Govern documentation

Five quick sections to get you running in 10 minutes.

1. Sign in at govern.inversify.live → an org is auto-created.
2. Generate an API key on the API Keys page (`inv_govern_...`).
3. POST to `/api/v1/evaluate` with a framework key, or build a graph and POST to `/api/v1/route/{id}/evaluate`.

A Route is a directed graph of nodes. Trigger → (any of: Redact, Classify, Judge, Decide, Webhook) → End.
The engine walks the graph, threading a working envelope (text, scores, decision) between nodes. Decide nodes set the final decision based on configurable allow/warn thresholds.

Govern ships with 5 preset frameworks (read-only): NIST_AI_RMF, EU_AI_ACT, OWASP_LLM_TOP_10, ANTHROPIC_AUP, ISO_42001.

Each framework is a list of controls. A control has an `id`, `name`, an LLM `prompt`, and a `weight`. Custom frameworks (Pro+) let you add your own.

Auth: `Authorization: Bearer inv_govern_...`
Rate limit: 200 req/min/key (configurable per plan).

POST /api/v1/evaluate            — { input, frameworkKey, model? }
POST /api/v1/route/{id}/evaluate — { input, context? }
GET  /api/v1/routes
POST /api/v1/routes
PATCH/DELETE /api/v1/routes/{id}
GET  /api/v1/frameworks
POST /api/v1/frameworks          — { key, name, controls } (Pro+)
PATCH/DELETE /api/v1/frameworks/{key}
GET  /api/v1/evaluations
GET  /api/v1/evaluations/{id}

Optional `Idempotency-Key` header on POST /v1/evaluate*. Cached for 24h keyed on (org, key, inputHash).

Govern signs outbound webhook bodies with HMAC-SHA256.
Header: `X-Govern-Signature: t=<ms>,v1=<hex>`
Verify: hex(HMAC-SHA256(routeWebhookSecret, "<t>.<rawBody>")) === <v1>.
Reject if `Math.abs(Date.now() - t) > 300000` (5-minute window).

The `input` field accepts a string OR an object:

  { "input": { "text": "...", "images": ["data:image/...;base64,...", "https://..."], "pdfUrls": ["https://..."] } }

Up to 10 images and 5 PDFs per evaluation. PDFs are fetched server-side and text-extracted; combined text + image content is then judged. For frameworks that score image content (CSAM_SAFETY), Govern uses a vision-capable model (gpt-4o, gpt-4.1).

Images are SHA256-hashed for the audit trail. The original bytes are NOT retained — keep your own copies.

The CSAM_SAFETY preset framework scores six controls covering grooming language, exploitation indicators, sexualized minor imagery, real-vs-synthetic minor depiction, age-ambiguous + suggestive content, and solicitation.

When CSAM_SAFETY triggers, the response payload carries an `escalation` block:

  {
    "decision": "BLOCK",
    "scores": { ... },
    "escalation": {
      "recommended": "report",
      "framework": "CSAM_SAFETY",
      "guidance": "...",
      "cyberTiplineUrl": "https://report.cybertip.org/",
      "statuteRef": "18 U.S.C. § 2258A — ..."
    }
  }

**Govern is not a registered electronic service provider under § 2258A.** Govern detects and scores; YOU are the legal reporter. Route the escalation block to your incident pipeline and submit reports via the NCMEC CyberTipline as required by your operations.

Re-run any past evaluation against the current route OR a different framework:

  POST /api/v1/evaluations/{id}/replay
  { "useCurrentRoute": true }

Or override the framework:

  POST /api/v1/evaluations/{id}/replay
  { "frameworkKeys": ["NIST_AI_RMF"] }

Replays count against your monthly cap and link to the source via `parentId`.

Both SDKs wrap the public API with idempotency, exponential-backoff retries, and typed responses.

Node:
  npm install @inversify/govern
  const client = new Govern({ apiKey: "inv_govern_..." });
  const r = await client.evaluate({ input: "...", frameworkKey: "OWASP_LLM_TOP_10" });

Python:
  pip install inversify-govern
  from inversify_govern import Govern
  client = Govern(api_key="inv_govern_...")
  r = client.evaluate(input="...", framework_key="OWASP_LLM_TOP_10")