389bc9e78ead915abad0c6b6cbe64fb1faa7822d587be939cd03a8.c46fa7c3956351de4721d1e68e5fc4a58a74bd6a7075d562849e31abe8c45c47.Two earlier pre-effect failures are preserved: a second-handle read of the already locked controller file, repaired with a peer-reviewed held-handle reader and a real disposable Windows fixture; and ignored bytecode left by post-merge tests, moved intact to a reviewed quarantine with matching file hashes. The 99 reviewed application files remained unchanged. Neither failed attempt issued signatures or performed a phase effect.
No new model call, source admission or assertion publication occurred. The historical unknown model reservation remains consumed and is not retried. The owner-effect launcher is being reviewed separately; #90 and the empirical qualification issues remain open.
CompleteDotTech commentedon Oct 1, 2026
Actual held quiescence/fence completed; later stages remain outstanding
The root-owned operation completed with exit 0 under the exact reviewed owner release and two verified stage authorizations. It made a fresh live ownership observation while all three old locks were held, retained four database handles excluding writers through the durable admission fence, and archived the exact five historical state files.
Fresh root verification confirmed all five archived byte hashes, restoration of their original effective file DACLs, the exact parent deny-create change, and absence of the old fixed state path. A fresh attempt to recreate that path failed with Windows access denied (
WinError 5). The file denies protect the admission/rename sequence; the durable old-parent deny protects the old fixed namespace afterward. This is the documented cooperative application boundary.An independent assistant reviewer separately rehashed all five archived files and the quiescence/barrier/fence/owner receipts, with zero mismatches. The reviewer did not repeat the root's namespace-creation probe.
Evidence SHA-256:
41b725cf4d9669b598ce1c0bd7e93d7c17373d38f1d2e86fd5aabcdf69e3d170.133d37ba79379514d25cc2155385080f9bfe659bf308f51c132bcf5ec8edbf50.062fa6f0e6ec7d132962b491339832841855a462f6b65d2584f3dedb4508626c.8c313280b1dbb29b1a23645dae83537ce8bcb78f5e2ebe0169893270c9484e99.aec8d8cca76a0d7e04eb9cc7d9bc23abc17608a4096f85e14ed020e02b90b31d.578feb3fa7addd90891912a66fd60d7ec404cd2ce931ce6835d21b43444aeb22.The historical budget was reopened read-only and still contains the consumed model reservation of one. Its outcome remains
UNKNOWN_UNATTESTED_NO_RETRY. No new model call, source admission or assertion publication occurred. Restore, import, activation and reopened-state verification remain NOT RUN; their exact releases and authorizations are being prepared. #90 and epic #12 remain open.CompleteDotTech commentedon Oct 1, 2026
Delivery and actual recovery checkpoint
31b21b6c9035356a5f106a74c31d94eddb51d297: three main CI jobs and 31 Windows phase tests passed. Fix SQLite restore readback sidecars that block interruption recovery #102 stays open because its final criterion audit found two negative-test gaps.3313c3144011f8cc1244bf525b621f1963508e96. Seven focused checks pass; final-head full validation and hosted CI are running. The first archive-based full attempt failed because tests require Git metadata and is preserved as failed.Temporary checkouts and private test artifacts are retained in issue-specific directories outside the project source tree. The original modified checkout and historical signed phase source are preserved.
CompleteDotTech commentedon Oct 1, 2026
Progress update following #102: the SQLite restore readback repair is delivered and #102 is closed. PR #104 merged at
7f01f338; postmerge CI passed on Python 3.11, 3.12 and 3.13. The final signed-head gate ran 1,403 tests: 1,395 passed, eight documented optional skips, no failures or errors. PDF coverage passed. All 619 source-package files matched signed Git.The current helper read the actual closed historical pilot capsule without changing it: four backup databases matched their byte and logical pins, and the original event chains, source statuses and one reserved model call remained accounted for. That readback did not restore or activate a new phase, call a model, or resolve the historical provider outcome. The earlier frozen phase and its sidecars remain preserved; #90 tracks the remaining real transition.
The 200-paper source-verified result selected 97 of 111 complete first-page abstracts and preserved the original 91 source-verified cases. This remains a historical measured configuration: a reviewed current-code mapping is disabled and has not run as a new acceptance gate. The legacy encoding failure remains disclosed. #20 unseen blinded quality, #22 full source-bound retrieval qualification and #23 real paper-to-graph-to-cited-answer qualification are still open. The 10,000-paper producer has 2,333 preserved rows; its original cutoff expired, so no restart or extension is inferred.
An application-owned status signing step for the phase transition has been launched separately. Its terminal result was pending readback when this draft was prepared, and signing alone grants no restore, activation, model or graph authority. Epic #12 remains open.
Separate actual predecessor receipts already verify the recovered import, activation and fresh historical reopen. They preserve the old reservation and UNKNOWN provider intent. The public #102 repair does not replace those receipts or authorize a repeat restore. The initialized status anchor remains revision zero with no committed event; the sign-only step is a separate pending operation.
CompleteDotTech commentedon Oct 1, 2026
Current delivery and execution checkpoint
Current
mainis7f01f3389e3360ba89ac780f585c9f83c9ec2aae; hosted default-branch checks passed. #102 is closed after the reviewed sidecar repair, package validation, and post-merge readback. Epic sub-issues #13–#19 and #21 are closed; #20, #22, and #23 remain open.The genuine 200-paper gate was verified again in its pinned historical runtime. It retains 97/111 complete abstracts accepted, zero false proposals, and source-verified preservation of all 91 original success IDs. The legacy reference-encoding failure remains reported unchanged. A reviewed code-identity mapping connects the unchanged measured extraction methods to current main; this is evidence mapping, not a new extraction run or unseen qualification.
The first reviewer-status commit failed before its event write because a required validity-window field was missing. The failure was preserved, the mandatory-window repair was tested and reviewed, and the actual same-receipt recovery committed durable revision 1. Root and separate assistant reviews verified the public signature, event, pointer, anchor, and unchanged historical target files. The receipt subsequently reached its fixed expiry; an actual read-only check confirmed that it is rejected as current authority. A separately reviewed chained renewal is required before later execution.
Current work is completing and reviewing the current-code pilot dispatch, complete historical-state readback, chained status-renewal path, and metadata-only unseen-cohort preparation. No unseen cohort has been selected or inspected and no new model call or graph write occurred in these recovery steps. Current resource checks fail the registered RAM and GPU-idle thresholds. Full assessment-map completion, genuine unseen qualification, full retrieval acceptance, and the real generated-answer pilot remain outstanding. Historical unknown provider attempts remain reserved and are not retried.
Recovery evidence: root
0844180d14d7d342fd4a88c1d7297230c5a0a9a032d4566390d024b461dbb327; independent assistant review00e096664fbcc1d2b6d99160b86f8503945131da615fb9f248c81fbe1d0caf14. See the detailed #90 recovery checkpoint.