The Sweep Playbook — the standing guardian

0. The two truths

1. 07The gates are the floor; the sweep is the ceiling. The binary gates — format check, tests, lints, the conform gate, the specmap check — say pass/fail and MUST be green before any sweep work. The collector's advisory facts (coverage gaps, danger bands, backlogs) sit above that floor and say what to harden next. A green sweep on a red floor is a contradiction: fix the floor.
2. The gate is truth; the collector is a guide. When the collector says a unit is ready to gate, the gate confirms or refutes it. Trust the gate.

08Law 2 governs the sweep's shape: explanation capital must be runnable capital, so the sweep is collector-first — a deterministic, no-LLM fact-gatherer (the stack's health tool) computes the objective state, and the operator (human or agent) acts on facts, not vibes.

09Where a check cannot be mechanised yet (the WISH rules), the sweep names it and routes it to judgment — it never pretends prose is a gate.

1. Tier structure

10Tier 0 — the hard floor (every run, binary, first). Never sweep on a red tree.

11Run the stack's floor command (format → build/tests → lints → conform gate → traceability check → test-gate where a baseline registry exists → fast-loop budgets where enforced).

12If any gate is red, the only work is making it green.

13Tier 1 — the ratchet (every run; act on the collector's facts). Run the health collector and work its output in cheapest-win-first order.

14The canonical moves, each driven by a collector field, never by memory:

• 15Split the danger band before it trips — files approaching the length budget are landmines; measure with the rule (physical lines), not the eye.
• Widen a gate for free — a unit the collector shows at zero drain (e.g. full doctest coverage) enters the corresponding gate list with zero work; the gate run confirms the prediction.
• Drain the smallest backlog — take the smallest-gap ungated unit, bring it to zero, then flip it into the gate (the expand-as-you-conform rhythm: a unit enters a gate only after it drains to zero — a flip must never widen a baseline).
• Re-justify the deviation debt — walk each recorded deviation: does its reason still hold? A deviation whose invariant has since been encoded in a type/checker is removed and restructured. A deviation with no live reason is a defect.
• Catch census regressions — for gated units the violation censuses must stay zero; a non-zero is a fresh violation that slipped a local check. Restructure beats testify: encode the invariant in a type or contract rather than recording an excuse.

16Tier 2 — drift (weekly; the slow rot the gates don't see).

• 17Debt registry: run the tripwire tool; re-disposition any touched-and-open entry (BROWNFIELD §3). New deficiencies found while sweeping are filed there, not left as prose.
• Doc/code drift: state checkpoints vs reality (see WAL convention — freshness rule); architecture docs vs the real layout; roadmap self-staleness. A stale doc that governs code is filed as stale-doc debt.
• Marker census: every TODO/FIXME/REVIEW/HACK is a WISH or a deferred decision — load-bearing ones graduate into the debt/intent registries, trivial ones are resolved or deleted. Prose that promises and doesn't deliver is false training signal.
• Golden transcripts: characterization oracles must fail loudly and be re-captured deliberately, never auto-updated — an auto-updated golden is a test that always passes.
• Traceability kind hygiene: a code edge into an unmarked spec unit prompts marking the unit's kind/revision in the same change — a unit that code points at is not informative.

18Tier 3 — deep judgment (weekly/biweekly; the WISH rules). The rules no checker mechanises yet — reviewed by a human or strong agent so they are not silently skipped, each a candidate to graduate into a real checker.

19The stack's GUIDE and cards carry the concrete list (typed seams, cell isolation and oracles, uniformity, contract-first ordering, lying prose, closed-vocabulary naming).

20Tier 3 may also launch a full raid when a Tier-1 backlog grows campaign-sized — plan it per the Campaign Form.

2. The collector contract

21The health collector is the sweep's instrument.

22It MUST be:

• 23deterministic (same tree → byte-identical output, so its committed snapshot's diff IS the health trend),
• advisory (it never fails the build — the gates do),
• no-LLM,
• and single-sourced (it reads the gate policies — the conform and specmap configs — never hardcoded lists, so its numbers cannot drift from what the gates enforce).

24Count the list, not the record: any inventory the sweep acts on comes from the collector or a config, never from memory or a document's stale snapshot.

25The promotion ladder — how the rule set grows: a Tier-3 WISH rule becomes mechanisable → add it to the collector as a census (cheap, advisory) → once proven, promote it to a blocking gate rule.

26Extending the collector is itself ratchet work.

3. Cadence

28The daily run is light: floor green, then one or two cheapest Tier-1 wins.

29The weekly run adds the drift and judgment tiers.

30Any single item is a safe stop — the sweep is incremental by construction, and the collector re-derives the remaining work each run.

4. Output of a sweep

31Each sweep that changes the tree lands topic-grouped commits (one logical unit each), citing the sweep item that drove them.

32Refresh the committed health snapshot in the same run — its diff records the trend.

33Resume pointer (see 06-WAL-CONVENTION):

• 34With a WAL (recommended): bump the WAL's standing line at any sweep that moves a milestone (a gate widened, a backlog unit cleared); the health snapshot plus the WAL is the next sweep's starting point.
• Without a WAL: the sweep's closing commit message carries the summary — floor state, the Tier-1 items taken, the next candidate — and the committed health snapshot is the resume pointer. Nothing may live only in the operator's head or the session transcript.

5. What the sweep deliberately does NOT do

• 35It does not replace the gates (Tier 0 is the floor, the sweep hardens what sits above it).
• It does not auto-fix. The collector gathers facts; the operator acts. Mechanising a fix is a separate, deliberate promotion (a new gate rule or a codemod).
• It does not touch owner-frozen surfaces or owner-court decisions without sanction; drift found in them is filed as debt, not fixed.
• It does not measure effectiveness (deferred by design); it records objective state.

6. Instantiating for a project

36A consumer project runs the sweep from the shipped skill (the Rust stack ships rust-ai-native-sweep) against the shipped tools; the only per-project material is the policy files the gates already require and, optionally, a thin local instance document recording project snapshot numbers and machine-scoped quirks.

37Keep the three layers apart: method (this document), language idioms (the stack's GUIDE/cards), project instance (the consumer's own notes).

38A machine quirk is not project fact; a project number is not method.