Evidence records¶

Maintained documentation describes current behavior and reproducible workflows. Investigation logs belong with their artifacts under analysis/, identified by artifact SHA256 and implementation commit. Completed plans are retired after their lasting contracts and evidence have been incorporated into the reference pages. Git history retains obsolete plans and unsupported reports.

Record a comparison¶

For each investigation, keep:

• Artifact paths and SHA256 values for the .crd replays and any .cdt traces.
• Binary provenance, producer version, plus the candidate implementation commit.
• Exact health, recording and comparison commands, with output paths.
• First mismatch for each channel, affected entity/field, and float bit/ULP details where relevant. Record caller-attribution diagnostics separately.
• Confirmed cause, landed fix, validation span, and any unresolved next probe.

Continue the same record when the artifact SHA is unchanged; distinguish each candidate commit and comparison result. A new artifact gets a new record. Do not assume two playthroughs have the same absolute tick timeline.

Trace comparisons¶

CDT traces come from the port alone, so a trace diff localizes a regression between two port revisions or two playthroughs of the same replay. Use the CDT contract as the version authority. Regenerate obsolete recordings; do not add migrations or salvage incomplete runs.

Run dbg health on both CDTs before interpreting a diff. Both selected windows must be ready for comparison. Run the full-channel dbg diff first, then dbg bisect or dbg focus for localization. A caller-label-only difference is an attribution diagnostic when RNG values and state transitions agree.

Evidence kinds¶

Passing a replay fixture, agreeing with the original code under the native execution oracle, passing the recovered core gate, matching recovered code, and visually playtesting a run establish different things; state which evidence supports each claim.