Delivery discipline

<!-- delivery-discipline: exempt --> <!-- This file DEFINES the delivery-discipline. It is documentation about the discipline, not a closeout declaration subject to it. The exemption is legitimate and is logged as `delivery_discipline_exemption_used` by Wave 4 so exemption usage remains auditable. Do not remove this comment unless you are replacing SKILL.md with a proper attestation block. -->

Delivery Discipline

Overview

Agents trained to optimize for completion will ship 60% of what's needed and declare the work closed. The pattern is reproducible. Scientific truth is protected by the existing hooks (confounder harness, R2 review gate, Salvagente Rule). Delivery completeness is not. This skill closes that gap.

You (the agent) must apply the 4 rules below BEFORE writing any deliverable that will declare anything CLOSED, PASS, DONE, SHIPPED, COMPLETE, FINALIZED, or READY in a commit-relevant markdown file. A PreToolUse hook will block your write if you skip the attestation block defined below. A CI validator will catch bypasses.

When To Use

Activate before writing:

• Any *closeout*.md, *status*.md, *summary*.md, *verdict*.md
• Any phase*-*.md, wave*-*.md, sprint*-*.md
• Any skill*.md, SKILL.md updates
• Any README.md or CHANGELOG.md entry that declares a version/release/completion
• Any document declaring a gate PASS, a phase CLOSED, a release SHIPPED

Do NOT activate for:

• Normal prose that mentions "done" in casual context ("I closed the file")
• Code files (.js, .py, etc.) — the skill governs deliverables, not source
• Exploratory notes that do NOT declare closure

The 4 Non-Negotiable Rules

Rule 1 — Surface mapping before writing

Before you write a deliverable that will declare anything CLOSED / PASS / DONE / SHIPPED / COMPLETE, you MUST have a mechanical inventory of what should be covered.

• If the inventory was produced by a subagent (Explore, general-purpose, etc.), you verify at least the critical claims directly before trusting it. Read the actual files the subagent describes. API signatures, file paths, function exports get verified first-hand, not assumed.
• Optimistic sampling is failure. "I covered the main cases" means you haven't looked at the edge cases.

Rule 2 — Explicit scope cuts

Every deliverable MUST declare, in its own body, what was LEFT OUT and why.

• A scope-cut list of 0 items is suspicious by default. If your solution covers "everything", you probably haven't looked hard enough.
• Force yourself to name at least one cut. If you genuinely cannot find one, state that explicitly and pass self-review (Rule 4) first.

Rule 3 — Completeness proof by enumeration

When you declare closure, list the concrete items verified. Not "all tests pass" but "26 items in the surface map, 22 verified directly with Read tool, 4 delegated to subagent report and cross-checked". Enumeration beats summary.

Rule 4 — Adversarial self-review before closure

Before you write "CLOSED" / "PASS" / "DONE", ask yourself: "What would a skeptical reviewer attack?" List at least 3 findings. If you find 0, assume you haven't looked hard enough and go back to Rule 1.

The Attestation Block (MUST appear in every deliverable declaring closure)

Format: a ## Delivery Attestation section (case-insensitive, ## or ###) containing a fenced json block. The hook extracts the first fenced json block under that heading and validates it against skills/vibe/assets/schemas/delivery-attestation.schema.json.

## Delivery Attestation

```json
{
  "covered": [
    "Concrete item 1 verified (not a summary)",
    "Concrete item 2 verified",
    "..."
  ],
  "scope_cuts": [
    {
      "item": "Thing left out of this deliverable",
      "reason": "Why it is outside this deliverable (not 'no time')"
    }
  ],
  "self_review_findings": [
    "Adversarial finding 1 (a skeptical reviewer would attack X because...)",
    "Adversarial finding 2",
    "Adversarial finding 3"
  ],
  "external_review_status": "pending"
}
```

Field semantics:

• covered: array of strings, ≥1 item, each ≥3 chars. Concrete items, not summaries.
• scope_cuts: array of {item, reason} objects. Empty array allowed but suspicious.
• self_review_findings: array of strings, ≥3 items required, each ≥20 chars.
• external_review_status: enum "pending" | "cleared" | "blocked".

Quick Reference

Intent	Action
Writing a phase closeout	Include the ## Delivery Attestation section with filled JSON block
Writing a skill file	Include attestation even if short skill — the skill itself is a deliverable
Updating CHANGELOG	Attestation required when declaring a release / version
Updating README status section	Attestation required when declaring pass/fail or coverage
Exempting a file legitimately	Add <!-- delivery-discipline: exempt --> comment near the top. Wave 4 will log this as a delivery_discipline_exemption_used governance event; frequent use becomes an audit target at that point.

Rationalization Table — DO NOT fall for these

Excuse	Counter
"This is just a small doc, discipline is overkill"	Small docs are where premature closure starts. The attestation takes 60 seconds.
"I don't have time for scope cuts"	The cut list takes 30 seconds. Skipping it costs 30 minutes when user pushes back.
"Self-review feels self-indulgent"	It's the cheap version of R2. Skip it and you pay in R2 corrections later.
"The block only triggers on 'CLOSED' — I'll say 'finalized' instead"	The regex covers synonyms: CLOSED, DONE, PASS/PASSED, SHIPPED, COMPLETE/COMPLETED, FINALIZED, READY, and the failure/partial forms FAIL/FAILED, BLOCKED, PARTIAL, FALSE-POSITIVE, and the review-verdict forms ACCEPTED, REJECTED. Wave 4 will log attempted evasions as governance events; until then, the block itself is the sanction.
"I'll put my closure claim below the legacy-boundary marker in CHANGELOG/README"	The legacy-boundary marker grandfathers pre-Phase-8 content ONLY while its hash=<sha256-hex> attribute matches the LF-normalized below-marker content. Appending any new closure below the marker changes the hash, the boundary becomes invalid, and the whole file is re-subjected to enforcement. To legitimately re-bless the legacy region you must update the hash in a reviewed commit (git diff shows both the new content and the hash change side-by-side). Bare markers without hash= are rejected outright.
"I verified with a subagent, that's enough"	Subagents sample too. Spot-check critical items with Read tool before trusting.
"External review status: cleared" (without actually running review)	Lying about external_review_status is still traceable — when R2 / R3 runs later, it produces external-review-record entries that do NOT match, and the inconsistency surfaces at the next audit. The Wave 3 CI validator checks schema shape (enum + required fields); full governance-event cross-referencing is Phase 8.1 scope, not Wave 3. Don't rely on "nobody will notice" — the record trail is public.
"I'll add the attestation at the end, after writing everything"	Writing the attestation FORCES you to enumerate. Skipping that step defeats the whole point. Write it as you go, not as an afterthought.

Red Flags — STOP and apply the skill

• About to write a deliverable containing "CLOSED", "PASS", "DONE", "SHIPPED", "COMPLETE", or synonyms in declarative context
• About to mark a TodoWrite task completed after a large deliverable
• About to stop the session with a "here's what I did" summary
• About to commit a markdown file whose name matches *closeout*, *status*, *summary*, *verdict*, *phase*, *wave*, *skill*, readme, changelog
• About to declare a release or version bump

All of these mean: stop, map the surface, declare scope cuts, write the attestation block, then continue.

Common Mistakes

• Fake attestation: filling fields with generic strings ("looked at everything"). Wave 2 hook enforces minLength + minItems on the attestation block at write time. Wave 3 CI validator will re-enforce schema shape on tracked markdown at commit time. Full audit-log cross-reference (checking whether external_review_status: "cleared" matches a real R2/R3 review record in governance_events) is Phase 8.1 scope, not Wave 3. Don't assume the current enforcement catches semantic lies — it catches structural ones.
• Skipping scope_cuts because "nothing was cut": if you genuinely cut nothing, you probably didn't look hard enough. Use the self-review step to find a scope cut.
• Putting attestation in a comment or footer: it must be in a ## Delivery Attestation section with a fenced json block. Other placements won't be parsed.
• Writing attestation as YAML or markdown list: must be fenced json block. Other formats fail schema validation.
• Adding exemption comment to bypass: exemptions ARE allowed and don't block. Wave 4 will log each use as a delivery_discipline_exemption_used governance event; frequent use becomes an audit target once that logging is live. Strict mode (VIBE_SCIENCE_STRICT=1) additionally refuses to allow any bypass path (including exemption) when the governance DB cannot be opened or when the governance_events table is missing — i.e. when the Wave 4 audit trail would silently drop records. The strict-mode probe actually opens the DB and checks the schema, not just whether the module imports.

Why This Exists

vibe-science v7.0 protects scientific truth (confounder harness, R2 gate, Salvagente Rule). It did NOT protect delivery completeness — agents could still ship 60% skills, declare phases closed prematurely, sample the surface optimistically.

The pattern was observed repeatedly: Phase 6.1 "closed" but re-retracted when fresh-eyes review found P1s. Phase 7 Wave 1 shipped at 60%. The initial Phase 8 skill was missing 40% of the surface.

This skill (Wave 1) + the PreToolUse hook (Wave 2) + the CI validator (Wave 3) + the governance event log (Wave 4) together force the same discipline on delivery that the existing hooks force on scientific claims. Wave 2 enforcement is active today: a hook that blocks the write, not a review that catches the failure after the fact. Waves 3-4 add CI-time scanning and audit-trail logging so bypass attempts become tracked events rather than silent drift.

When in Doubt

Apply the rules. The cost is 60-120 seconds per deliverable. The cost of skipping is the "user pushes back → retract → patch" cycle we've seen multiple times.

Default: write the attestation. It's cheaper than the correction loop.