The grade we failed, with its diagnosis.
The bar, published first: before trusting our mistake-classification rule, we required two independent graders to agree above a chance-corrected 0.70 over the same population: the statistic named per dataset, with raw agreement and class prevalence reported beside it, and no metric chosen after labels exist.
The measurement, dated 2026-07-30: 88 mistakes, two independent graders, neither seeing the other's output before both were fixed. Cohen's kappa 0.410 against the 0.70 floor. Raw agreement 0.648. Fail, and held: the bar did not move, the graders were not tuned toward each other, and nothing built on those labels ships.
The diagnosis: in 44% of cases the mistake record did not capture enough for either grader to work from. Most disagreement was one grader finding what the record barely held, not two graders splitting on a judgment call. A perfect rule would still fail on a record that thin. So the fix is the record, not the grade.
What re-opens enforcement: re-measure on the improved record and report three numbers separately: the chance-corrected statistic, raw agreement, and the split between record-coverage disagreement and genuine boundary disagreement. The floor is then set from the observed distribution, not imported from another task's precedent, which is the mistake that set 0.70 the first time.
The re-measure, dated 2026-08-15: two independent graders, working blind and separately, labeled the same 26 mistakes: agreement 0.944, raw 0.962. The first pass of the 0.70 floor this project has recorded. Both numbers stand, and the movement between them is the finding: 0.410 with one grader on the full 88 (2026-07-30), 0.944 with two graders on 26 (2026-08-15). Same rule, two records: the earlier one too thin to grade from, the later one written by people narrating their own mistakes. That is the diagnosis confirmed from the opposite side, not a grade that improved.
What the newer number does not say, stated because a number without its selection is a claim without a denominator: the 26 were chosen for having detailed records, so the reading describes a well-recorded subset and not the corpus. Enforcement still waits on a re-measure of the whole record at that thickness. Any figure quoted from a subset carries its selection in the same sentence: the earlier 0.720 reading, for instance, is on 25 of 88 rows, selected on resolvability, and is never quoted bare.