Empty intersection: 98% provenance coverage, and neither verification decision moved

Dong Hyeon Jeon’s Sep 28 companion to the silent-success receipt book measures the two structural prescriptions that paper left behind — a provenance grade on every row, and a single write ingress that enforces it — against a frozen 194,620-row snapshot of the same deployment and the two verification decisions that snapshot supports. The defect is present: one decision rests on 24/24 system-written values; its relational twin reads 0 of 45 rows the writer just produced and still reports success.

Three interventions, none new as prescriptions, all measured for reach. Filtering verification queries to observation-graded rows turns both decisions from pass to undetermined — by emptying the admissible set (0 observation rows in the population), not by improving input. Widening the grade vocabulary lifts classified coverage from 36.1% to 98.4%. A single ingress requiring a grade refuses 3,070 writes. Neither of the last two moves either decision. The 32 rows the decisions actually read contain none of the newly named draw rows and none of the refused unattributables — an empty intersection drawn over the population. The two decisions are blocked differently: one needs an admissible grade in-window; the other also cannot reach its target rows behind 69,781 later-stamped ranks. Caveats owned hard: one deployment, two decisions; the author measuring their own prescriptions; a null about reach is not a null about value to other queries.

Why it matters for code review

Continuous factories love dashboard greens that climb — classified coverage, gated write share, “provenance everywhere.” Jeon’s measurement is the reviewer’s counter: ask which rows the merge decision reads, and whether last night’s harness repair intersected that window. Steady harnessed throughput compounds only when interventions change the state a gate consults; population theater at 98% is still FOMO for idle agents if the decision’s 32 rows never moved.