[SignalLayerLabs]

Problem and behavior

PR14 deployed successfully, but its first operational check found four Forex lanes with 30 attempted blocks each and zero learned outcomes, and no USA lanes. Yahoo's separate live quote row was persisted as another candle at each distinct timestamp. Discovery also depended on incomplete coverage metadata and rejected whole histories for a single invalid bar.

Quarantine off-grid and unclosed Yahoo observations, finalize legacy partials only through a closed refetch (including daily PriceHistory captures), and rebuild frozen curriculum sources under an explicit clock policy. USA daily lanes read stored PriceHistory directly. Preserve invalid canonical bars and missing intervals as blockers, retain old source/checkpoint/journal evidence, and defer replacement while a prospective decision is pending. Source replacement uses revision compare-and-swap. The stored Research report now distinguishes productive learning from attempted blocks and shows rejection reasons.

Validation and limits

• 1,012 backend/migration tests passed; 41 frontend tests and the production static build passed. Historical-source and HF metadata audits passed.
• Isolated production-worker replay: 60 blocks across USA and Forex, 1,109 dependent expert outcomes, 45 productive blocks; a new database session added zero duplicate updates.
• Development returns remain weak (roughly −0.39% EURUSD, −0.97% GBPUSD, +0.34% QQQ, −0.34% SPY; 1–7 selected-policy closures). No alpha claim, final confirmation, qualified promotion or real-money routing.
• Production still runs PR14 until this correction is explicitly approved and deployed. Post-deployment learning and source replacement require fresh verification.

Reproduction, source hashes, preserved-data contract and release acceptance: docs/research/curriculum-source-recovery.md, docs/evidence/curriculum-source-recovery.json.