Feedback Loops (L8)

Evaluation signals mapped to actions across layers. In 0.1 only two automatic loops are enabled; others produce recommendations.

Signals → Actions by Layer (0.1)

L1 Noosphere (retrieval)
- Signals: recall@k, mrr@10, groundedness/faithfulness, source_hygiene
- Auto actions:
  - rerank_tweak: adjust semantic/BM25 weights; adjust k within safe bounds
  - source_blacklist(id, ttl): blacklist low‑quality/poisoned sources (TTL)
- Recommendations: reindex segments, refresh embeddings, normalize duplicates
L3 Workshop (tools)
- Signals: tool_success_rate, error_rate, p95_latency, error taxonomy
- Auto actions:
  - tool_blacklist(tool_id, ttl)
  - retry_policy_tune (limits only)
- Recommendations: fallback chain updates, parameter tuning, deprecation candidates
L5 Memory
- Signals: contradiction/contamination, suspect entries, consolidation score
- Recommendations: mark_suspect(ids), reconsolidate, tighten retention policy
Orchestration (ACS/CEO/HCS)
- Signals: e2e p95, strategy failure patterns
- Recommendations: strategy switch, stop criteria updates, routing changes

evaluation_observation: raw facts (latency_ms, tokens, tool outcome, retrieved_ids)
evaluation_result: metric/value/threshold/verdict + links to observations
feedback_action: { type, scope, ttl, reason, request_id, ts }