Every 5 resolved cases, the system clusters past resolutions and asks Gemini to distill the recurring PATTERN into a reusable methodology. Methodologies become first-class KB entries every agent retrieves.
Claims where no KB entry was confident enough. Reviewers should backfill the KB.
👍/👎 from real customers. Cited KB entries get quality boost / penalty automatically.
Every auto-resolved claim becomes a labeled training pair. Gold cases (supervisor approved + customer accepted) export to a one-click fine-tuning dataset for Vertex AI Gemini, OpenAI, or Anthropic. After ~2000 gold pairs, the merchant fine-tunes a smaller, faster ClaimsForge-shaped model — typically ~30-50% token + latency reduction.
Business-editable supervisor rules from data/hard_rules.json.
Toggle a rule off here to disable it without a code deploy — cache picks up the change within 60s.
Tier-1 rules (pHash collision, duplicate-order, multimodal mismatch, frequency cap) stay in
code and cannot be toggled — they're the constitution.
When the supervisor or verifier escalates, the structured briefing lands here so agents can act without re-reading the transcript. P0 = legal/regulator (act in <5 min) · P1 = high emotion (<30 min) · P2 = normal (same day).