Find wasted effort

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.

	/find-wasted-effort (this)	/token-waste-audit (FR-26)
Axis	process-per-outcome (ceremony)	tokens-per-rule (consumption)
Inputs	Flow artifacts ON DISK: gate/deploy reports, handoffs, approval artifacts, git log, round structure, prevents: annotations	Claude Code transcripts (~/.claude/projects/.../*.jsonl)
Question	"Which ceremony step bought no safety outcome?"	"Which tokens bought no information?"
Shared	candidate/false-positive header · read-only-first posture · state/audit/<date>.md gitignored output	(same)

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)

Verdict	Meaning	Requires
confirmed	the waste signature held AND the required contrary evidence was searched for and not found	the contrary-evidence search, stated
dismissed	contrary evidence found — the ceremony bought a real outcome (a gate that blocked, a control that fired, a deliberate exception)	the contrary evidence, cited
inconclusive	signature present but the window is too small / ambiguous / a catastrophic-low-frequency control	why it can't be confirmed

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.

#	Rule	Signature	Required contrary evidence (→ dismissed)
1	Unused gate stops	N rounds where the gate approved every deviation, no deviation blocked	ANY blocked-gate counterexample in the window ⇒ dismissed (never suggest Middle eligibility for that class). Suggestion is eligibility for review, never auto-reclassify.
2	Per-commit full-suite habit	full-suite runs ≫ 2/round with identical fail-sets each run	a run whose fail-set DIFFERED (the suite caught a real regression mid-round) ⇒ dismissed for that round
3	Artifact duplication	the same rule/evidence block copied verbatim in ≥3 round artifacts	intentional self-bootstrapping (handoff role-prelude, template scaffolding) ⇒ dismissed
4	Context-rebuild overhead	CUT — net-new	already shipped in /token-waste-audit's v3.21.0 cross-session aggregate. POINT at that report; do not re-implement the signature.
5	Lane misclassification	small diff + zero design decisions but Full ceremony was run	any design decision / risk surfaced in the round ⇒ dismissed. Ambiguous size/risk ⇒ inconclusive; never auto-reclassify a lane.
6	Ratchet inventory	a ceremony element with NO prevents: annotation (A3) AND no firing in the window	a prevents: annotation present, OR a firing in the window, ⇒ dismissed. catastrophic-low-frequency element on a clean window ⇒ inconclusive (expected to sit idle), never confirmed. Output = a review candidate, never "remove".
7	Watch-vs-read waste (cross-session ceremony layer ONLY)	cross-session ceremony where a later session re-watched/re-derived what an earlier session already recorded durably	the durable record was absent (a real observability gap, not waste) ⇒ dismissed. Scoped to the cross-session ceremony layer — do NOT duplicate FR-26's execution-layer record-then-read/polling signature (token-economy / smoke-testing own that).

Inputs (already on disk — read-only)

• docs/specs/*/verification-gate.md, gate reports, deploy reports (docs/tmp/handoff/*, docs/handoff/*)
• approval artifacts (state/approvals/*.json), git log (per-task/per-round commit shape)
• round structure (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.

Field	Meaning
proposal_id	stable <rule>-<verdict-kind>-<hash> id (deterministic)
rule	source rule number
verdict	confirmed (rules 1,2,3,5,7) or prune_review_candidate (rule 6 — never an auto-prune)
raw_evidence_refs	pointers to raw on-disk artifacts — never a prior audit report/proposal (self-output quarantine)
target_kind / target_path	what / where a human could change (often an (operator decision))
exact_patch	the concrete change text a human could apply (description, not an applied diff)
operator_confirmation_required	always true
source	always "audit"

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.