/AUDIT-AUTHORITY.local

Audit authority ladder (local supplement, not synced)

Not synced from upstream. Local supplement kept outside SKILL.md / reference/audit.md so sync-impeccable.sh (verbatim wipe) does not touch it. The upstream audit flow is the source of truth for how to audit; this file documents how audit's findings reconcile with the OTHER three quality authorities in the frontend plugin.

Reachability (known gap). This file is injected by two channels only: the design-md-first UserPromptSubmit hook (when DESIGN.md is present) and the frontend-expert coordinator's Report step. The upstream-synced reference/audit.md cannot reference it (verbatim policy). Consequence: a standalone frontend:impeccable (argument: audit) run with no DESIGN.md and no coordinator will not see this ladder — the ladder only matters when another quality authority is also running, so apply it when you have cross-authority findings to reconcile, otherwise audit's own heuristic output stands.

The four quality authorities measure different axes

When the frontend plugin runs multiple quality checks on the same UI, four authorities can each emit findings on contrast / typography / spacing / a11y:

Authority	Skill / Agent	Evidence type	What it actually measures
design-md lint	frontend:design-md (lint --format json)	computed	Token resolution, WCAG contrast ratios (deterministic)
impeccable audit	frontend:impeccable (argument: audit)	heuristic	Spacing "cramped", hierarchy, composition (LLM judgment)
web-design-guidelines	frontend:web-design-guidelines	standard-citation	Fetched standards compliance checklist
anti-patterns agent	frontend-anti-patterns agent	pattern-match	AI-slop tells (gradient text, icon-tile stacks, etc.)

Reconciliation by evidence type (not by static priority)

A static priority (design-md > impeccable > ...) is the wrong abstraction because these authorities measure different axes. Reconcile by evidence type:

• computed supersedes heuristic on the same node. design-md's contrast ratio of 3.1:1 beats impeccable audit's "contrast looks okay" for the same color pair. design-md is deterministic fact; audit is LLM estimate.
• pattern-match is additive, not substitutive. The anti-patterns agent's slop findings (e.g. "gradient text detected") ADD to the audit's findings; they don't replace them. A slop hit on a UI that impeccable audit "passed" is redundancy catching an LLM omission, not two authorities in conflict — re-run the relevant impeccable sub-command (e.g. quieter / typeset) on the flagged area.
• standard-citation is advisory. web-design-guidelines findings SUGGEST; they don't block. Surface them as recommendations, not as P1 blockers.
• Same evidence type, same node, conflicting value → the more specific / more recent authority wins. In practice this is rare; default to design-md (computed) for token-level conflicts.

Where audit emits

impeccable audit is the single reconciliation outlet for heuristic findings: it should incorporate (not duplicate) computed findings from design-md lint and additive pattern-match findings from the anti-patterns agent, then emit one consolidated finding set. The coordinator's Report step (no longer the reconciliation point under Option B) consumes that set.

Absolute Bans (upstream authoritative)

The text bans live in upstream SKILL.md's ### Absolute bans (8 rules, synced verbatim). The executable detector rules live in scripts/detector/registry/antipatterns.mjs (~40 ids). Note: hero-metric and glassmorphism-as-default are text-only bans with no corresponding registry rule; the other bans map to advisory registry rules.