forge-quality — Quality Intelligence Layer
Layer-2 instruction (progressive disclosure): this file is small; domain knowledge lives in layer 3
(.claude/config/quality/cards/*.md, 0 tokens until loaded). The module is
.claude/forge-bin/forge-quality.cjs; the canonical domain list is
.claude/config/orchestration/domain-catalog.json — on a conflict the CATALOG wins and the
diverging seam gets fixed, never a second list started.
Commands
node .claude/forge-bin/forge-quality.cjs analyze "<the mission>" # THE entry point: profile + playbook + lenses + omissions + card descriptors + council in one JSON (F-14)
node .claude/forge-bin/forge-quality.cjs profile "<the mission>" # the MissionProfile only
node .claude/forge-bin/forge-quality.cjs card <slug> # one knowledge card (layer 3, on-demand)
node .claude/forge-bin/forge-quality.cjs drift # catalog vs 4 seams — every difference is a finding (exit 3 on drift)
node .claude/forge-bin/forge-quality.cjs kernel # the fixed Quality Kernel (~600-800 tokens)
node .claude/forge-bin/forge-quality.cjs gate --domain <domain> # the domain's quality gate card as a PASS/FAIL/NOT_RUN checklist (--json; unknown domain -> GENERIC or PROVISIONAL card, exit 0; exit 2 invalid name, 3 invalid card)
node .claude/forge-bin/forge-quality.cjs gate --list # every gate card with its check counts (--schema prints the card schema)
Workflow (router Step 1 — required for BUILD tasks)
- Run
analyze; machine output beats any human-written table.classification_confidence: nonemeans: the fallback is an ASSUMPTION — resolve the type explicitly (intake) before building. - Lenses: every lens gets EXACTLY one disposition (RELEVANT/NOT_APPLICABLE/DEFERRED/OWNER_GATED) with a reason. Silently skipping does not exist; neither does silently adding one.
- Omissions are candidate requirements (pre-validated;
omissions_validmust be true — a hand-edited card goes back throughvalidateRequirementCard). A RESEARCH_HYPOTHESIS never becomes a hard requirement directly — promotion requiresconfirmed_by{type: local-validation| official-source|experiment, ref}. - Knowledge cards: load only descriptors with
exists: trueviacard <slug>— max 6 cards, max 3 retrieval rounds. A missing card is an honest notice, never invented content.
Council (selective, never the default)
councilTrigger decides: NONE is the default; FULL only on an explicit request or high impact +
high uncertainty. Real dispatch runs through the Agent tool; the CouncilDecisionRecord is validated
with validateCouncilRecord — that validation is shape_only (labels don't prove a principal; real
provenance is owner-gated, see OWNER-GATED.md). Council consensus is NEVER proof — only
tests/measurements are.
Quality gate cards (Verify Boss)
One card per catalog domain in .claude/config/quality/cards/ as <domain>.json (the .md files there
are the knowledge cards above). A card holds at most 12 checks, each with claim, evidence_required,
how_to_run, severity (block|warn) and honesty_note, plus hard_gates, not_applicable_when and
gaps. Every check cites a section of the domain playbook (or a FORGE_HARD_RULES.json rule) and a
fragment that must still be found there: gate re-checks the citations and prints STALE CARD when the
playbook changed. Where a playbook has no verifiable gate, the card says so in gaps instead of inventing
one. gate only lists: Verify Boss marks each check PASS / FAIL / NOT_RUN with evidence, and a block check
that is FAIL or NOT_RUN fails the gate.