Found by the motir run MOTIR-2865 parent-run, 2026-08-16. The correction is already applied (the story's PR reports the true partition and hands the misplaced rows to MOTIR-2871); this card is the telemetry.
MOTIR-2865's own words: "166 are writes refused by a policy's WITH CHECK where the write is issued by application or script code, not by a test fixture" — and then, one section later, an explicit hedge on a different axis:
⚠️ The scan is a corroboration, not the specification. It only sees
db.$transactionblocks … Each child re-derives its own call sites; nobody builds to this table alone.
So the card knew its instrument was weak about WHERE the write lives, said so, and told every child to re-derive it. It said nothing about WHO issued the write — which is the claim that decided the card's existence, its parent story, and its boundary ("no child's PR changes a test fixture's choice of client").
Measured on the branch, per child, by tracing every failing test to its statement:
| card | rows the card claimed | rows actually issued by application/script code |
|---|---|---|
| MOTIR-2866 | 49 | 49 |
| MOTIR-2867 | 39 | 17 (12 notification + all 10 watcher are fixtures) |
| MOTIR-2868 | 42 | 41 (1 is a fixture's context, not its client) |
| MOTIR-2869 | 18 | 8 (importSeam's 10 are the fixture helper at :107) |
| MOTIR-2870 | 18 | 0 — every site is a fixture, and all four named services were already bound |
| total | 166 | 115 |
Three of five children were mis-sized, and one had no product work in it at all.
An error message names a TABLE. It never names a caller. new row violates row-level security policy for table "watcher" is byte-identical whether lib/ or a test issued it, so no amount of care with that string could have produced the writer axis. The partition did not have a weak instrument for WHO — it had none, and wrote the sentence anyway.
That is the asymmetry worth keeping: the card hedged the axis it could see and asserted the axis it could not. The hedge reads as rigour and makes the unhedged sentence read as measured.
The cheap check existed and is one command per group — grep -n 'db\.\$transaction' <the failing test file>. For watcher it would have returned ten hits in three test files and zero application writers, before the card was written.
MOTIR-2870 was sized at 5 points / 85 minutes on the strength of a premise — "these four tables are refused with nothing in the scan to blame, so the writes reach the database by some uncharacterised route" — that was false. The route was the fixtures. A card whose whole reason to exist is a mystery should verify the mystery is real before it is estimated, because the estimate is priced on the mystery.
plan-rules/ — a partition cut from a machine-readable signal must state, per axis, whether the signal CARRIES that axis. An axis the signal does not carry is not "approximate", it is absent, and a card may not assert it. Where the axis decides scope, ownership or a boundary, the plan pass owes one verification command per group and its output.grep per failing file.notes.html gets the lesson (#257 is the ancestor — a classifier that cannot tell an application defect from test debt — and this is the sharper form: it could not tell them apart in principle, and the card said it had).