A report is derived from the record, and carries what produced it

A report is rendered from findings and digests that were themselves derived from a trace, and it records the run id, configuration hash, seeds and engine commit it came from. It is never assembled beside the record from numbers a script gathered along the way.

This is not a style preference; it is a defect already observed. The first real study's report was a dictionary built by its own evaluation script, and it drifted from the trace immediately: the facts that mattered — that word of mouth fired 113 times, that every turn moved a belief, that turns split across two channels — were in the trace and never reached the report, while the report's own probe_disagreement: 0.0 overstated what 18 easy questions established. Nothing tied the two together, so nothing could notice.

Every finding resolves to trace ids at the moment it is authored, by asking the view to resolve them. A finding whose evidence does not exist therefore fails where it is written rather than where it is rendered, and report cannot introduce a claim because it only formats findings that already passed that check. The run's calibration is stated once, as its TrustStatement; findings carry only their own confidence, so "this audience was small" is never confused with "this engine has never been benchmarked".

Considered options

Letting each consumer gather what it needs from a run was rejected: that is the arrangement that produced the drift above, and the previous design's duplicated objection clustering before it. Recording provenance only in the registry was rejected because a report travels — it is the artefact that leaves the machine, and a reader holding it should not need the registry to know which run it describes. Rendering findings without resolving their evidence was rejected because an unresolvable citation discovered at render time is discovered too late to fix the finding.

Consequences

report stays thin and testable: same findings in, byte-identical documents out, markdown and JSON carrying the same set. A report names the engine commit that produced it, so two reports from different code cannot be mistaken for a difference in findings. And the evaluation script that started this ADR is replaced by the digest it should have quoted.