Standard: Owner Mandates Become Ledgers

An owner's multi-part directive is a checklist, and the mesh's rule for checklists — checklists-are-contracts — binds it exactly as it binds a standard's ## Verify table. This standard is the owner-directive case, spelled out, because it is where the mesh most visibly bleeds: a request given in a first message that took three prompts to finish, each round feeling "pawned off", each round proving the last round was capable all along.

Canonical, project-agnostic standard (the version other repos copy). Distilled from papermc-despawned-items' 2026-07-21 mandate-execution failure analysis. The generic checklists-are-contracts machinery previously bound only hub standards' checklists — this closes the gap that a spoken owner mandate fell straight through.

The mechanism it prevents

A directive is lossy-compressed at intake and the loss is never audited. "Every command from the different permissions, used correctly and incorrectly, verified individually" — ~9 load-bearing words — becomes a task named "permission matrix tests", which keeps 3. Once the summary task is checked off, the un-transcribed words are unrecoverable because nothing re-reads the original message. A request that doesn't become an entry becomes optional.

Three compounding effects make it worse: deferrals carry plausible reasons but no evidence of an attempt; the "NOT done" disclosure list feels like compliance but carries no obligation forward, so a repeated mandate is re-derived from scratch instead of escalated; and "do as much as you can" quietly re-anchors "done" to a shippable milestone (a release!) rather than to exhaustion of the mandate.

The rule: transcribe verbatim, one row per clause

A multi-part owner directive is transcribed verbatim into notes/plans/<date>-mandate.md (template: ../templates/mandate-ledger.md) before execution — one row per clause:

• Status is one of done · blocked-with-evidence · awaiting-owner. A completion claim cites the rows, not a paraphrase. The phase-end check diffs the delivered work against the owner's original words, not against the plan file (the plan can already have dropped a clause).
• Deferral requires falsification (the checklists-are-contracts rule, restated here): blocked demands a recorded attempt (command run, error text, version checked) and a retest trigger — never a plausible reason alone.
• Disclosure is stateful. An awaiting-owner row survives into the next session as a tracked row; the first action under the same owner next session is to re-present the open awaiting-owner rows for a keep/descope decision — not to re-derive them from scratch. A repeated mandate escalates every awaiting-owner row to do-now, and a second repetition of the same item is itself a reportable process failure (it means the loop failed, not that the scope grew).
• No milestone-anchoring under an open mandate. A release may ship mid-mandate, but the completion claim must state "mandate rows remaining: N" — a green release never implies the mandate is done.

Convert lessons into rules at write time

Same rule as checklists-are-contracts, because this failure proved it twice in one day: when the run records a lesson ("probe the mock first"), the same commit adds the behavioural rule to CLAUDE.md or the relevant standard. A lesson left as prose does not transfer, even hours later on the same repo.

Verify (is it being followed?)

The per-standard slice the compliance audit aggregates — report done/partial/missing: