Audit the evidence
Use five questions from TRACE to inspect this study. This is an evidence map, not a TRACE certification or a new protocol. Records show what was retained; they do not prove an explanation sound.
Scope: one C2 model-building example in the indexed
publication, commit c070d990be882. These records come from
the parent run before the C4 Field branch. They do not establish
coverage of every stage or community. Saved-run links below require
repository access. Full
instructions and source
hashes are readable on this site; instructions describe what should
happen, not proof that it happened.
Evidence register: exact file versions, hashes, checks and limits. It records this inspection, not an acceptance decision.
T. Track records: what happened and why?
Recorded example: Model history associates model changes with a leg, reasons and source communities. The C2 close receipt identifies a completed turn, validator and report hashes. The report hash was checked against its saved file for this guide on 24 October 2026.
Limit: a close receipt is a record of completion and checks, not a record of every inference or a resumable agent by itself.
R. Review against sources: what was tested?
Recorded example: The C2 review retains nine objections and discloses the ten model sections omitted from the review packet. Reviewer metadata binds the reviewer to a packet hash; that hash and the returned model hash match their saved files.
Limit: the reviewer saw selected sections. Neither a favourable review nor valid file shapes establish the model's truth. C2's full instructions explain the review criterion.
A. Accept independently: who could decide?
Recorded example: QC2-001 identifies a handler ruling under J's delegation, including what the reviewer may see. The handler protocol separates aggregator, handler and PI duties.
Gap: this demonstrates a delegated scope decision, not independent acceptance of the entire model. The aggregator answers the critic's objections. We have not established a separate, version-bound acceptance decision for every output. Critique, close and scientific acceptance must not be read as synonyms.
C. Control transitions: what may advance?
Recorded example: QC2-002 states J's direction for the final C2 pass and how to treat the stopping criterion. The close report preserves a warning even though the mechanical check passed. Read the ruling beside the warning; do not turn either into a blanket approval.
Limit: a stage gate differs from the decision to publish a checkpoint. A readable record does not imply a downloadable bundle or a live interview. See available routes.
E. Evaluate corrections: did the repair work?
Recorded example: The C2c response accepts the reviewer's objection to the label “Present” and specifies a correction. Pass 4 records that replacement as “order fails in”. The earlier version is retained. This audit checked the revised wording, not the underlying counts.
Gap: the later close report does not prove an independent, targeted recheck of every correction. That wider claim remains unverified.
Use a checkpoint interview
Ask which evidence changed the account, what an objection exposed, and what remains uncertain. Check the reply against the saved record. New interview answers are private, later responses; they do not amend the original evidence or certify it. Interview guide.