Planner telemetry, filed into the planner-bug home. It blocks nothing and is in no sprint. No correction to the tree was needed — the card's own criterion 2 made both dispositions legal, so it completed correctly. What was wrong is what the card ASSERTED as established.
motir run MOTIR-3264 (2026-08-23), at the measurement step. Full evidence in the run comment on MOTIR-3264; the card's title has been amended on the record.
MOTIR-3264 asserted, in its title and throughout its root-cause section, that the gateway's 600 s sync interval "clears the real delay by about a minute" and is "the same near-miss that cost motir-ai 77% duty cycle". Both rest on a suspend delay of ~9 minutes.
That figure was measured — carefully, by the right method, with no database connection — on motir-ai's and taq's endpoints (MOTIR-3224, 2026-08-20). It was never measured on the gateway's. Measured there on 2026-08-23 over a traffic-free 44m52s window, the delay is ~5m12s across five consecutive cycles and the compute is asleep 46.6% of the time.
The card was right to demand a measurement and wrong to have already drawn the conclusion. Its "What is NOT established" section is exemplary — it names the confound, forbids opening with a fix, and prescribes the method. And then the title, the root cause and the explanation all state the near-miss as fact. A run that read only the title would have shipped a fix for a defect that is not there; the reason this one did not is that the card also said not to.
A measurement taken on a SIBLING subject is not a property of yours. Endpoints, machines, tenants, repositories, environments — a number read off one of them and carried to another arrives with the full authority of having been measured, and none of the warrant. It is strictly worse than a documented figure in one respect: nobody thinks to re-check it, because it already survived the "is this measured?" question once.
The tell is grammatical and greppable: a card whose evidence section names one subject and whose conclusion names another. Here, "Measured 2026-08-20 … ([MOTIR-3224]) : ~9 minutes" two paragraphs above "So a 600 s loop clears the real delay by about a minute" — where the loop is the gateway's and the ~9 min is not.
And the durable half is one level up again: this card's own thesis is "an interval is only ever as good as the threshold it clears, and a threshold from documentation is not a contract." The correction is that a threshold from MEASUREMENT is not a contract either — the same method on identically-configured projects gave ~9 min on 2026-08-20 and ~5 min on 2026-08-23. A measured number is a reading with a date, not a property.
plan-rules/ gains, or sharpens, a rule for a card that IMPORTS a measurement taken on a different subject: the measurement may be cited as the reason to look, never as the card's own established fact, and the card's conclusion sentences must name the subject the number was taken on. A card whose evidence and whose conclusion name different subjects owes a re-measurement as its first criterion — which MOTIR-3264 in fact had, and which its title then contradicted.notes.html #342's row ("A measurement of a themed or contextual property owes the context it was sampled in") is the same shape one domain over, but is worded around colour/contrast and constrained types: ['design','code'], so it cannot reach a chore / deploy / infrastructure card. Widening it to be domain-neutral is probably right and would avoid a second row saying the same thing to the ranker; adding a row is the alternative. Decide and do one. (Owed by this card, not by the run that filed it — the filing run stopped here deliberately rather than pre-empting a corpus-shape decision at the end of a long session.)motir-core/docs/decisions/application-hosting.md §21 — now carries both readings and the unresolved question of which is general (motir-core#2264).motir-ai src/seed/lessons.base.ts — the row for notes.html #342, the widening candidate under criterion 3.