Find Wasted Effort (A2 — process-per-outcome ceremony audit)

The process-per-outcome ceremony audit — the sibling of the tokens-per-rule /token-waste-audit (FR-26). Different axis, different inputs, shared discipline.

It is not standalone-with-no-sibling. It reuses the shipped substrate — see flow-skills/token-economy/SKILL.md and hooks/local/token-waste-audit.py for the FP-header text, the read-only posture, and the state/audit/ output convention. Do not reinvent or duplicate token-waste-audit's detection.

Guardrail first (same floor as FR-21 / FR-26)

Ceremony drops, safety never. This audit flags ceremony that is outcome-neutral in the observed window — that is a review candidate, not a remove instruction. A clean window is not proof a control is worthless: a gate stop can be low-frequency/high-severity (catastrophic-low-frequency, policies/ratchet-governance.yml). Every finding states the contrary evidence that would dismiss it; absence of contrary evidence in a short window is inconclusive, never confirmed.

Read-only-to-the-project is load-bearing (D4)

The analyzer writes only inside the gitignored state/audit/ directory — NO memory writes, NO overlay/spec/decisions/provider/policy edits, NO prune/remove application. This holds in Phase 1 (read-only) AND Phase 2A (proposal-output): Phase 2A enriches the output (a Proposed memory entries report section + an optional gitignored state/audit/ proposals JSON) but applies NOTHING. The audit describes candidates and proposes changes a human could apply; the PO owns subtraction (A3 prune protocol). The actual memory write-apply is Phase 2B (DEFERRED, consumer-repo prototype, AC2b). If asked to "just remove the dead gate" or "apply the proposal," refuse: "Per the A2 read-only-to-the-project posture (D4), I surface review candidates and propose patches; removal/apply is Phase 2B + the PO's call via the policies/ratchet-governance.yml prune protocol."

Verdict vocabulary (every rule emits one)

The 6 active rules (rule 4 is CUT)

Per-rule input shapes, signatures, and false-positive examples: references/rule-signatures.md and references/false-positive-examples.md. Cite them; don't restate.

Inputs (already on disk — read-only)

• docs/specs/*/verification-gate.md, gate reports, deploy reports (docs/tmp/handoff/*, docs/handoff/*). A report feeds window verdicts only when linked to a selected commit: a ## Conclusion: record (Outcome/Task/Commit), or the gate-report template's **Status:** line (presence check) plus a Per-task commit table row whose SHA resolves to a selected commit whose subject cites that T-number. One linked report = one window body; each §7 Gate satisfaction FAIL row is a gate block and its §6 Deviations text supplies approvals, so fixed template text (the §9 relay block) never counts. Everything else stays historical.
• approval artifacts (state/approvals/*.json; link only via explicit task + commit fields, so Flow-minted approvals, made before the commit, stay historical), git log (per-task/per-round commit shape)
• round structure: a commit joins the spec-slug round of the one linked gate report whose own Per-task commit table row resolves it and whose **Task range:** covers that row; a range never claims a commit missing from its table. Else the subject slug, else the commit; change-notes docs/changes/*, future round-files
• prevents: annotations + policies/ratchet-governance.yml (A3 taxonomy + coverage map)

Output

• Mode-A chat summary (totals + per-rule confirmed/dismissed/inconclusive + coverage + proposal count).
• Gitignored report state/audit/find-wasted-effort-<date>.md carrying the FP header (reuse token-economy's), the per-rule findings, a coverage section naming which prevents:-annotated controls were in scope (D5 — silence is not safety), and (Phase 2A) a Proposed memory entries section.
• Phase 2A (read-only-safe): a Proposed memory entries report section + an optional gitignored state/audit/find-wasted-effort-proposals-<date>.json sibling (skip with --no-proposals-json). Both live only under state/audit/.
• NO writes beyond state/audit/ in any phase here: no memory, no overlay, no spec/decisions/provider/policy edits, no prune application. The write-apply is Phase 2B (DEFERRED, consumer-repo, AC2b).

Proposal schema (Phase 2A)

A proposal is a change a human could apply — the audit emits it, never applies it. Emitted from confirmed findings (rules 1,2,3,5,7) and rule-6 per-element review candidates; inconclusive/dismissed emit none.

Self-output quarantine (Codex #5): the evidence collectors do NOT read state/audit/, so the audit can never cite its own prior output as evidence. Proposals cite only raw artifacts.

Coverage statement (mandatory in every report)

The report states its own coverage: which artifacts it read, which rules it could evaluate, and which ratchet-governance.yml coverage entries were in scope. An un-annotated control NOT in the coverage map is reported as coverage gap, not as safe.

Growth rule

A ceremony-waste pattern that recurs across audits and matches no rule above → add one rule row (with its required contrary evidence + an FP example in references/) via skill-authoring. Project-specific ceremony patterns stay in project docs/skills, not here.

Anti-patterns

• Treating a finding as a verdict to act on — they are candidates; the PO prunes (A3), gated on FP fixtures (P2).
• Confirming a catastrophic-low-frequency control as waste on a clean window — that inverts the guardrail.
• Duplicating /token-waste-audit's execution-layer signatures (rule 4 cut; rule 7 scoped to cross-session ceremony) — different axis, no overlap.
• Auto-reclassifying a lane or auto-removing a gate — out of scope in every phase of this skill.

Clean-room note

Original Fusebase Flow content. Designed after reviewing public AI coding workflow patterns; no third-party code, prompts, skill files, or hook scripts are copied. Reuses the in-repo token-economy substrate by reference. See docs/source-map.md.