cesnimda
12 KiB
PRE-P10-UX cumulative acceptance — 2026-10-29
Decision: CONDITIONAL GO for development review and the next QA handoff. Production: NO-GO. Phase 10 implementation is not authorised. This dated report supplements the historical A/B/C/D reports; it does not replace their original evidence or promote conditional acceptance.
Scope and source
The accepted CV-QA baseline is 4a8c907d70cbed235728f7d690b5409e72b93abe, containing CV-UX 4a0af62. The cumulative A–D application source is 3255da9fc6e5d885432bed60f37f4cc47b7dd00d; the broad regression freeze is 994bf8f8bbf0df3c53111ca7f66af97fe0680013. Changes through that freeze affect test compatibility and evidence, not application behavior. A subsequent real local smoke exposed a strategy JSON contract conflict, corrected in caa0c6a2efea5a6e9d8910f55fb00fd982cd48fd; see the narrow addendum. The machine-readable evidence summary binds the broad results and retains limitations. Raw evidence is under .artifacts/pre-phase10-ux/ and .artifacts/cv-ux-01/ in this worktree.
Published checkpoints:
| Branch | Published source/checkpoint |
|---|---|
codex/cv-builder-blank-undo |
663023a088573e44abea8da1c1858da934af77e9 |
feature/cv-template-library-review |
3f4db8ca8183704945915ae4d2053e641bafa5b2, activation addendum aea7c29 |
feature/dashboard-interactive-grid |
0bd290e788fc525157fe59ca830d39d5748fd192 |
feature/admin-processing-monitor |
3255da9fc6e5d885432bed60f37f4cc47b7dd00d |
test/pre-phase10-ux-acceptance |
final evidence and operational handoff continue from 994bf8f |
No main merge or production deployment occurred. Unrelated changes in the primary checkout are preserved.
Delivered behavior
Blank CV creation now starts genuinely blank, with explicit Career Profile selection and bounded document undo/redo. Existing inherited CV contracts remain compatible. New or explicitly applied template designs have increased title spacing, full-height sidebars, actual rendered gallery previews, matching cover letters and bounded photo handling. Historical designs remain unchanged until the user explicitly applies an updated design.
The dashboard uses revision-safe persisted geometry, actual drag/resize and keyboard controls, responsive linear layout and compact disclosures. The administration monitor exposes owner-scoped, redacted operation state and structured history with bounded polling, filters, safe job links and genuine failure timestamps. Worker operations propagate correlation IDs to the sidecar.
The former upload-to-queue dead end has an enabled local document worker and real Docling extraction. The supported workflow is upload → actual processing states → source-grounded review/correction → explicit approval → revision-safe Career Profile → CV Editor → PDF. Canonical profile facts, frozen CV versions, CV overrides and application associations remain independently protected. External AI consent and account entitlements still apply; worker configuration does not waive them.
Fresh automated verification
| Lane | Result |
|---|---|
| .NET backend | 1,447 passed, zero failed/skipped |
| React frontend | 1,031 passed in 112 suites |
| Ordinary browser | 380 distinct cases qualified by exact-identity reconciliation |
| Specialized browser | 53 cases: AI/auth/rate/ingress/admission/preferences/RD21/diagnostics/volume |
| Real Docling browser | 8 distinct journeys, including long CV and historical-document preservation |
| Real MariaDB provider | 90 passed, temporary-schema cleanup confirmed |
| Additional MariaDB browser | 3 passed: save/conflict/network recovery and concurrency observations |
| Linux browser | 8 passed, immutable image/source/DLL binding and cleanup |
| Python | 81 offline contract/security tests; 14 real Docling tests |
| Other deterministic checks | benchmark oracle 11; structural helpers 9; Node helpers 25; Linux lifecycle 11; seed guard 29 |
| TypeScript / lint / frontend build | passed; local /api build separately passed |
These totals are separate lanes, not one additive test total. The ordinary browser result is a composite: integration-04 retained two old smoke assumptions; integration-05 reran all nine corrected smoke cases and the remaining 95 cases. Only the two failed identities are replaced; the overlapping seven are not counted twice. Specialized diagnostics retained four obsolete always-visible-control assumptions; all 18 passed after a test-only update in integration-06. No automated retries were used. The assertions now exercise explicit Career Profile selection and keyboard-opened dashboard disclosures; source/runtime binding proves application behavior was unchanged between these recovery runs. See smoke compatibility and diagnostic compatibility.
The real zoom extension harness passed both language/theme cases and six repeated cases after correcting extension startup and network-guard behavior. One long-path diagnostic launch succeeded; that alone does not exclude every path-related failure. The short owned temporary profile remains a robustness measure.
After the narrow strategy correction, the complete backend suite passed 1,448 tests and the complete offline Python lane passed 83 tests. One intervening backend run retained an outdated untyped strategy mock; its failure and the corrected full rerun are preserved. The frontend source/rendering/persistence schemas did not change. Earlier frontend, browser, provider and PDF results remain bound to their original freeze; they were not all rerun after this narrow AI-mode change. Actual local queue/Docling and direct sidecar checks provide the additional runtime evidence, with limitations in the local report.
RD21 checks cover stale writes, retry/recovery, transaction rollback, ownership and confirmed persistence. The extra MariaDB observation explicitly records non-idempotent repeated creation bodies; it is not evidence of creation-intent idempotency.
Actual PDF evidence and remaining failure
Each platform passed the original 24 CVs/60 pages and four cover letters/14 pages. Revised output covered 36 CVs/122 pages and 18 matching letters/38 pages. Source completeness, bounds and name wrapping passed; sidebar raster coverage passed all 43 relevant pages on each platform. Additional imported PDFs retained their original semantic assertions, including three long documents with 142 source markers and 12 employment-record associations each.
The unchanged revised S01 coordinate oracle remains PARTIAL: Windows 8/36 pass, Linux 16/36 pass. The remaining 28/20 cases fail only S01's 2 mm positional threshold. A separate visible-ink check within three pixels and inspected raster contact sheets do not waive that failure. Human typography, printing and physical-device acceptance remain open. No claim that all PDF fidelity checks passed is made.
Extraction benchmark and performance
The frozen synthetic corpus contains 18 logical cases in PDF and DOCX, 36 actual source-hash-bound files. Docling versions are 2.142.1, core 2.104.0 and parse 7.23.0. Structural conversion ran without models or network, with read-only storage and CPU/memory/time bounds. All 36 converted; 89.283 seconds total, median 2.4381 seconds, p95 2.89724 seconds, maximum 3.63939 seconds. Maximum observed child RSS was 150,408 KiB; extracted output totalled 31,664 bytes. This is not whole-application resource or production-load qualification.
| Metric | DOCX | |
|---|---|---|
| Literal facts | 134/134 | 134/134 |
| Typed facts TP / FP / FN | 100 / 0 / 22 | 122 / 0 / 0 |
| Typed precision / recall | 100% / 81.967% | 100% / 100% |
| Exact employment records | 13/16 | 16/16 |
| Exact education records | 4/6 | 6/6 |
| Provenance records | 24/27 | 27/27 |
| Ambiguity abstention | 6/6 | 6/6 |
The separate frozen ten-case subset per format has PDF typed recall 47/60 (78.33%) and DOCX 60/60. PDF unresolved-source detection produced TP16/FP27/FN2: precision 37.21%, recall 88.89%, indicating review burden. Zero typed false positives here is not a population hallucination guarantee. These results measure deterministic extraction, not model-assisted CV interpretation or the user's corrected mapping accuracy. Reviewed transactional application is covered separately by integration tests.
Supported qualification covers text PDF and DOCX, EN/Bokmål, multi-page, tables, columns, incomplete dates and bounded synthetic long histories. Scanned/image-only/OCR-heavy and unusual unsupported documents require explicit unreadable/review handling; OCR and general layout completeness are not qualified. Missing facts remain unknown. Near-100% factual extraction remains aspirational.
Independent decisions
| Criterion | Decision | Limit |
|---|---|---|
| Blank creation / undo-redo | PASS | automated scope; human interaction review open |
| Revised spacing | PARTIAL | unchanged S01 coordinate failures |
| Template library / matching letters / photo | PASS | automation; human design/print/device review open |
| Dashboard / admin monitor / event viewer | PASS | isolated development automation |
| Secure upload / Docling extraction | PASS | supported bounded fixtures; native parser fuzz qualification open |
| Structured mapping / review / progress-recovery | PASS | supported contracts; incomplete extraction remains visible |
| MariaDB persistence / RD21 revision safety | PASS | isolated provider and browser evidence |
| Deterministic extraction accuracy | PARTIAL | PDF recall and review burden above |
| Real CV AI compatibility / factual accuracy | BLOCKED | no immutable isolated CV-model quality qualification |
| Preservation | CONDITIONAL | all 57 original criteria retained: 56 PARTIAL, one DEFERRED; 33 original routes plus admin route |
| Security and privacy | CONDITIONAL | advisory, operator and parser limits below |
| PRE-P10-UX / CV-QA handoff | CONDITIONAL GO | explicit remaining acceptance debt |
| Phase 10 readiness | CONDITIONAL GO for candidate assessment/planning only | implementation remains unauthorised |
| Production | NO-GO | no deployment or blanket acceptance |
Security, privacy and owner actions
The sealed security diff reviewed 136 files through d52403d, with no reportable findings. The subsequent 3255da9 addendum was separately reviewed and its failure-timestamp finding fixed; it is outside the original sealed scan. See security evidence. npm production advisory count is zero; full npm has five low development-tool advisories. Python import lock audit has zero. NuGet identifies one distinct moderate MySqlConnector 2.4.0 advisory, repeated by its consumers: GHSA-473q-m89c-ghf8. Before release, qualify an upgrade to at least 2.6.1 or the documented trusted-CA configuration and rerun provider/RD21 checks. Production TLS configuration was not inspected. See dependency audit.
Owner acceptance still requires real keyboard/screen-reader and contrast review, physical desktop/tablet/mobile testing, English/Bokmål language review, actual printing, revised S01 disposition, isolated real CV-model qualification, external provider credentials/consent where applicable, and production filesystem/off-disk backup/model-cache/operator evidence. A healthy endpoint or synthetic local response does not satisfy these gates.
Local runtime activation and synthetic worker outcomes are recorded separately in local runtime acceptance. The QA handoff identifies the next bounded work. Do not merge, deploy or begin Phase 10 from this conditional report.