Filed by the motir run MOTIR-2833 of 2026-08-15, which could not execute the card. Recorded here per the planner-bug-home rule; the correction is already applied to the plan (MOTIR-2856, MOTIR-2857, and the amendment on 2833 itself), so this card blocks nothing.
MOTIR-2833 was written as a 3-point closing card: run one named suite, confirm it green, write eight verdicts, lower a ceiling to 0. It carried, in its own body, the two facts that disprove its plan — and read both of them as narrowing the answer toward the conclusion it wanted.
The quoted fact. The card says: "Against pg_policies, only project and work_item carry public arms — board, board_column, board_column_status, workflow_status, comment, public_request_vote and work_item_link do not. ... All eight reads target project or work_item, which is why they are plausibly public rather than plausibly broken."
What it actually establishes. The eight reads target project / work_item — and four of them JOIN a table on that unarmed list. findPublicRoadmapSubmitted, countPublicRoadmapSubmitted and findPublicRequestMatches each join workflow_status, which is named in the card's own sentence as having no arm. Under motir_app a join to an unarmed table returns zero rows and raises nothing, so the read is broken by precisely the fact the card quoted as reassurance. listPublicDirectoryRanked fails the same way through workspace → organization, two tables the card's list does not mention at all.
The second premise, independently false. The card names tests/publicProjects as the evidence that settles all eight. Three of the eight (listPublic, listPublicDirectoryRanked, maxActivityByProjects) are not exercised by that suite anywhere — they live in tests/projectSquare, which sits at 36 failures under the app role. One grep for each read's name over tests/ would have shown this at plan time; it takes about a minute.
The card's step 1 says: if the suite is not green, STOP — the failures belong to MOTIR-2796 / MOTIR-2797. Both merged before this run and both are done, so that disposal routes a real, latent PRODUCTION defect to two closed cards. Followed literally, the run stops and reports "blocked on 2796/2797", the reader sees two done cards, and the finding evaporates. What is actually behind it is that at MOTIR-2515's cutover the public roadmap, the duplicate-detection pre-check and the entire project square go blank for every reader.
A card that names the tables carrying a policy arm has described the FROM clause, not the query. A read is admitted only if EVERY table it touches is admitted — the join is part of the predicate. When a plan-time fact is a list of what is armed, the verification it demands is a list of what each read JOINS, checked against it; reading the fact as "my reads are on the good side of this list" is where the assumption enters.
And an acceptance criterion may only cite evidence that can bear it. "Cite the green suite run that settles it" is a strong, correct bar. It is worth exactly as much as the check that the named suite actually EXERCISES the thing — which is one grep per read, at plan time, and was not done.
Logged as notes.html mistake #269 (motir-meta).