cesnimda
6.9 KiB
Integrated qualification progress
29 October addendum
The mandatory D owner-addendum audit identified missing diagnostic filters and identifiers in the initial checkpoint. D_ADDENDUM_ACCEPTANCE.md records the completed changes, reproduced failures and current independent evidence: backend 1,447, frontend 1,031 across 112 suites, TypeScript/lint/build exit zero, and four focused browser cases (including actual 100/200/400% zoom) passed. Final MariaDB is being rerun; the first final attempt could not start Docker and reached no tests. The source delta is separately reviewed and lies outside the original sealed security scan.
Fresh deterministic interpretation/scoring is now complete for the 36 source-bound structural documents: PDF typed precision 100%, recall 81.97%, F1 90.09%; DOCX typed precision/recall/F1 100%. Both formats recovered 134/134 expected literal fragments and correctly abstained in all six designated ambiguous cases. PDF provenance is 24/27 and DOCX 27/27. PDF unresolved-source classification remains limited (16 true positives, 27 false positives, two false negatives); text recovery alone does not establish record reconstruction. Evidence: .artifacts/pre-phase10-ux/integration-scoring-02/, including separate frozen-split results. No generative model was invoked; synthetic corpus results do not estimate population accuracy or qualify reviewed persistence/real AI.
The earlier rows below are preserved as historical evidence, not final-source results. Full ordinary/specialized browser, provider/PDF closure, local C/D activation and final handoff remain pending.
28 October checkpoint
28 October 2026. Cumulative branch test/pre-phase10-ux-acceptance contains A/B/C/D through source d52403d, test-fixture correction 4dcc07b and security evidence 019d0aa. This is an intermediate evidence checkpoint, not final acceptance or permission for Phase 10 implementation.
The verified preceding CV-QA baseline is backend 1,336, frontend 995 in 109 suites, ordinary browser 364, specialized browser configurations 53, real MariaDB 80 and Python 76 offline plus 14 real parser tests. Historical evidence remains in docs/cv-integration-qa/; CV-UX commit 4a0af62 is an ancestor. The newer C/D focused evidence must not be added to overlapping full-suite counts.
Fresh completed independent lanes under .artifacts/pre-phase10-ux/:
| Lane | Result | Evidence |
|---|---|---|
| Full backend | 1,424 passed, zero failed/skipped | integration-01/backend/backend.trx |
| Full frontend | 1,027 passed in 112 suites, zero failed/pending | integration-frontend-01.json |
| TypeScript / full lint | exit 0 each | integration-{typecheck,lint}-01.log |
| Offline sidecar | 81 passed, no model loading/network | integration-independent-01/python-offline.log |
| Real Docling parser | 14 passed | integration-independent-01/python-docling.log |
| Benchmark oracle self-checks | original 11, structural 9 | integration-independent-01/benchmark-*.log |
| Node helper contracts | 25 passed | integration-independent-01/node-helpers.log |
| Linux qualification lifecycle | 11 passed after synthetic-fixture correction | integration-linux-helper-02.log; original failure retained |
| Guarded operation fixture | 29 self-checks passed | existing seed-operation.py --self-test, temporary databases only |
| Structural conversion | 36/36 actual frozen documents, 89.283s total; peak child RSS 150,408 KiB | integration-structural-01/{manifest,corpus-binding,resources,process}.json |
| Security diff | sealed, 136 changed files, no reportable findings | SECURITY_REVIEW.md |
| Advisory checks | npm production zero; full npm five low dev-tool entries; Python import lock zero | integration-independent-01/ |
Structural conversion verifies the original document bytes and performs sequential real extraction in a read-only, network-disabled pinned container with bounded CPU/memory/time and no model. The annotation binding accounts only for Git's CRLF-to-LF checkout difference; expected annotations were not changed. Literal recovery, typed facts, record reconstruction and provenance scoring remain a separate subsequent step. RSS is a maximum observed child peak, not a whole-application memory bound.
Ordinary browser discovery is 380 tests in 66 files. The first serial group retained two 240-second zoom-extension startup failures. Trace shows login rendered and then waiting for the extension service worker; bundled Playwright defaults disable extensions. The scoped test-only correction explicitly excludes that default flag for this extension-owning browser context. It must pass both EN/light and NB/dark real zoom cases before publication; no zoom assertion, timeout or geometry oracle is weakened. The original failed run remains evidence. Passing unrelated cases will be reconciled by exact discovery identity; overlapping recovery runs are not added as extra tests.
The remaining ordinary groups, specialized browser contracts, final real Docling journeys/MariaDB, native Windows/Linux PDF proof, integrated source/runtime closure, final local source activation and acceptance/handoff are pending. The existing local app still serves qualified B source 3f4db8c at http://jobbjakt.local; C/D are not yet claimed as locally activated.
preservation.json carries all 57 original capability criteria without promoting their 56 PARTIAL/one DEFERRED statuses. Source membership preserves all 33 original routes and adds /admin/processing (34 total). Route membership alone is not functional acceptance. Human/device/language/contrast/manual-print debt, revised large-name PDF coordinate-oracle failures, CV generative accuracy, exact deployed model parity and production/operator qualifications remain open. Production remains NO-GO; no main merge, deployment or Phase 10 cleanup occurred.
Final dated addendum — 2026-10-29
The pending automated lanes above are now reconciled in ACCEPTANCE.md and evidence-summary.json: backend 1,447; frontend 1,031/112 suites; ordinary browser 380 distinct; specialized browser 53; real Docling browser eight; MariaDB provider 90 plus three additional browser cases; Linux browser eight; Python offline 81 and real Docling 14. The report explicitly preserves original failures and exact test-only recovery accounting. Revised PDF S01 remains PARTIAL. Frozen deterministic extraction scoring is complete with PDF typed recall 81.967%, DOCX 100%, and the documented PDF review burden. Real CV AI quality remains BLOCKED.
The guarded cumulative A–D local activation replaced only the development backend/frontend/sidecar using qualified immutable images. The running source is 994bf8f; actual served frontend bytes and backend /health were verified. Fresh real local worker and model outcomes, cleanup and source override instructions are recorded in LOCAL_RUNTIME_ACCEPTANCE.md. The historical B-only activation statement above remains a record of its original checkpoint.