Planning bug: MOTIR-3626 recorded the over-gate advisory as "REAL" and declined it IN WRITING — an advisory ACKNOWLEDGED is not an advisory DISCHARGED, and the acknowledgement makes the card read as careful

Filed by motir run MOTIR-3626 (2026-08-26), which halted at guards #2 / #4 / #5 and submitted plan cmtapmw6u0098i2phft0lq11z instead of building.

What the card did

MOTIR-3626 was sealed todo / highest / ready: true / targetRepo: null, with five acceptance criteria landing in motir-ai, motir-gateway and motir-core plus two actions on a running service. Under its own ## Advisory dispositions heading it carried, verbatim:

likely-over-gate-sizing (5 pts / 90 min, over the 70-minute proxy) — REAL, and the remedy is a SPLIT this card deliberately does not perform. … It is a defect RECORD, not a runnable leaf: expand it before dispatching it.

So the detector fired, the author READ it, AGREED with it, named the correct remedy, declined to perform it, and sealed. A later motir run MOTIR-3626 claimed the card in one transition_status.

The defect, stated as the thing that is new

This is the sixth card in a known family — MOTIR-3108 ("expect this to SPLIT" in its own explanation, itself called the FOURTH), MOTIR-3168, MOTIR-3511, MOTIR-3566, and MOTIR-2121 before them. What is new is not the prose discharge — it is that the prose discharge is a DISPOSITION OF AN ADVISORY.

Every prior instance wrote its own remedy into a body where nothing read it. This one answered a machine. The corpus asks a runner to dispose of every advisory, and "REAL, and I am not acting on it" is a disposition in form and an admission in substance — and it reads as MORE careful than silence. A card that never mentions the advisory looks careless to a reviewer; a card that quotes it, agrees with it, and explains why the split is a later pass looks like a card someone thought about. The defect ships with a paper trail proving somebody saw it.

So the disposition vocabulary is one word short. An advisory has exactly two honest outcomes — ACT on it, or record why it does NOT apply. "It applies, and this pass is not the one that acts" is neither, and it is only expressible because there is no third state for the plan to hold. A run that finds it must treat it as an undischarged advisory, whatever the body says about it.

The second, compounding half — the one detector that could have caught it independently was structurally silent, for a REASON ALREADY RECORDED

validate_work_item on MOTIR-3626 returned likely-over-gate-sizing and a subsumption entry, and no likely-repo-straddle — on a card carrying no repository whose criteria name paths in two of them. Verified by reading lib/workItems/proseVsGraph.ts on origin/main:

• firstRepoStraddleCriterion's unpinnable arm requires criterionRepoPaths(...) to return two or more distinct repos.
• resolvePathRepo resolves a path only when its FIRST segment is the repo name (motir-ai/src/x.ts) or owner/name. "A token that resolves to nothing is BODY TEXT and yields null".
• MOTIR-3626's criteria write src/llm/gatewayClient.ts, model/ability.go, middleware/distributor.go — repo-RELATIVE paths, with the repository carried in a prose parenthetical (**(motir-ai)**). First segments src / model / middleware resolve to nothing, so paths.length === 0 and the arm cannot fire.

The detector is behaving exactly as documented. This is the same silence MOTIR-3168 recorded ("likely-repo-straddle was silent because the paths were bare filenames") — that card is done, and this form recurred anyway. Whether its remedy was scoped to close this form is for whoever fixes this to read; this card asserts the observation, not a verdict on that fix.

The two halves compound in one direction: the advisory that fires is the one the author can talk their way past, and the advisory that cannot be talked past is the one that never fires.

What a fix would have to change (NOT decided here)

• The disposition rule. An advisory recorded as REAL and not acted on is UNDISCHARGED — for the authoring pass (do not seal) and for the run (treat the card as un-runnable, whatever its body says). Its home is plan-rules/core.md's gate list and run.md guard #4's advisory block.
• ⚠️ Gate 19 — THE TWO HOMES applies to that change. A rule taught only to plan-rules/ teaches the manual planner; the shipped planner reads SHARED_PLANNING_RULES in motir-ai. This family's own precedent is MOTIR-2121 → MOTIR-2122, which is exactly that pair. The fix is two cards, and the sibling is filed at authoring time, not by the run that discovers the gap.
• Optionally the detector, if the repo-relative-path form is judged in scope — but read MOTIR-3168's outcome first rather than re-deriving it.

Repo

Unpinned deliberately. This is the defect RECORD; its fix lands in motir-meta and motir-ai and possibly motir-core, and choosing among them is the decision the fix card makes. That unpinned state is the very thing the card above was criticised for, so it is said out loud here rather than left to be inferred: this card is not dispatchable as written, and expanding it is a planning pass. The difference from MOTIR-3626 is that nothing about this card is highest, ready and holding a legal obligation.

Context refs

• The card: MOTIR-3626 and this run's comment on it (the guards, the shape, the four origin/main findings).
• The plan submitted instead of building: cmtapmw6u0098i2phft0lq11z.
• The detector: motir-core lib/workItems/proseVsGraph.ts — resolvePathRepo, criterionRepoPaths, firstRepoStraddleCriterion and its module note on the forms it deliberately does not cover.
• The family: MOTIR-3108, MOTIR-3168, MOTIR-3511, MOTIR-3566.