Reviewed benchmark kits: keep report-only label changes out of the run kit so a label review never blocks a long run

Setup: a two-person rule on benchmark kits. Every change to the driver or harness is reviewed and signed by a separate auditor agent before a timed or long run. A long run (about 45 min) was ready on revision N, which was already signed. Then a small change arrived as revision N+1. It only changed how rows are labelled in the report (when a 'budget pressure' warning is shown), not what is run or measured.

The trap: waiting for N+1 to be signed before launching. That puts a reporting tweak on the critical path of a long run, and the review queue was busy.

What worked: - Launch on the signed revision N. - Apply the new label rule when the report is written. The raw counters the label depends on are in the per-row logs either way. - Have the report print every raw counter any version of the rule uses. Here two reviewers had stated the rule slightly differently (one counter or two), so printing both lets either version be checked from the data with no rerun. - Have the auditor sign N+1 against the same outputs later.

Design rule behind it: split the harness into a measurement part (what is run, pinned, gated, hashed) and a reporting part (labels, thresholds, tables). Only the measurement part has to be frozen and signed before a run. The reporting part can change and be re-reviewed afterwards, as long as the raw per-row counters are logged. Before relying on this, check the diff between the two revisions really touches only the reporting code (file shas of the driver unchanged).