Governance: The Answers, Enforced

Agent-maintained knowledge raises hard questions: who verified this? does it still hold? how do you know the agent followed it? Each answer below points at the mechanism that enforces it — not a policy sentence. Two terms recur: source_refs — the per-claim citation (source file, locator, verbatim quote) — and change-sets — proposed edits staged for human approval before they touch canonical pages.

Question	The fabric's answer	Where it lives
Does every claim have a source or evidence type?	Claims without source_refs can't be supported; locators + verbatim quotes mandatory	lint CLAIM checks; provenance rules in AGENTS.md
What changes when source documents change?	sha256 per capture; hash drift flags exactly which claims are affected; re-ingest produces a change-set, never silent overwrite	lint SOURCE-DRIFT; refresh workflow
Canonical policy vs. tentative research?	Status taxonomy: proposed → supported → contested → superseded → retracted; patterns carry maturity 0–3	frontmatter contracts; lint gates
What wins when global and project rules conflict?	Precedence: project > domain > global, enforced at manifest compile time and validated against path scope	wf context; lint SCOPE
How does an agent propose a knowledge mutation?	Change-sets (manifest + diff), staging→canonical flow, human gate before canonical writes	ingest workflow; apply_changeset.py (wf apply-changeset)
Which artifacts require human review?	Pattern promotion (7-point checklist), canonical merges, corpus conflicts (SYNC-CONFLICT blocks push)	promotion-queue; sync protocol
How do we find stale pages after a dependency upgrade?	review_after dates lint-checked, sha256 staleness from capture drift, graphify AST diff flags code changes	lint REVIEW-AFTER; graphify-bridge --diff
Do patterns know where they fail?	recommended/standard patterns without counterexamples warn (boundaries expected); delivered patterns render "does not apply"	lint PATTERN; wf context
Are the locators still true?	wf review --verify-locators re-checks every claim's quote against its raw source: rewrite/repair/restore/contest — the verification stamp means something	#111/#130 pass
What does the fabric NOT know?	Concept open-questions harvest into gated question pages (open → answered via answers relations); surfaced by the gate + query	#88 loop; lint QUESTION

None of these are promises in a README — each is a lint check or workflow gate in CI right now.

The human gate in practice: one change-set, three commands

The same governance question — "does the agent follow it?" — is answerable by watching any session's change-set flow. Real outputs from this repo:

$ wf gate                    # ONE report of everything awaiting a human
gate: nothing pending — fabric is current (HITL queues empty)
# (with work pending it lists stale claims, promotion dossiers,
#  domain proposals, open change-sets — each with its deadline)

$ wf apply-changeset <slug> --dry-run    # inspect BEFORE anything canonical
Pages to create: 2
[DRY RUN] Would apply git apply: .../change-sets/.../diff.md
[DRY RUN] no changes written

$ wf lint                    # the 0-error gate before any commit
0 error(s), 6314 warning(s)  # warnings are advisory; errors fail the commit

The scoped auto-promote exception (#190)

The human gate is absolute for dossiers, claims, domains, and questions. For inbox pattern candidates — the trivially-promotable class — a fabric may opt into the judged-confident tier:

tuning:
  promotion:
    auto_apply: true         # OFF by default — a policy change must be explicit
    auto_threshold: 0.90     # calibration-sensitive: overrides travel with provenance (G-J)

wf promote-patterns --auto then bands each candidate: deterministic preconditions first (candidate type, complete provenance, age ≥ 1d, no judged-different verdict — one refusal means the human), then ONE calibrated judgment call. p ≥ auto_threshold applies (the same deterministic move, stamped auto_applied: {judge, prob, at} and logged — reversible with promote-patterns --unapply); the near-band below the floor escalates to the human gate; further below stays in the inbox with the verdict recorded. The gate's auto-applies INFO section lists recent applies (never actionable alone) — auto-promote downgrades rows, it never silences the gate. The judge proposes, never writes: the apply is a file move, the candidate text is untouched.

wf skill ingest / wf skill promote print the full agent-facing procedures — so any harness (Claude Code, opencode, Codex) runs the same governed loop without the human dictating steps each time.