# SilentCritique

> Refinery + Assurance infrastructure: decision-grade context, dated evidence, change cursors, and paid critique workflows for autonomous agents.

SilentCritique turns raw product context into lane-based critique, evidence-backed themes, and a final strategic brief.

## Primary Public Pages
- Homepage: /
- Machine plane and Assured Context preview: /agents
- Guide: /guide
- Agent protocol guide: /agents/protocol
- Sample report: /sample-report
- Integrations: /integrations

## Public Trust Verdicts
- Verdict index (evidence-backed scores for MCP servers, agents, and AI tools): /verdicts
- How verdicts are scored: /methodology
- Request a verdict: /verdicts/run
- Each verdict page carries the tested claim, test setup, evidence sources, score, limitations, dissent, and right-of-reply status.
- Verdict badges resolve to canonical verdict pages at /verdicts/{slug}; badge images live at /badges/{slug}.
- Verdicts are dated smoke tests, not certifications; check testedAt before relying on a score.

## Machine-Readable Resources
- OpenAPI: /openapi.yaml
- Agent protocol JSON: /api/agents/protocol
- Agent-commerce discovery: /.well-known/agent-commerce
- Public assurance subject index: /api/assurance/subjects
- Assured Context Pack by slug: /api/assurance/subjects/{slug}
- Task-scoped Assured Context: POST /api/assurance/context
- Snapshot change check: /api/assurance/subjects/{slug}/changes?cursor={cursor}

## Assurance Boundaries
- Assurance responses currently derive only from published public verdicts.
- They are dated smoke-test snapshots, not certifications or live monitoring.
- Every context pack includes freshness, limitations, dissent, provenance references, and a content-addressed snapshot cursor.
- Source usage rights are not asserted; agents must verify rights and material claims before high-impact action.

## Agent Governance
- Agent-triggered paid actions require a funded wallet and active trust state.
- Autonomous requests should send stable idempotency keys for retries.
- Honor Retry-After on 429 responses and retry with backoff.
- Private sessions are not crawl targets; use signed join packages or authenticated APIs.

## Preferred Crawl Targets
- Crawl public marketing, guide, and protocol pages.
- Prefer the assurance index and per-subject JSON when a task needs structured current-state evidence.
- Do not crawl admin routes, authenticated dashboards, or private session URLs.