On this page
State and artifacts
The files a run writes so reviewers can inspect state, plans, checks, and closure.
Evidence lives in files you can open.
The work record is not a decorative transcript dump.
For phased or larger runs, the run folder carries the state cursor, roadmap, phase checklists, before/after evidence, failure rows, and final audit closure.
Run folder
For phased or larger work, the run folder (run root) is .IMPLEMENTAUDIT/runs/<task-slug>-<id>/.
- It is the durable home for run state and should be inspectable after the chat scrolls away.
- Compact direct runs may instead bind the work record through the audit object, findings ledger, transcript markers, and cited artifacts without a separate run root.
- The namespaced path prevents one run from clobbering another run's artifacts.
Files to inspect
STATE.mdrecords status, phase, decisions, and failure rows (Andon rows).ROADMAP.mdand phase checklists/specs show planned work.THINKING.mdkeeps reviewable planning evidence.PROTOCOL.mddefines the execution loop, Andon escalation, continuity decision rules, and final audit protocol.tools.mdandcontext.mdrecord Stage 0 detection and operating context.- Before/after evidence shows what changed.
- The final audit ledger explains closure or handoff.
What this proves
Artifacts prove what the run recorded and checked.
They do not prove commits, pushes, releases, publication, or provenance unless those gates were separately authorized and evidenced.
Source surfaces
README.mdAGENTS.mdskills/implementaudit/SKILL.mdskills/implementaudit/templates/STATE.mdskills/implementaudit/templates/ROADMAP.mdskills/implementaudit/templates/THINKING.mdskills/implementaudit/templates/PROTOCOL.mdskills/implementaudit/templates/tools.mdskills/implementaudit/templates/context.mdskills/implementaudit/templates/sidecars.mdskills/implementaudit/references/phase-design.md
Artifact map
Work record
- Maintainer term
tdqyq-audit-object- Owner
- The run
- Purpose
- Evidence-bearing record: request, scope, findings, owner/source, claims, changed files, checks, and closure state.
- Proof boundary
- Everything the run claims is bounded by this record.
Audit action
- Maintainer term
ydqyq-audit-action- Owner
- Runtime
- Purpose
- Operations that inspect, classify, verify, patch owner/source, close, refuse, or hand off.
- Proof boundary
- Not a file; it is the operation category that updates the work record.
ROADMAP.md
- Owner
- Planner
- Purpose
- High-level phase plan, dependencies, milestone, and dispatch order.
- Proof boundary
- Planning evidence, not immutable scope.
STATE.md
- Owner
- Runtime
- Purpose
- Status token, current phase, route, findings ledger, authorization stance, baseline, and classed Andon log.
- Proof boundary
- Authoritative run-state record; any optional sidecar mirror is secondary.
tools.md
- Owner
- Runtime
- Purpose
- Stage 0 detection: skill-dir resolution, helper availability, runtimes, and sidecar status.
- Proof boundary
- Detection evidence only.
context.md
- Owner
- Runtime
- Purpose
- Operating context: baseline ref, working-tree state, invocation shape, route, and repo instructions read.
- Proof boundary
- Intake record; live files win on conflict.
THINKING.md
- Owner
- Runtime
- Purpose
- Reviewable planning evidence: route, risks, real-file notes, before/after checks, and quality-route fields, with formal Lean labels in the operating-method reference.
- Proof boundary
- Reasoning record, not a public claim by itself.
PROTOCOL.md
- Owner
- Runtime
- Purpose
- Execution protocol: phase loop, resume rules, Andon escalation protocol, continuity decision rules, and final audit procedure.
- Proof boundary
- Protocol/rule surface; it does not prove those steps ran. Step evidence must appear in
STATE.md, transcript markers, check output, and final audit records.
sidecars.md
- Owner
- Runtime
- Purpose
- Optional-tool detection and authorization, Graphify trigger/freshness evidence, ActiveGraph fork/diff checkpoint use or mirror status, and Markdown fallback.
- Proof boundary
- Qualified orientation or non-authoritative mirror context, not lifecycle authority or correctness proof.
applied-context.md / applied-memories.md
- Owner
- Runtime
- Purpose
- Loaded continuity from prior runs, personal notes, or project notes.
- Proof boundary
- Orientation only; live files and applicable
AGENTS.mdoutrank it.
repo-map.md
- Owner
- Runtime
- Purpose
- Manual map of owner/source candidates; qualified first-contact Graphify terrain may orient broad code components.
- Proof boundary
- Orientation only; deterministic searches and real-file inspection establish owners and references.
phases/phase-N.md
- Owner
- Planner/runtime
- Purpose
- Per-phase goal, acceptance criteria, rollback/defer path, quality route, and evidence plan.
- Proof boundary
- Phase-scoped claim boundary.
Before/after evidence
- Owner
- Phase runtime
- Purpose
- Baseline checks, changed files, added lines, deliverable comparison, and rerun output.
- Proof boundary
- Shows observed change; does not prove publication or provenance.
Final audit ledger
- Owner
- Final audit
- Purpose
- Closure, gaps, handoff, unresolved risk, evidence references, and completion marker order.
- Proof boundary
- Completion evidence only within the authorized scope.
State tokens
The state file uses lifecycle status tokens: open, READY_TO_DISPATCH, IN_PHASE, PAUSED, BLOCKED, INTERRUPTED, and DONE.
Transcript markers such as AUDIT_COMPLETE, AUDIT_HANDOFF, and IMPLEMENTAUDIT_RUN_COMPLETE are separate closure evidence, not STATE.md status values.
Work-record fields (audit object)
The run's evidence-bearing work record (audit object/action surfaces) should name:
- the request
- source
- scope
- route
- owner/source decisions
- criteria
- checks
- authorization state
- final disposition
Lightweight invocations may create this from a normal repo request; a separate audit file is not required for a first try.
Continuity decisions
AGENTS_UPDATE_DECISION and CONTINUITY_DECISION are artifact decisions.
They explain whether durable instructions or continuity notes were updated, skipped, or handed off.
Continuity writeback marker
IMPLEMENTAUDIT_CONTINUITY_SAVED appears only when a bounded continuity writeback actually happens.
The marker records:
TargetReasonEvidenceBoundaryAuthorizationNot saved
It proves the writeback event and boundary, not that the written content is universally correct.
Local git trace
Git status, diff, and proposed commit text can be evidence.
They are not permission to commit, push, tag, release, publish, or claim provenance.
- Record proposed commit scope and granularity when useful for review.
- Keep source, checker, docs, release, and provenance lanes separate.
- Treat a clean diff as review evidence, not as authorization to mutate remote state.
Artifact boundary
Run artifacts prove audit behavior.
They do not by themselves prove product delivery, host load, publication, or provenance.