C.33:6 - Bias-Annotation

Bias	How C.33 counters it
Carrier completeness bias	Require captured selected structure, lost or hidden structure, admissible use, and non-admissible use before relying on the carrier; add expected but uncaptured structure when the next use depends on it.
Metric bias	Treat entropy, epiplexity estimate, benchmark score, dashboard value, dependency F1, and invariant F1 as readings only when `C.16` or `C.32.ACE` has opened that claim.
Source-label ontology bias	Keep source labels such as layer, router, cache, expert, pruning, distillation, block, DSM cluster, and architecture-search result as source wording. When a label carries an unresolved FPF structure or relation claim, use `C.30.STRAT`, `C.30.TFS-REL`, `A.6.M`, `C.31`, `C.32`, or another direct subject pattern to recover the object and predicate for that claim.
Observer-belief bias	Record observation class, confidence, active-passive gap when present, budget boundary, and unexplored regions for agent-produced or probe-produced carriers. Do not infer internal belief, safe change, or assurance from a map.
Decision-memory bias	Treat ADR-like records as decision descriptions and method expectations. Use `C.32.PAD` or `C.32.ADR` for decision and projection claims, and use C.33 only for what structural content the record carries or loses.