Construct behavioral knowledge from the supplied original public evidence and the validated state_layout. Published tool contracts are fixed. Return exactly claims, text_rules and python_source according to output_schema. Do not repeat or modify state_schema, tools or initialization_requirements. Historical values describe examples, never the current instance.

Claims use statement, status, source_refs, scope and origin. Status is observed, documented, inferred, contradicted or unknown. Observed claims cite actual episode source records; documented claims cite actual description/contracts. Inferred rules retain uncertainty outside supported contrasts. source_refs contain source_id and an exact relative JSON pointer from source_locators. A model-only hypothesis uses origin=model_prior without fabricated citations. Citation resolution does not prove entailment.

Ensure claims and text_rules agree. Describe guard order only when documented or distinguished by jointly invalid observations. Account for every documented state effect, including related entities; preserve unrelated state. Do not turn absent examples into universal refusals or assume conventional allocation, rounding or error literals. Explicitly retain unknown behavior.

Fictional format example, not target evidence: {"claims":[{"statement":"The recorded read returned null.","status":"observed","source_refs":[{"source_id":"episode:sample","pointer":"/0"}],"scope":"That recorded read.","origin":"evidence"}],"text_rules":"Use current instance values. The one observed null result does not establish that every read returns null.","python_source":null}. Use only actual supplied source IDs. Supply nonempty generated Python only when the requested representation requires it; it will not be executed here.
