Motir

Found while running MOTIR-4056, on review by Yue, after the pull request was already open (motir-ai#366).

What the card specified

The card drafted its core rule text with this sentence, and it shipped verbatim into the shared tier:

"The requirement is above you in this message … and it states what the user asked for; your target's own body is what THIS level narrows it to. Neither of them is yours to rewrite."

Its stated premise for the split was:

"Phrased once per phase, because only the completion differs — a whole CARD for author, a complete LEVEL for lay."

Why it is wrong

The target's body plays OPPOSITE roles on the two phases, and the card assumed one role for both.

So the sentence is correct on lay and forbids the deliverable on author: a rule in the tier both phases compose, telling author not to rewrite the very body it exists to write.

Why it matters more than a wording slip

It is the SAME defect this card was re-scoped to fix, recurring inside the re-scoped draft. MOTIR-4131 records the first occurrence: motir-ai#351 addressed a start-gate to the phase being gated. This is that mistake one clause down — a sentence delivered to two readers that is only true for one — written by a card whose whole subject is who a rule is addressed to, and which carries an explanation section about exactly that failure.

That is the finding. The lesson A rule stated where its actor never reads it is not a rule covers a rule in a file nobody loads; this is the sibling case it does not reach — a rule in a file everybody loads, that is only true for some of them. Placement was correct and the ADDRESSEE was still wrong.

Why nothing caught it

• INVARIANT B — disjoint cover asks whether a line is in two packs. A line in core is in ONE pack and reaches every cell, which is the shape it is designed to permit.
• The routing baseline asks WHO receives a segment, and the answer — every cell — was correct and intended.
• The conservation classifier asks whether surviving text changed. It was a pure addition.
• The card's own suite asserted the rule reaches every cell and nothing about whether it is true there.

No guard in this corpus asks whether a core segment is TRUE on both surfaces. The act test in planningPhaseAxisGate.test.ts asks the adjacent question — can each surface DISCHARGE the rule, by tool name — and a claim about which body you may rewrite names no tool, so it passes.

Acceptance criteria

• A guard exists that fires when a core segment makes a claim that is phase-specific. The mechanism is the deliverable of this card and is not pre-decided here; the act test's "can this surface discharge it" framing is the nearest shipped precedent, and the honest starting point is that this instance names no tool and so escapes it.
• The guard is proved against THIS instance: the withdrawn sentence, restored to core, goes RED — asserted against the real corpus, not a toy string.
• If no mechanical guard is defensible at acceptable cost, that verdict is written down WITH its reasoning and the card closes on an authoring rule instead — a core segment states in its own doc comment which surfaces it is true on, asserted by presence. A verdict of "not mechanisable" is an acceptable outcome; leaving it undecided is not.
• Whatever ships does not go red on the ~12 core segments already in the corpus, and that is measured rather than assumed.

Context refs

• motir-ai src/llm/planningRulePacks.ts — PLAN_AGAINST_THE_REQUIREMENT and its two per-phase segments, as corrected
• motir-ai tests/part2PlansAgainstTheRequirement.test.ts §2 — the regression, pinned in both directions plus a negative on core
• motir-ai tests/planningPhaseAxisGate.test.ts §1 — the act test, the nearest existing guard, and why it does not reach this
• motir-ai src/llm/authorSession.ts · src/llm/treeGeneration.ts — the two sessions the roles are read off
• MOTIR-4131 — the first occurrence of this family, on the same card