Compliance

EU AI Act enforcement, Federal Compliance (NIST, FedRAMP, CMMC), Governance Proof Tokens, the model inventory, and the Policy Registry.

EU AI Act Compliance

The EU AI Act's high-risk requirements apply from 2 December 2027 for Annex III systems, and from 2 August 2028 for AI in Annex I products (Digital Omnibus on AI, Regulation (EU) 2026/1744). GaaS implements five governance policies covering Articles 9, 10, 13, 14 and 15, giving high-risk AI deployments a compliance-ready audit trail from day one.

High-risk requirements apply 2 December 2027
Maximum Fine €15M or 3%
Turnover Cap 6% global annual turnover

Implemented Policies

Policy ID Article Name Failure Mode
pol_euaia_001 Art. 9 Risk Management System FAIL
pol_euaia_002 Art. 10 Data Governance & Management CONDITIONAL
pol_euaia_003 Art. 13 Transparency & Information CONDITIONAL
pol_euaia_004 Art. 14 Human Oversight FAIL CRITICAL
pol_euaia_005 Art. 15 Accuracy, Robustness & Cybersecurity FAIL
High-Risk Domains: These policies activate automatically when the intent target falls within a high-risk domain: healthcare, critical_infrastructure, education, employment, essential_services, law_enforcement, migration, justice.

API Endpoints

Compliance Status Response

// GET https://api.gaas.is/v1/compliance/eu-ai-act
// X-API-Key: your_api_key

{
  "compliance": {
    "org_id": "org_abc123",
    "assessed_at": "2026-09-28T18:30:00",
    "overall_status": "partial",
    "implemented_count": 6,
    "partial_count": 4,
    "requires_action_count": 1,
    "not_applicable_count": 0,
    "high_risk_ai_detected": true,
    "enforcement_deadline": "2027-12-02",
    "articles": [
      {
        "article": "Article 9",
        "title": "Risk management system",
        "requirement": "Establish, implement, document and maintain a risk management system throughout the lifecycle of the high-risk AI system.",
        "status": "implemented",
        "gaas_capability": "GaaS governance membranes define the risk boundary for each AI agent. The 5-stage pipeline (intent validation → enrichment → policy evaluation → deliberation → decision assembly) constitutes the risk management system. Audit trail provides lifecycle documentation.",
        "evidence": [
          "GET /v1/membranes/current",
          "GET /v1/audit/chain/verify",
          "pol_euaia_001"
        ],
        "notes": null
      }
      // … one entry per article
    ]
  }
}

Federal Compliance

GaaS maps governance controls to four major US federal frameworks, enabling AI systems to satisfy procurement and security requirements for government agencies and defense contractors.

NIST AI Risk Management Framework 1.0

Maps to all 4 NIST AI RMF functions (GOVERN, MAP, MEASURE, MANAGE) across 23 subcategories. GaaS provides automated evidence generation for each subcategory based on live pipeline data.

NIST SP 800-53 Rev. 5 Moderate Baseline

5 enforcement policies (pol_nist800_001–005) covering 4 control families: Access Control (AC), Audit & Accountability (AU), System & Information Integrity (SI), and Incident Response (IR).

FedRAMP Moderate Baseline

5 enforcement policies (pol_fedramp_001–005) for cloud services selling to US federal agencies. Includes 3PAO-ready evidence packages with NIST SP 800-53A assessment procedures.

CMMC 2.0 Level 1–2

4 enforcement policies (pol_cmmc_001–004) covering Level 1 (basic safeguarding of Federal Contract Information) and Level 2 (advanced protection aligned to NIST SP 800-171 Rev. 2) for Defense Industrial Base contractors.

Federal Procurement Summary

A single endpoint that aggregates compliance status across all four frameworks, plus 3PAO evidence packages per control.

SOC 2 Control Mapping

GaaS maps 20 SOC 2 Trust Services Criteria controls (CC1–CC9 and Availability) to concrete platform capabilities, each with API evidence endpoints and [EXAMINE/TEST] assessment procedures. This is a control mapping to support your own SOC 2 audit — it is not a SOC 2 attestation of GaaS, and statuses are honest: controls with open gaps (e.g. a formal penetration-testing program) report partial, not implemented.

Fetch any control's evidence package via GET /v1/compliance/federal/evidence/{control_id} (e.g. CC6.1 logical access, CC8.1 change management, A1.2 recovery infrastructure). Evidence endpoints include the hash-chained audit export, chain verification, SLO reports, and the policy registry.

ISO/IEC 42001 Control Mapping

GaaS maps 15 ISO/IEC 42001 Annex A controls (AI policy, internal organization, resources, impact assessment, life cycle, data, transparency, responsible use, and third parties) to platform capabilities with API evidence endpoints and [EXAMINE/TEST] assessment procedures. This is a control mapping to support your AI management system audit — it is not an ISO/IEC 42001 certification of GaaS, and statuses are honest: incomplete capabilities (e.g. independent model validation, automated sub-processor notification) report partial.

Fetch any control's evidence package via GET /v1/compliance/federal/evidence/{control_id} (e.g. A.2.2 AI policy via governance membranes, A.5.2 impact assessment via six-dimensional risk scoring, A.6.1.2 life cycle via the membrane DRAFT→SHADOW→LIVE progression).

Exporting Evidence for Your GRC Platform

Pull every control evidence row — across NIST AI RMF, NIST SP 800-53, FedRAMP, CMMC, NIST CSF, SOC 2 TSC, and ISO/IEC 42001 — in one call for manual upload to Vanta, Drata, or any GRC platform: 

GET /v1/compliance/evidence/export?framework=SOC%202%20TSC&format=csv

Omit framework for all frameworks. format=csv returns one row per control (id, family, title, requirement, status, capability, evidence endpoints, assessment procedures, last assessed) with a download filename; format=json returns the full evidence packages. Operator role required.

Governance Proof Tokens

When a GaaS deployment has a signing key configured, every live decision carries a Governance Proof Token (GPT) in decision.governance_proof_token: a signed record of what was decided, for which agent, and which audit record anchors it. Every field is signed, so changing any of them — the verdict, the organization, the audit hash — breaks the signature. Shadow and test decisions get no token.

Anyone holding a token can check it: an auditor, a regulator, an insurer, or the other party to a transaction. They need no GaaS account.

Token Fields

Field Description
token_id Random UUID4. Anyone holding it can check the token.
issued_at When GaaS issued the token (UTC).
decision_id / intent_id The decision and intent this token covers.
agent_id / org_id The agent that asked, and its organization.
verdict The decision's verdict as the API returns it: approve, approve_modified, block or escalate.
pipeline_mode Always live: shadow and test decisions get no token.
policies_evaluated / policies_passed How many policies ran, and how many passed.
risk_score / risk_classification The decision's risk score (0.0–1.0) and class.
audit_ref / audit_hash The audit record the decision is anchored to, and that record's hash in the tamper-evident audit chain.
key_id Which signing key; matches kid in /.well-known/gaas-audit-keys.json.
gaas_signature ECDSA P-256 with SHA-256, DER-encoded, hex — over every other field (see below).

Checking a token online

Example token

{
  "token_id": "ea3fb059-42d6-4663-b9a5-366aefa536f0",
  "issued_at": "2026-09-28T18:45:17.265864Z",
  "decision_id": "dec_001d59151d7b4d629c482c5544e6c93f",
  "intent_id": "int_94fb1d4e6e034df287c112d2dd81340d",
  "agent_id": "agent_farm_bot",
  "org_id": "org_default",
  "verdict": "approve",
  "pipeline_mode": "live",
  "policies_evaluated": 10,
  "policies_passed": 10,
  "risk_score": 0.0001,
  "risk_classification": "low",
  "audit_ref": "aud_05f7496b8a614707b135b822905368f2",
  "audit_hash": "25a4fef2a81c9872b30609828ad092a50ae69f5da03f873f8c0e340399999fa5",
  "key_id": "b3e18dff8282417d",
  "gaas_version": "0.3.0",
  "gaas_signature": "3046022100a4d7519db157a1e049526e7f391509607d6a05ddc5e4141ee63be486a8e4d5420221009ac5c64e2dde91721b140150fe1161d165dd4e4d45205ea6646ff56e985e3a60"
}

Example check

// GET https://api.gaas.is/v1/verify/proof/ea3fb059-42d6-4663-b9a5-366aefa536f0  (no API key; "token" field omitted here)
{
  "token_id": "ea3fb059-42d6-4663-b9a5-366aefa536f0",
  "valid": true,
  "signature_valid": true,
  "chain_integrity": true,
  "verdict": "approve",
  "agent_id": "agent_farm_bot",
  "org_id": "org_default",
  "issued_at": "2026-09-28T18:45:17.265864Z",
  "key_id": "b3e18dff8282417d",
  "error": null
}

Checking a token offline

The signed bytes are the token's JSON with gaas_signature removed, keys sorted, and no whitespace. The signature is ordinary ECDSA P-256 / SHA-256 over those bytes, so any crypto library can check it: 

# Python (pip install cryptography)
import json
from cryptography.hazmat.primitives import hashes, serialization
from cryptography.hazmat.primitives.asymmetric import ec

token = json.load(open("token.json"))
signature = bytes.fromhex(token.pop("gaas_signature"))
payload = json.dumps(token, sort_keys=True, separators=(",", ":")).encode()
key = serialization.load_pem_public_key(open("gaas-audit-key.pem", "rb").read())  # "pem" in the JWK
key.verify(signature, payload, ec.ECDSA(hashes.SHA256()))  # raises InvalidSignature if anything changed

# Or with jq 1.7+ (it keeps numbers exactly as written) and OpenSSL
jq -cS 'del(.gaas_signature)' token.json | tr -d '\n' > payload.json
jq -r .gaas_signature token.json | xxd -r -p > signature.der
openssl dgst -sha256 -verify gaas-audit-key.pem -signature signature.der payload.json   # Verified OK

Model Inventory

GaaS auto-generates an inventory of every AI agent in your organization from agent profiles and live audit data — purpose, owner, validation status and decision history — for your model risk management, with no spreadsheet required.

On SR 11-7: the inventory was built for the Federal Reserve's SR 11-7 and OCC Bulletin 2011-12. Both were rescinded on 17 April 2026 by SR 26-2 and OCC Bulletin 2026-13, “Revised Guidance on Model Risk Management”, which states that generative and agentic AI models are not within its scope. GaaS does not claim the inventory meets a supervisory letter.

Validation Status Mapping

GaaS Agent Status Inventory Validation Status
CERTIFIED VALIDATED — Model has passed full independent validation
VERIFIED IN_VALIDATION — Validation in progress; compensating controls active
REGISTERED PENDING_VALIDATION — Model registered but not yet submitted for validation

API Endpoints

Example

# GET https://api.gaas.is/v1/model-inventory/agent_fintech_01
# X-API-Key: your_api_key

{
  "agent_id":            "agent_fintech_01",
  "model_name":          "Payment Authorisation Agent v2.3",
  "model_type":          "decision_engine",
  "owner":               "treasury@acme-bank.com",
  "purpose":             "Real-time payment fraud scoring and authorisation",
  "sr117_status":        "VALIDATED",
  "gaas_status":         "CERTIFIED",
  "last_validation_date": "2026-01-15",
  "next_review_date":    "2026-07-15",
  "decisions_last_90d":  48391,
  "block_rate":          0.023,
  "escalation_rate":     0.004,
  "audit_coverage":      "100%"
}
Automatic Updates: The inventory refreshes in real time as decisions flow through the pipeline. Validation status, decision counts, block rates, and escalation rates all reflect live data — no manual entry required.

Policy Registry

The Policy Registry is npm for AI governance — a curated catalogue of versioned, composable policy packs that you can install into your membrane with a single API call. Packs are maintained by the GaaS team and the community, pinned to a semver, and verified for compatibility before installation.

Seed Packs

Pack ID Description Policies
gaas-core-v1 Foundational governance policies for all AI deployments 10
gaas-healthcare-v1 HIPAA patient access, minimum necessary, TCPA consent for patient communications 3
gaas-financial-v1 SOX, AP2 payment mandates and spend limits, PCI DSS, PSD2 SCA, AML velocity 8
gaas-privacy-v1 GDPR lawful basis, CCPA opt-out, FERPA education records 3
gaas-eu-ai-act-v1 EU AI Act Articles 9/10/13/14/15 for high-risk systems 5
gaas-nist-csf-v1 NIST Cybersecurity Framework 2.0 5
gaas-nist-800-53-v1 NIST SP 800-53 Rev. 5 Moderate Baseline (AC, AU, SI, IR) 5
gaas-fedramp-moderate-v1 FedRAMP Moderate Baseline with 3PAO evidence packages 5
gaas-cmmc-v1 CMMC 2.0 Level 1–2 for Defense Industrial Base 4

API Endpoints

Installing the EU AI Act Pack

# Install the EU AI Act pack
# POST https://api.gaas.is/v1/policy-registry/gaas-eu-ai-act-v1/install
# X-API-Key: your_api_key

{
  "membrane_id": "mem_xyz789",
  "version":     "1.0.0"
}

// Response
{
  "status":           "installed",
  "pack_id":          "gaas-eu-ai-act-v1",
  "version":          "1.0.0",
  "policies_added":   ["pol_euaia_001", "pol_euaia_002", "pol_euaia_003", "pol_euaia_004", "pol_euaia_005"],
  "membrane_id":      "mem_xyz789",
  "effective_from":   "2026-08-01T00:00:00Z",
  "requires_restart": false
}
Zero Downtime: Policy pack installation is hot-reloaded — your membrane applies the new policies to incoming intents immediately without any service interruption or redeploy.

Audit Reconstruction & Verification

GaaS provides auditor-facing endpoints for independently verifying governance decisions without live system access. Download self-contained verification bundles (ZIP archives) or retrieve structured decision reconstructions with inline hash verification.

Verification Bundles

Verification bundles are ZIP archives containing audit records, governance proof tokens, a self-contained Python verification script (verify.py), and a manifest. Auditors can verify hash chain integrity and cryptographic signatures offline.

Decision Reconstruction

Returns a structured JSON report for a governance decision, including all 5 pipeline stages, inline SHA-256 hash re-verification, and chain position validation. Suitable for regulators who need to understand exactly how a decision was made.

Auth: All reconstruction endpoints require operator role or above.

Related Pages

Ready to achieve EU AI Act compliance?

Provision your governance membrane in under five minutes and have all five EU AI Act policies active before August 2026.

Get started free →