Skip to content

current-handoff.md

Latest commit

 

History

History
85 lines (69 loc) · 5.22 KB

current-handoff.md

File metadata and controls

85 lines (69 loc) · 5.22 KB
Raw

Current Handoff - V.I.O.L.E.T.

PR #136 is merged in origin/main at f6cae3483f4cf75974746a4cc82222f28e399b96. SCV2-ML2 has completed its isolated SourceConcept-only execution; final PR adjudication remains with the project lead.

Canonical State

Item Value
Merged baseline PR #136 / f6cae3483f4cf75974746a4cc82222f28e399b96
Current phase SCV2-ML2: Multilingual Identity Candidate Closure
Accepted R2R DB blombooru_scv2_r2r_dryrun_test_20260710 (immutable)
Accepted SCV2-ML1 DB blombooru_scv2_ml1_acquisition_test_20260712 (immutable)
Fresh review-fix ML2 DB blombooru_scv2_ml2_identity_closure_reviewfix_test_20260715
Superseded ML2 DB blombooru_scv2_ml2_identity_closure_test_20260714 (preserved, immutable)
SourceConcept / needs_review 1677 / 0

Accepted predecessor chain: PR #133 / A1R, SCV2-R2 (target_met_constraint_aware_r2), PR #135 / R2R, and SCV2-ML1: Multilingual Alias and Source-Metadata Closure. Provider metadata acquisition ended in ML1; production, Entity bridge, and downstream routes remain unauthorized. See R2R evidence reuse policy.

R2R's 3,319 dispositions and 12,249 signals remain immutable. ML2 made no provider, Pixiv, gallery-dl, LLM, production, Entity, or media_tags truth call or write. Search-result union is not identity union, and additional query terms use media-level AND intersection.

The superseded first ML2 execution reported 1214 candidate pairs. That count is preserved only as historical evidence and is not current execution truth; the fresh review-fix execution rebuilt the ledger from the new database and produced the actual 1213 pairs reported below.

ML2 Final Evidence

ML2 rebuilt creator identity candidates from the accepted ML1 snapshot using (provider, stable_creator_id, creator_role) as the deterministic source-layer anchor. It retained every non-empty canonical/account/historical name as an alias signal, used star-topology candidate generation, and never expanded into all-pairs alias comparisons.

  • identity families: 606 = 12 already materialized + 594 new + 0 cannot-link + 0 deferred + 0 fragmented deferred;
  • candidate pairs: 1213 = 1213 must_link + 0 cannot_link + 0 deferred;
  • accepted R2R accounting: 3319 = 1522 must_link + 1791 cannot_link + 6 deferred_nonblocking, reconstructed cache-only from the immutable database and compatible private evidence with zero reuse conflict;
  • active-concept reuse: 0 inactive references/reuses, 310 partial historical references retained only as private diagnostics, and 0 pre-existing active-component fragmentations;
  • concept-media membership: 2065 / 2065 exact trusted support rows and distinct expected media, with zero duplicate, missing, unsupported, or media_count mismatch;
  • SourceConcept-only runtime: 606 families / 1213 alias cases at 1.0 expected-media coverage, zero fallback, zero search-inert concepts, and 25 / 25 media-detail visibility samples passed;
  • candidate-generation gaps: 30 before / 0 unexplained after;
  • ML2 touched graph: 606 complete active components, largest 8; all 12 accepted historical components were audited with their old links, with no multi-stable-ID, direct/transitive cannot-link, contamination, cross-role, unknown-role, or duplicate-active-identity violation;
  • creator-context benchmark: 93 / 93 evidence-supported cases succeeded; one evidence-absent case is deferred_nonblocking_evidence_absent;
  • search-only families: 3642 before / 3642 after, with zero regression, unsupported, rejected-only, superseded-only, AND leakage, or search mutation;
  • second execution: zero mutations, no duplicate media support, and identical component/disposition fingerprints;
  • public redaction and review-pack integrity passed fail-closed publication gates.

Final contract: target_met_multilingual_identity_candidate_closure, target_met=true, safe_to_merge=true, route_approved=false, active_blockers=[]. The accepted ML1/R2R databases remained immutable; only the fresh review-fix ML2 dev/test clone received allowlisted SourceConcept/source-name observation writes. The first ML2 database remains preserved as superseded evidence.

Next Boundary

The evidence recommends project-lead review before any separately governed Controlled Scale Validation over roughly 10k-15k media. This PR does not start that phase and does not authorize a provider run, production/full-library run, PX1-B, Provider-2/PX-REC1, Entity bridge, truth promotion, media import, AI tagging, classification, localization, or source/iCloud mutation.

The six ML1 hardening milestones remain unchanged: PRE-NEXT-PROVIDER-EXECUTION-HARDENING, CONTROLLED-SCALE-AUDIT-DEBT, and PRE-NONWAIVED-PROVIDER-CREDENTIAL-HARDENING. None is exercised by ML2 because ML2 performs no provider execution.

Durable Links

Canonical roadmap path: docs/roadmap/current-mainline-roadmap.md.