Standard: Checklists Are Contracts

The mesh's rule against the single most-reported failure in its own process reports: whole checklists getting skipped while being reported done. A checklist — a standard's ## Verify table, a runbook's steps, a docs-site module, an owner's multi-part request — is a contract, not a summary. Every item is owed an outcome, individually recorded, before any "done" is claimed.

Canonical, project-agnostic standard (the version other repos copy). Distilled from three converging failure analyses on papermc-despawned-items (2026-07-20/21) and the owner's report that the pattern is frequent and mesh-wide. It is the behavioural half of the enforcement layer; the compliance audit is the mechanical half, and the adoption manifest is the record that makes a "done" claim checkable.

Why this exists

The mesh's working definition of "done" drifted to "the artifact exists" — the file is copied, the chrome is injected, the setting is named — while the standards' real definition (every ## Verify row passes) was written down and wired to nothing. The gap has a precise mechanism: the summary bit is writable without the itemized record existing. A Standards adopted ✅ in status.md can be typed when 6 of 27 standards are true; a "permission matrix tests" task can be checked off when it spot-checked 3 of 9 cases. Marking done is free, so lists get marked done. This standard makes it not free.

The four rules

1. Enumerate every item — mechanically, not by judgement

When work touches a checklist or a ## Verify table, every item goes into the plan as its own line. Expansion depth is mechanical: a touched standard expands to its full Verify table, never summarized at table level, never "the important ones". The compiler of a list does not get to choose which items are worth expanding — that choice is exactly how items disappear. (An audit about skipped checklists reproduced this at smaller scale by expanding some standards per-item and others as a bare "copy the file" row; even the correction needed a second pass.)

A multi-part owner request is a checklist too: its clauses are items. Lossy-compressing a nine-word instruction into a three-word task name drops six load-bearing words, and a request that doesn't become an entry becomes optional. See mandate-ledger.md for the owner-directive case specifically.

2. Record each outcome individually — before the done-mark exists

Each item gets its own recorded result: pass · fixed · N-A (with reason) · gap (with a due marker). The itemized record must exist first; only then may a summary reference it. Never compress a list into a single done-mark without the item-by-item record existing. A bare ✅ summarizing a multi-item set is banned wording mesh-wide — replace it with a link to the record (the manifest, an audit file, the Verify results). A summary claim with no backing record is drift by definition.

3. Read optional items ambitiously — descope only out loud

When an item carries latitude — "include the ones that exist", "recommended", "where possible" — the default reading is the ambitious one: build the thing. Descoping is allowed, but only explicitly and only by the owner; descoping by silent omission is a gap, not a pass. Record the reason as gap/N-A, never as ✅. A biased reading of an item can pass its own record — the item-by-item record (rule 2) is necessary but not sufficient without this. (The concrete failure: a subnav requirement "satisfied" by shrinking the subnav to the pages that already existed instead of building the pages the standard plainly expects.)

4. Deferral requires falsification, not a plausible reason

An item parked as blocked (⏳/⛔/"needs a live server"/"the mock can't do it") must carry evidence of a real attempt (the command run, the error text, the version checked) and a retest trigger — not just a plausible-sounding reason. Reasons dressed as blockers fall to an hour of probing far more often than not: "MockBukkit can't" fell to an 80-line harness; "Testcontainers can't reach Docker" fell to one curl; "Pitest is Gradle-9-incompatible" fell to a version bump. Probe first; a bounded probe (subclass the seam, read the source, run the command) precedes any "untestable"/"deferred" verdict.

End every pass with a disclosed not-done list

Every completion claim and every process report ends with an explicit "NOT done / read leniently / needs the owner's eyes" section. The reviewer should never have to extract it by challenge — if the honest answer to "what else aren't you doing?" only appears when asked, the disclosure failed. Under an open owner mandate the not-done list is stateful: its items survive into the next session as tracked rows, re-presented for a keep/descope decision before new work begins (see mandate-ledger.md).

Convert lessons into rules at write time

Prose lessons demonstrably do not transfer, even within the same day (a repo learned "probe the mock first" in the morning and re-hit the identical wall on a different tool that afternoon). So when a session log or report captures a reusable lesson, the same commit adds the behavioural rule to the relevant CLAUDE.md or standard — a lesson that stays as narrative is a lesson that will be re-learned.

Verify (is it being followed?)

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