000 Jeremy Content Consistency Validator

Universal document consistency auditor. Deterministic drift checks across docs, code, tests, and CI — conflicts resolved through a per-fact-class authority registry, never a global ranking. Read-only: it reports, it never fixes.

What Changed in v3

v3 folds the full consistency-audit engine into this plugin and retires the v1/v2 shell (design council, issue #991):

• The "website is truth" axiom is gone. No source is globally authoritative anymore. Authority is declared per fact class, as data, by a human — in sot-map.yaml.
• Deterministic and LLM-judged checks are structurally separated. Judged findings are advisory-only: never Critical, never blocking.
• Email/CRM sources are out of scope for v1 of this model. No WebSearch, no WebFetch — filesystem and git only.
• The multi-agent roster is deferred pending open design decisions. Phase 1 ships one skill plus a thin command; no agents.
• Fixture-gated. The quality gate is a golden fixture corpus with seeded drifts; evals assert exact findings ("exactly these N, zero invented"), not procedure.

How It Works

1. Load the authority registry — sot-map.yaml (default ~/000-projects/intent-os/sot-map.yaml, path configurable). Each row maps a fact class (e.g. version-string, license, ci-commands) to the artifact class that owns it.
2. Inventory artifacts — README, CLAUDE.md, CHANGELOG, 000-docs/, docs/, planning/, CI workflows, package manifests.
3. Run the 9 drift checks (below), each tagged deterministic or LLM-judged.
4. Resolve conflicts through the registry. A conflict on a registered fact class is filed against the non-authoritative artifact. A conflict on an unregistered fact class is emitted as unowned fact-class — human adjudication needed — no winner is ever guessed.
5. Report — Part A deterministic findings (Critical/Warning/Info), Part B advisory findings (LLM-judged), unowned-fact-class callouts, and bootstrap-drafted registry rows for human review.

The 9 Checks

The Registry Model

sot-map.yaml is the source of truth for who owns which fact — declared as data, adjudicated by a human:

version: 1
fact_classes:
  version-string:
    authority: package-manifest
    adjudicated_by: jeremy
    adjudicated_on: YYYY-MM-DD

• No row → no guess. Unowned fact classes are reported for human adjudication.
• Bootstrap mode. Without a registry, the skill still detects every conflict, names no winners, and drafts proposed rows (using project-type detection and the retired legacy hierarchy purely as a drafting heuristic). A human reviews and commits the rows — the skill never writes sot-map.yaml.
• Truth invariant. The governed brain arbitrates asserted company/doctrine claims and the rules — never generated facts. The validator never reads the brain as ground truth for generated facts, and the brain ingests only human-adjudicated findings.

Full specification: skills/validate-consistency/references/sot-registry.md.

Installation

/plugin marketplace add jeremylongshore/claude-code-plugins
/plugin install 000-jeremy-content-consistency-validator@claude-code-plugins-plus

Usage

/validate-consistency

Or naturally: "check consistency", "validate docs", "audit documentation", "doc drift check".

Also invoked automatically by /release during Phase 1.6 — deterministic 🔴 Critical findings feed the release blocking gate; advisory findings are surfaced for review and never block.

Read-Only Guarantee

The skill's tool grant is Read, Glob, Grep, Bash(echo:), Bash(git:), Bash(diff:*) — no Write, no Edit, no WebSearch, no WebFetch. It reports discrepancies with file paths and line numbers; you decide what to fix.

Scope (Phase 1)

• One skill (validate-consistency) + one thin command. The multi-agent roster from the original design is deferred pending open decisions.
• Sources: local filesystem + git. Email/CRM auditing is out of v1 entirely.