Content Credibility Engine

Meta Museum treats content generation and syndication as a credibility system, not a volume system.

This document defines four operational layers:

1. Trust layer: provenance, authorship, traceability, rights.

1. Originality layer: semantic novelty and source-grounded synthesis.

1. Distribution layer: channel orchestration, SEO metadata, syndication.

1. Consistency layer: cadence, voice control, and review discipline.

Scope and compatibility

This policy extends existing repository rules in:

• `CLAUDE.md`
• `docs/roadmap.md`
• `docs/linked-art/LinkedArtModel1.0-Reference.md`

If any conflict appears, Linked Art fidelity, rights safety, and "cite or refuse" remain non-negotiable.

1) Trust layer requirements

Trust metadata is persisted in:

• `provenance/ledger.json`
• `provenance/source-map.yaml`

Required metadata per publishable artifact:

• `artifactId` stable identifier
• `contentHashSha256` over canonical source bundle
• `author` and `reviewers`
• `createdAt` and `publishedAt`
• source list with URL, provider, retrieval timestamp
• citation coverage evidence
• rights/reuse disposition

Minimum enforcement:

• No publish without at least one source reference.
• No publish with unresolved rights status.
• No publish if citation coverage fails policy threshold.

2) Originality layer requirements

Originality records are persisted in:

• `semantic-core/originality-index.json`

Baseline novelty policy:

• compute embedding-space novelty against recent published corpus
• classify as "meaningfully original" when cosine distance is above `0.18`
• require one "unique insight" note per artifact (human-written or curator-approved)

Guardrails:

• template-only outputs are rejected
• uncited paraphrase outputs are rejected
• all generated claims must map to source evidence or be removed

3) Distribution layer requirements

Distribution control artifacts:

• `distribution/schedule.yaml`
• runtime queue database at `distribution/queue.db` (gitignored)

Channel set (initial):

• Web
• LinkedIn
• Medium
• Email
• API feed/syndication

Required pipeline behavior:

• per-channel publish caps and minimum cadence checks
• snippet extraction and summary variants
• SEO metadata generation (title, description, canonical URL, tags)
• rights-safe media checks before enqueue
• queue-worker orchestration across `web`, `linkedin`, `medium`, `email`, and `api` channels

4) Consistency layer requirements

Voice and cadence controls:

• `generation/style-profile.md`
• `distribution/schedule.yaml`

Baseline policy:

• minimum 2 outputs per week across channels
• maximum 1 output per day per channel by default
• style profile versioned with last-updated reviewer

5) Monitoring and audits

Monitoring baseline persisted in:

• `monitoring/metrics.json`

Tracked dimensions:

• trust score
• originality score
• distribution reach
• engagement velocity

Weekly audit checks:

• citation/link integrity
• novelty drift
• rights/compliance regressions
• broken outbound links
• engagement velocity minimum threshold
• trust/originality score regression thresholds vs configured baseline

Automation:

• Local run: `pnpm credibility:audit`
• Gate run (non-zero on failure): `pnpm credibility:audit:check`
• Weekly GitHub Action: `.github/workflows/credibility-audit.yml`
• schedule: every Monday
• checks: validation drift, eval relevance threshold, markdown broken-link scan, engagement/trust/originality eval-threshold alerts
• artifact: `artifacts/credibility-audit/latest.json`

6) Implementation notes for this repository

• OpenTelemetry is the default tracing substrate for this workflow.
• Postgres is the system of record where queue/provenance services graduate from file-backed prototypes.
• Redis remains the short-lived cache layer for scoring and queue coordination.
• FastAPI service extensions are allowed where they fit current C2/C4 service boundaries.
• Additional AI framework dependencies (for example LangChain/LlamaIndex) are optional and must pass the standard "ask before acting" dependency gate in `CLAUDE.md`.

7) Definition of done for credibility-ready content features

1. A test exists that fails when citations or rights metadata are missing.

1. A test exists that fails when originality metadata is absent for generated artifacts.

1. Queue/schedule policy is enforced by code, not just documentation.

1. Metrics are emitted and trace-linked for generation, review, and publish operations.

1. Human approval is explicit before publication actions.