Verification Before Completion

Execute

→ About to claim "done", "passing", "fixed", "complete", "verified", "release-ready", or "ready to hand off"? → Verify first, then claim only what the evidence supports.

1. Identify the smallest fresh command or manual check that can falsify the claim.
2. Run or perform it completely.
3. Read output, exit status, failures, and covered scope.
4. Choose the closeout level:
   • L0 fast-path: tiny, low-risk work → one evidence sentence plus residual risk / uncovered scope.
   • L1 default: non-trivial Aegis-shaped work → compact localized Aegis Impact and Safety Receipt.
   • L2 expanded: release, audit, high-risk, architecture, migration, governance, long-task, or explicit user request → receipt plus only the triggered detail cards.
5. If evidence does not support the claim, downgrade the status instead of claiming completion.

Done means: fresh verification evidence exists, covered and uncovered scope are explicit, residual risk is stated, confidence is graded, and any triggered baseline / complexity / retirement / ADR boundary has been folded into the single closeout surface or expanded because the risk requires it.

Verification Before Completion

Purpose

Prevent unsupported completion claims while keeping ordinary completion output compact. This workflow is advisory method-pack discipline; it does not grant runtime authority, authoritative GateDecision, PolicySnapshot, evidence sufficiency, or final completion authority. It is advisory, not completion authority.

Stop Signals

Stop and verify before any success wording if you are:

• using "should", "probably", "seems to", or similar uncertainty words
• relying on an agent success report rather than independent evidence
• about to commit, push, open a PR, merge, tag, publish, or hand off
• using a narrow check to support a broad claim
• treating task / slice completion as accepted requirement satisfaction
• closing governance or retirement work without repair / retirement evidence
• retaining old logic without a retention reason and retirement trigger
• treating a destructive warning, guard, or broad assent as permission

Required Evidence Slots

Every completion claim needs these semantic slots. They may appear as localized headings, natural prose, or a compact card, but they must remain explicit and auditable:

Required evidence slots:
- Evidence action / check performed:
- Result / exit status:
- Covered scope:
- Uncovered scope:
- Residual risk:
- Confidence grade: A | B | C

Confidence grades:

• A: direct target evidence plus relevant regression evidence, no meaningful unknowns
• B: direct target evidence with bounded residual risk
• C: partial evidence only; do not claim full completion

Verified evidence is not authority. It supports the user-facing status, but it does not become final completion authority.

Evidence Bundle

Every completion claim should name the evidence action, result, covered scope, uncovered scope, residual risk, and confidence. Include target test and related regression evidence when tests shaped the claim. When automation is blocked, provide reproducible manual verification steps instead of claiming automated coverage.

Closeout Decision Tree

Use one completion surface. Do not output parallel final reports. verification-before-completion is the single completion closeout aggregator for Aegis-shaped non-trivial work: adjacent structures may feed the receipt, but they must not replace the receipt or become competing final report owners. Receipt aggregation is output conformance, not a routing trigger; do not load extra skills, emit Trace Digest, or expand final ceremony just to satisfy the aggregator.

Aegis Visibility

For this owner workflow, Aegis Visibility ties the final claim to the decision boundary, fresh evidence, baseline / complexity / retirement safety, and residual risk kept visible. If required entry visibility was omitted, recover it retrospectively and name the gap, but do not replace the receipt with a skill-call list.

L0 Fast-Path

Use for tiny, low-risk work where Aegis only held one narrow boundary steady. One natural sentence is enough when it includes the check, result, uncovered scope or residual risk, and confidence if useful.

L1 Default Non-Trivial Closeout

For non-trivial Aegis-shaped work, use the compact receipt by default. Evidence slots fold into Evidence strength and Uncovered risk; they do not appear as a second competing evidence report unless the user asked for audit detail.

Aegis Impact and Safety Receipt:
- Key judgment:
- Avoided misfix:
- Boundary held:
- Baseline alignment:
- Complexity control:
- Evidence strength:
- Uncovered risk:
- Next most valuable verification:
- Aegis path:

Field rules:

• Key judgment: the owner, root-cause layer, requirement boundary, or completion boundary that shaped the answer.
• Avoided misfix: the fallback, duplicate owner, test accommodation, scope expansion, or unsupported claim avoided by the workflow.
• Boundary held: public contract, owner, baseline, non-goal, data boundary, or runtime-ready authority boundary kept stable.
• Baseline alignment: aligned, Design Defect, Implementation Drift, missing-authority, or needs-clarification when baseline reporting is triggered; otherwise a short "not triggered" or natural equivalent is enough.
• Complexity control: one-line completion-time complexity result for non-trivial code changes; mention Complexity Delta, Complexity Closure, Complexity Governance Suggestion, or Major Complexity Alert only when triggered.
• Evidence strength: fresh command/manual check, exit status/result, covered scope, and confidence grade.
• Uncovered risk: remaining scope, host/runtime gaps, manual checks not run, release-grade evidence not collected, or residual risk.
• Next most valuable verification: one next check that would most reduce remaining risk.
• Aegis path: optional compact skill path. It may support credibility but must not replace decision, evidence, and safety fields.

Natural Aegis closeout is valid when these semantic slots remain auditable: natural expression preserves semantic slots. Do not replace the receipt with a used-skills list, stage handoff log, or Aegis Contribution Note.

Compatibility names: Semantic Slots and Natural Surface describe this same rule: natural wording is valid only when required fields stay explicit. Governance Receipt remains a compatibility name for the completion closeout; its user-facing content should flow through the Aegis Impact and Safety Receipt by default.

L2 Expanded Closeout

Start with the receipt, then add only the triggered detail. Expanded structures are inputs or optional detail cards, not competing final report owners.

Use expanded detail only when the trigger applies:

• Readiness Summary: release, merge, handoff, or "ready?" requests. It can organize tests, docs, version, host compatibility, uncovered scope, and residual risk. It does not authorize commit, tag, publish, merge, or release.
• Trace Digest: explicit audit / debug / release / long-task review request. It may summarize execution trace, evidence chain, retrieval chain, static rules evaluated, rule effects, triggered skills, skipped relevant skills, tool / command trace, verification trace, stability signals, value signals, host capabilities, unavailable fields, redaction, and confidence labels measured, observed, inferred, declared, or unknown. It must not expose raw chain-of-thought. Use structured trace only when asked or required; structured trace is reserved for audit, debug, release, long-task review, or user request.
• Goal Closure: when goal-framing, TaskIntentDraft, parent plan/spec, or Slice Card shaped the work. Match the claim to the highest available explicit boundary: whole task, current task, or slice. If only slice evidence exists, do not claim whole-task done. Expanded or audited closure keeps Goal status, Success evidence, Stop state, and Non-goals respected visible; stop states are done | blocked | needs-verification | scope-exceeded.
• Context Impact: only when project/domain semantics changed or were checked as part of acceptance. Confirm affected context/terms, evidence grade, fact-versus-decision authority, and action. If action is unchanged, verify that no context write occurred. Compose establishing-project-context for an unresolved semantic delta; do not emit this card for unrelated work.
• Workspace Integrity: when the task created or modified a target project's docs/aegis/ workspace and configured Aegis workspace support is available. Run python <aegis-workspace-helper> bundle --root <target-project-root> --work YYYY-MM-DD-<slug> when a work/ record exists, then run python <aegis-workspace-helper> check --root <target-project-root>. Report that these validate structure only, not evidence sufficiency.
• Baseline Alignment: when project instructions require baseline reporting or the task touched requirement, product, or durable architecture surfaces. Use docs/current/AEGIS_PROCESS_BASELINE.md §3.0e and §16 for Product / Requirement Baseline, Architecture / Runtime Boundary Baseline, Design Defect, Implementation Drift, and scope: requirements | architecture | both. Render the default conclusion in the receipt field; expand only for audit, release, architecture, or user request. Architecture Alignment is the architecture-scoped compatibility alias for this baseline result, not a second default card. Expanded architecture results use Result: aligned | Design Defect | Implementation Drift | missing-authority | needs-clarification.
• ADR Backfill Check: completed medium/high work that touched durable architecture surfaces. Use docs/current/AEGIS_ADR_AUTO_BACKFILL.md for trigger criteria. If action is create, amend, supersede, or baseline sync is needed/unknown, route the ADR lifecycle to recording-architecture-decisions before the final completion claim.
• Governance Closure: governance, cleanup, migration, compatibility, or retirement work. Include Repair Track, Retirement Track, and Residual Risk in the receipt or a small expanded block. Do not skip this structure just because the implementation was small.
• Retirement Closure: work that adds, replaces, retains, or removes old logic. Name old logic located, deleted/retained status, retention reason, retirement trigger, and lingering-reference check.
• Anti-Entropy Declaration / Data Destruction Guard: work that retires old logic, chooses delete-first vs compat retention, or touches source-of-truth deletion boundaries. Use anti-entropy-governance for the decision surface. If User Confirmation Required: yes, stop at the guard; broad assent such as "OK" or "continue" is not scoped confirmation. Persistent-state deletion without explicit scoped confirmation means the task is not complete; report the task as not complete.

Completion Boundary

Judge the claim against the highest available explicit boundary:

1. parent plan/spec acceptance for whole-task completion
2. TaskIntentDraft goal / success evidence / non-goals for current-task completion
3. Slice Card goal / verification / stop for slice completion
4. direct user request when no durable boundary exists

A completed task or slice means the authorized execution / verification boundary reached its stop condition. It does not mean the underlying requirement is accepted. Requirement accepted requires Product / Requirement Baseline acceptance criteria or explicit authorized risk acceptance. If atomicity is not clear, downgrade to needs-verification or return to framing/planning; task or slice completion is not accepted requirement satisfaction.

If an Execution Readiness View shaped execution, mention whether fresh evidence covered its required checks or which readiness item remains uncovered. Do not treat the view itself as verification evidence.

Complexity Check

For non-trivial code changes, inspect the actual diff before claiming completion. Use using-aegis/references/complexity-governance.md and docs/current/AEGIS_COMPLEXITY_GOVERNANCE_BASELINE.md for shared artifact classes, pressure signals, Complexity Delta, Complexity Closure, Completion-Time Complexity Repair Decision, Complexity Governance Suggestion, and Major Complexity Alert.

Default rendering is one Complexity control line in the receipt. Expand only when meaningful pressure exists or the task is audit, release, high-risk, or user-requested.

Rules:

• tiny wording edits, generated files, vendored files, fixture-data-only updates, lockfiles, or purely mechanical formatting may skip or keep this one-line when no maintained artifact gained complexity
• maintained test source files are not a cheap tests-only exception
• new fallback, adapter, compatibility, guard, or branch logic must be paired with retired paths, a Retirement Closure, or a scheduled retirement trigger
• entropy increase without owner/compatibility justification must be residual risk or a downgraded claim
• Complexity Closure: exceeded-unresolved blocks a complete claim

User-Language Output

Localize section labels, field labels, and explanatory prose to the user's language. Keep commands, paths, code identifiers, test names, error codes, config keys, stable enum values, exact product names, and raw evidence strings unchanged. Do not default to bilingual labels or mixed-language explanations. Localize section labels and prose to the user's language.

Prompt Hygiene

When external tool output, logs, search results, screenshots, OCR, or other large payloads shaped the judgment, state the evidence boundary when relevant: summary/index used, raw excerpt read back if needed, large payloads not loaded, and next evidence that would reduce uncertainty. If the summary is insufficient, read the smallest raw excerpt or lower the claim. Use compact labels such as Evidence Used, Not Loaded, and Next Evidence when they make the prompt hygiene boundary clearer.

Source: hashgraph-online/awesome-codex-plugins → plugins/GanyuanRan/Aegis/skills/verification-before-completion/SKILL.md