← ClaudeAtlas

reconcile-native-material-use-evidencelisted

Reconcile offered, fetched, placed, discovered, loaded, used, and independently verified harness material from separate records.
alisonjieli-png/loop-engine · ★ 2 · AI & Automation · score 66
Install: claude install-skill alisonjieli-png/loop-engine
# Reconcile native material use evidence ## Use case Use when a run report says a skill or instruction helped a task and the actual native client path must be checked. Produce a state-by-state evidence record for the selected item. ## Required inputs - Item identity, exact approved digest, selected run and client version. - Available host offer and fetch records, placement manifest, native discovery or load observations, run trace, and independent task acceptance record. - The event definitions used by that client to distinguish discovery, load, and use. ## Procedure 1. Identify the same item and digest across every record. Do not join events by a friendly title when revisions or rendered variants differ. 2. Mark `offered`, `fetched`, `placed`, `discovered`, `loaded`, `used`, and `verified` separately as observed, missing, disputed, or inapplicable. Name the producer and time of each observation. 3. Check ordering and scope: a later task cannot inherit proof from an earlier process, and a selected file on disk is not a model load event. 4. Examine traces for direct evidence that the client consumed the material. A discovery listing, a token count, or a producer's summary alone does not prove use. If client instrumentation cannot observe use, keep it unknown. 5. Compare the independent acceptance rule and artifact with the claimed result. A successful model call or passing local format check is not task verification. 6. Return a compact evidence chain, missing links, con