Motir

Found by running MOTIR-4056 to completion. The correction is submitted as a plan against MOTIR-3942; this card is the telemetry.

Two defects, one card, and each of them looked like rigour.

1. The WHAT was declared out of scope on an argument that only covers the HOW

The card's boundary read "OUT: WHAT COUNTS AS ENOUGH. That is the planner's judgement, delegated by the decision-authority ladder", and its body argued the point at length: enumerating it would take back a decision the corpus has delegated, and a checklist becomes a ritual.

Both arguments are correct about the HOW and neither reaches the WHAT. DECISION_AUTHORITY_LADDER delegates architecture, slicing, sequencing, sizing and placement — and then says the opposite about the other axis in its own words:

"For WHAT there is no rung at all — each one answers 'how a thing like this is normally built', and NONE of them can tell you WHICH thing the user meant."

So the card applied the delegation argument to the one axis the ladder explicitly refuses to delegate, and deleted the specification for it. What a card must CONTAIN to be buildable is the planner's; what the REQUEST must pin down to be plannable is not, and it is as writable as any requirements standard.

The tell was in the card and read as a virtue. A boundary that argues for its own omission at length, citing a rule that supports it on a neighbouring axis, is the shape to check — an omission nobody defends gets noticed, and one defended in three paragraphs does not.

2. The gate was addressed to the phase it gates

The rule was "before you open a phase, you must have enough information to FINISH it", delivered to phase-author and phase-lay. Those packs are composed only into sessions that have already opened. The instruction reached every reader except the actor that decides to open one.

It went green: the rule was present in all 22 cells, per-cell asserted, and the guards were satisfied. Nothing in the suite tests who a sentence is addressed to — presence and delivery are both checkable, and audience is not.

The takeaway, as the rules it should have been planned by

1. A rule about whether to BEGIN something is placed with the actor that decides to begin it, never with the activity it gates. Naming the reader is part of authoring the rule.
2. The decision-authority ladder delegates the HOW. A card that cites it to omit something must first say which AXIS the omission is on — on the WHAT the ladder is an argument for specifying, not for delegating.

Both are placement rules of the same family as MOTIR-4128, filed against the same card from the same run.

Acceptance criteria

• motir-meta's rule corpus carries both rules where a planner authoring a rule-pack card reads them, stated as "in situation X, do Y" rather than as an account of this incident.
• The first names the check concretely: before sealing a rule, name the SESSION that reads it and confirm that session can perform the act the rule asks for.
• The second is written as a limb on the existing decision-authority guidance rather than a free-standing rule, because it is a misreading of that ladder and belongs beside it.
• REGISTER.count.py's cautionary count does not rise.
• The conservation baseline is refreshed only if rule text genuinely relocates, with the word-diff in the PR.

Scope BOUNDARY

• OUT: the CORRECTION to MOTIR-3942, which is submitted as a plan and is where the product change lives.
• OUT: motir-ai's mirror — MOTIR-3944 runs the mirror pass over the epic.
• OUT: the INVARIANT B placement rule — MOTIR-4128's, and the two land in the same file if both are approved.

Context refs

• motir-ai src/llm/planningRulePacks.ts — DECISION_AUTHORITY_LADDER's "for WHAT there is no rung at all", and CORPUS_ORDER's pack tags, which are what decide a rule's audience.
• motir-ai src/llm/authorSession.ts · src/jobs/planningTarget.ts — the sessions and the seam that make the addressee question concrete.
• motir-ai#351 — the shipped-then-drafted PR this was found in.
• MOTIR-4128 — the sibling planning bug from the same card.