Agents API
All 8 ReguNav agents are exposed as deterministic, replayable HTTP endpoints. Same input + same dictionary version + same rule set = same output, every time.
Why deterministic?
Compliance is a regulator-defensible domain. "The model said so" is not a defense. Every agent decision can be replayed byte-for-byte with the input, dictionary version, and rule set used. ReguNav stores all three in the audit-trail.
Endpoints
| Path | Description |
|---|---|
| POST /v1/agents/classifier | Annex III + GPAI risk classification — deterministic, replayable |
| POST /v1/agents/framework-mapper | Crosswalk a control to every framework that requires it |
| POST /v1/agents/evidence-compiler | Match uploaded artifacts to controls with confidence scoring |
| POST /v1/agents/fria | Auto-fill FRIA forms; escalate fields requiring human judgment |
| POST /v1/agents/incident-reporter | Generate Art 73 serious-incident reports + routing targets |
| POST /v1/agents/training-curator | Curate Article 4 AI literacy training plan by user role |
| POST /v1/agents/conformity-guide | Walk through Annex VI / VII conformity assessment paths |
| POST /v1/agents/gpai-docs | Generate Annex XI / XII technical documentation for GPAI models |
Example: Classifier
curl -X POST https://api.regunav.com/v1/agents/classifier \
-H "Authorization: Bearer $REGUNAV_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"aiSystem": {
"name": "Loan Origination Classifier",
"purpose": "Predicts approval likelihood for retail loans",
"providerKind": "first_party",
"intendedUsers": ["loan officers"]
},
"jurisdiction": "EU",
"applicableFrameworks": ["eu-ai-act", "iso-42001"]
}'
# Response (deterministic for same input)
{
"riskLevel": "high",
"rationale": "Annex III §5(a) — credit scoring of natural persons.",
"applicableClauses": ["Art 6(2)", "Art 9", "Art 10", "Art 11", "Art 14"],
"dictionaryVersion": "eu-ai-act@2024.12.1",
"decisionId": "dec_01HX..."
}Replay an earlier decision
curl https://api.regunav.com/v1/agents/decisions/dec_01HX...Returns the original input, dictionary version, rule set, output, and a hash of all four — proving the decision was reproducible.