Policy Gateway backend guide

Policy Gateway splits management APIs (config, projects, history) from the enforcement endpoint.

All endpoints require Authorization; use a user JWT for management and a policy API key (ak_...) for enforcement.

• GET/POST /api/policy-gateway/config - read/write the saved policy config.
• POST /api/policy-gateway/simulate - compute decisions for category scenarios.
• GET /api/policy-gateway/history - list revision/simulation/enforcement entries.
• GET/POST /api/policy-gateway/projects - manage projects and monthly limits.
• GET/POST /api/policy-gateway/projects/{project_id}/keys - create scoped policy keys.
• DELETE /api/policy-gateway/keys/{key_id} - revoke a key.
• POST /policy/chat/completions - enforce policy on live traffic.

POST a config object to store the policy and create a revision entry in history.

• You can send the config directly or wrap it under a top-level config key.
• Missing fields are filled with defaults; list values are trimmed and capped.
• Set org_controls.audit_logs to false to disable enforcement history writes.

curl https://api.abliteration.ai/api/policy-gateway/config \
  -H "Authorization: Bearer $ABLIT_JWT" \
  -H "Content-Type: application/json" \
  -d '{
    "config": {
      "policy_id": "policy-gateway",
      "name": "Support Policy",
      "rules": {
        "allowlist": ["refund policy", "account support"],
        "denylist": ["illegal instructions"],
        "response_pattern": "rewrite"
      }
    }
  }'

Policy Gateway applies a consistent evaluation order so outcomes are predictable and testable.

• Allowlist is strict: if an allowlist is configured and nothing matches, the decision becomes refuse.
• Denylist + flagged categories: either can trigger a non-allow decision.
• Rewrite flag: if response_pattern is rewrite and rewrite_instead_of_refuse is false, the decision becomes refuse.
• Reason codes: the first code containing the decision string is used, otherwise defaults apply.
• Moderation gate: when enabled, sexual/minors and self-harm categories are rejected before policy evaluation.

Send OpenAI-compatible chat completions through the policy endpoint.

• policy_id is optional; if provided it must match the saved (or inline) policy id.
• Provide policy_user and policy_project_id (or headers) for quotas and audits.
• Inline policy overrides are supported via config or policy in the request body.
• Responses include a policy object with decision metadata and rollout state.

curl https://api.abliteration.ai/policy/chat/completions \
  -H "Authorization: Bearer $POLICY_KEY" \
  -H "Content-Type: application/json" \
  -H "X-Policy-User: user-12345" \
  -H "X-Policy-Project: support-bot" \
  -H "X-Policy-Target: support-bot" \
  -d '{
    "model": "abliterated-model",
    "messages": [
      { "role": "user", "content": "Summarize our refund policy." }
    ],
    "policy_id": "policy-gateway"
  }'

Rollouts are configured in the policy config and targeted with X-Policy-Target or policy_target.

• Shadow rollouts mark sampled requests as rollout_mode=shadow and do not enforce.
• Canary rollouts enforce only the sampled percentage; unsampled requests are allowed.
• Shadow takes precedence over canary when both are enabled.
• If automatic rollback is enabled and enforcement spikes, Policy Gateway temporarily disables enforcement.

Create a project per app or agent, issue a scoped key, and attach policy_user for per-user quotas.

• Project ids are normalized to lowercase slugs; the response contains the canonical id.
• If org_controls.project_keys is true, API-key requests must include a project id (or use a scoped key).
• Project monthly limits override config project_quota and always use a monthly window.
• User quotas use policy_user (or API key id when omitted) and honor the configured window.

curl https://api.abliteration.ai/api/policy-gateway/projects \
  -H "Authorization: Bearer $ABLIT_JWT" \
  -H "Content-Type: application/json" \
  -d '{ "name": "Support bot", "monthly_token_limit": 10000000, "monthly_request_limit": 20000 }'

curl https://api.abliteration.ai/api/policy-gateway/projects/support-bot/keys \
  -H "Authorization: Bearer $ABLIT_JWT" \
  -H "Content-Type: application/json" \
  -d '{ "label": "Support bot prod" }'

Use simulation for dry runs and history for audits.

• Simulation compares categories against rules.flagged_categories only.
• Set persist: false to skip writing a simulation entry.
• History results are newest-first and capped by POLICY_HISTORY_LIMIT (default 50).

curl https://api.abliteration.ai/api/policy-gateway/simulate \
  -H "Authorization: Bearer $ABLIT_JWT" \
  -H "Content-Type: application/json" \
  -d '{ "categories": ["self-harm/intent"] }'

curl "https://api.abliteration.ai/api/policy-gateway/history?type=enforcement&limit=20" \
  -H "Authorization: Bearer $ABLIT_JWT"