OMNIX QUANTUM LTD · Harold Nunes, Editor · May 2026
📖 Browse the Protocol Website → — RFC Index · Public Verifier · Conformance Program
AI agents act. But can you prove who authorized them — and that the authorization was still valid when they acted?
The Agent Trust Fabric (ATF) is an open protocol for cryptographically verifiable AI agent authority governance. It solves three problems that every regulated AI deployment faces:
| Problem | ATF Solution |
|---|---|
| No proof of who authorized an agent action | Delegation Receipts — PQC-signed, chain-traceable to a human principal |
| Authorization may expire or degrade mid-execution | Runtime Continuity Records — continuous authority health at nanosecond precision |
| Evidence is not independently verifiable by regulators | Forensic-grade archive pipeline + self-contained Evidence Packages |
Any auditor with the issuer's public key can verify the complete authority chain — no access to OMNIX infrastructure required.
flowchart TD
HUMAN["👤 Human Principal\n(authorizes agent)"]
AIR["L1 — Agent Identity Record (AIR)\nUnique agent identity + public key\nRFC-ATF-1 §4"]
DR["L2 — Delegation Receipt (DR)\nTask scope + authority budget\nPQC-signed by delegator\nRFC-ATF-1 §5"]
TAR["L3 — Temporal Admissibility Record (TAR)\nDR validity confirmed at execution time\nnanosecond-precise admission timestamp\nRFC-ATF-1 §6"]
RCR["L4 — Runtime Continuity Record (RCR)\nCES = T×0.30 + B×0.30 + D×0.20 + I×0.20\nSampled throughout execution lifecycle\nRFC-ATF-2"]
EAP["L5 — Evidence Archive Pipeline (EAP)\nMerkle-chained, immutable receipt log\nHOT → WARM → COLD retention tiers\nRFC-ATF-3"]
OEP["L5 — Evidence Package (OEP)\nSelf-contained forensic ZIP\nOffline-verifiable by regulators\nRFC-ATF-3"]
HUMAN -->|"issues"| AIR
AIR -->|"signs"| DR
DR -->|"verified at"| TAR
TAR -->|"sampled during"| RCR
RCR -->|"archived to"| EAP
EAP -->|"exported as"| OEP
style HUMAN fill:#2d5a9e,color:#fff
style AIR fill:#1a472a,color:#fff
style DR fill:#1a472a,color:#fff
style TAR fill:#4a1a72,color:#fff
style RCR fill:#4a1a72,color:#fff
style EAP fill:#7a3a00,color:#fff
style OEP fill:#7a3a00,color:#fff
ATF-INV-001 (Monotonic Authority Reduction): Authority budget granted to an agent MUST NOT exceed the delegator's own budget at any delegation depth. Enforced cryptographically at every layer.
| Layer | Artifact | Standard | Invariants |
|---|---|---|---|
| L1 | Agent Identity Record (AIR) | RFC-ATF-1 | ATF-INV-001–006 |
| L2 | Delegation Receipt (DR) | RFC-ATF-1 | ATF-INV-001–006 |
| L3 | Temporal Admissibility Record (TAR) | RFC-ATF-1 | ATF-INV-006 |
| L4 | Runtime Continuity Record (RCR) | RFC-ATF-2 | RGC-INV-001–008 |
| L5 | Evidence Package (OEP) + GPIL + EAP | RFC-ATF-3 | 26 new invariants |
40 total formally specified invariants across three RFCs. Algorithm: ML-DSA-65 (Dilithium-3, FIPS 204) — designed to resist classical and quantum attacks per NIST FIPS 204 (ML-DSA-65).
The ATF stack is composed of three formally specified, interoperable open standards. Each RFC extends the prior without replacing it — the full stack is required for ATF-FEI-Compliant designation.
Defines the cryptographic foundation of the ATF stack: Agent Identity Records (AIR), Delegation Receipts (DR), the Trust Lattice model, Monotonic Authority Reduction (MAR), and post-quantum signing via ML-DSA-65. Six core invariants: ATF-INV-001–006.
| Status | ✅ Published — May 2026 |
| Designation | ATF-Compliant |
| Invariants | ATF-INV-001–006 (6) |
| DOI | 10.5281/zenodo.20155016 |
| SSRN | 6757339 |
| Depends on | — (foundation layer) |
| Specification | RFC-ATF-1.md |
Extends RFC-ATF-1 for long-running agent executions: the Continuity Eligibility Score (CES) formula, Runtime Continuity Records (RCR), Authority Fragmentation Guard (AFG), and the Reauthorization Challenge (RC) protocol. Eight new invariants: RGC-INV-001–008.
| Status | ✅ Published — May 2026 |
| Designation | ATF-RGC-Compliant |
| Invariants | RGC-INV-001–008 (8) · 14 total with RFC-ATF-1 |
| DOI | 10.5281/zenodo.20241344 |
| SSRN | 6763978 |
| Depends on | RFC-ATF-1 — ATF-Compliant implementation required |
| Specification | RFC-ATF-2.md |
Adds the Forensic Evidence Infrastructure layer: policy interoperability across sovereign governance runtimes (GPIL), evidence lifecycle classification with HOT/WARM/COLD retention tiers (ELR), immutable Merkle-chained archive pipeline (EAP), self-contained cryptographically sealed forensic packages (OEP), and export authentication with key isolation (FEA + FVP). Twenty-six new invariants across six families.
| Status | ✅ Published — May 2026 |
| Designation | ATF-FEI-Compliant (highest tier) |
| Invariants | GPIL·ELR·EAP·OEP·FEA·FVP (26) · 40 total with RFC-ATF-1 + RFC-ATF-2 |
| DOI | 10.5281/zenodo.20247342 |
| Depends on | RFC-ATF-1 + RFC-ATF-2 — ATF-RGC-Compliant implementation required |
| Specification | RFC-ATF-3.md |
RFC-ATF-1 ────────────────────────────────── FOUNDATION STANDARD
│ ATF-Compliant (6 invariants)
│
└──▶ RFC-ATF-2 ──────────────────────── RUNTIME GOVERNANCE STANDARD
│ ATF-RGC-Compliant (14 invariants)
│
└──▶ RFC-ATF-3 ───────────── FORENSIC EVIDENCE STANDARD
ATF-FEI-Compliant (40 invariants)
An implementation cannot claim
ATF-RGC-Compliantwithout first satisfying all RFC-ATF-1 invariants. An implementation cannot claimATF-FEI-Compliantwithout satisfying all RFC-ATF-1 and RFC-ATF-2 invariants. Conformance is verified using the ATF Conformance Suite — 92 vectors, zero external dependencies.
The ATF Conformance Suite (atf-conformance-suite/) is the standalone,
platform-independent verification harness for all 40 ATF invariants
across RFC-ATF-1, RFC-ATF-2, and RFC-ATF-3.
- 92 test vectors — positive and negative cases for every invariant
- 3 profiles —
BASE(6 invariants) ·RGC(14) ·ALL(40) - Zero dependencies — Python 3.8+ stdlib only
- Tamper-detectable results — every run produces an
ATFCR-*artifact with SHA-256 integrity
git clone https://github.com/Costenho19/atf-protocol-standard
cd atf-protocol-standard/atf-conformance-suite
python run_conformance.py --profile ALL --output result.jsonATF Conformance Suite 1.0.0
Profile: ATF-FEI-Compliant
Run: 2026-05-16T21:40:35Z
Result ID: ATFCR-FE46C0F921F6EFE4
✓ ATF-INV-001--006 16/16 vectors
✓ RGC-INV-001--008 20/20 vectors
✓ GPIL / ELR / EAP / OEP 42/42 vectors
✓ FEA / FVP 12/12 vectors
Verdict: PASS (86/86)
Hash: sha256:fe6f17e8287f562093018245430b0a4e7055065775f62ea2d3df11a0af719d3d
Full documentation: atf-conformance-suite/README.md
Conformance program: CONFORMANCE.md
| Property | ATF | Open Policy Agent | SPIFFE/SPIRE | JWT / OAuth 2.0 |
|---|---|---|---|---|
| Cryptographic delegation chain (who authorized what) | ✅ PQC-signed receipt per action | ❌ Policy only — no per-action proof | ❌ Identity only — no task scope | ❌ Bearer token — no chain |
| Authority budget monotonicity (MAR) | ✅ ATF-INV-001 — enforced cryptographically | ❌ | ❌ | ❌ |
| Runtime authority health monitoring | ✅ CES formula sampled nanosecond-precise | ❌ | ❌ | ❌ |
| Post-quantum cryptography (FIPS 204) | ✅ ML-DSA-65 | ❌ | ❌ | ❌ |
| Offline regulatory verification (no platform) | ✅ EAP-INV-005 — public key only | ❌ Requires OPA runtime | ❌ Requires SPIRE API | ❌ Requires issuer |
| Forensic evidence archive | ✅ Merkle-chained OEP | ❌ | ❌ | ❌ |
| AI agent-native design | ✅ | ❌ | ❌ | ❌ |
| Published threat model | ✅ THREAT_MODEL.md | Partial | Partial | Partial |
ATF is not a replacement for OPA or SPIFFE. It is a complementary layer: OPA decides whether an action is authorized; ATF proves that it was authorized, by whom, with what budget, and whether authority remained valid throughout execution.
ATF is architecturally aligned with — but does not constitute compliance with — the following frameworks:
| Framework | Requirement | ATF Artifact |
|---|---|---|
| EU AI Act Art. 9 | Human oversight measures for high-risk AI | Delegation Receipt chain traces every agent action to a human principal |
| EU AI Act Art. 12 | Record-keeping for high-risk AI systems | Evidence Archive Pipeline (EAP) + OEP forensic packages |
| NIST AI RMF PR.AA-06 | AI system authorization | ATF-INV-001 (MAR) — authority budget tracked and bounded |
| NIST AI RMF GV.RR | Risk response governance | RGC escalation protocol — MONITORING → WARNING → CRITICAL → HALT |
| ISO/IEC 42001 §9.2 | AI management system audit | OEP self-contained forensic packages — offline-verifiable by regulators |
| NIST CSF 2.0 DE.CM | Continuous monitoring | Runtime Continuity Records — sampled throughout execution lifecycle |
ATF provides the cryptographic evidence layer that these frameworks require. It does not replace a compliance programme — it makes one auditable.
Verify a receipt offline (zero platform dependency):
pip install pypqc
python verifier/verify_receipt.py examples/delegation_receipt.jsonRun the protocol conformance test suite:
pip install pytest pypqc
pytest tests/ -vValidate a receipt against the JSON Schema:
pip install jsonschema
python -c "
import json, jsonschema
schema = json.load(open('schemas/delegation_receipt.schema.json'))
receipt = json.load(open('examples/delegation_receipt.json'))
jsonschema.validate(receipt, schema)
print('VALID')
"Use the reference implementation:
from atf_core import create_delegation_receipt, verify_receipt
dr = create_delegation_receipt(
delegator_id='HUMAN-harold.nunes',
delegate_id='AID-FINANCE-9B8C7D6E5F4A3B2C',
task_scope={'action': 'equity_order_execution'},
budget_granted=60.0,
budget_delegator=100.0,
)
result = verify_receipt(dr)
print(result['verdict']) # PASSatf-protocol-standard/
├── RFC-ATF-1.md ← Delegation protocol (6 invariants)
├── RFC-ATF-2.md ← Runtime continuity (8 invariants)
├── RFC-ATF-3.md ← Evidence lifecycle & forensic (26 invariants)
├── examples/
│ ├── delegation_receipt.json
│ ├── temporal_authority_record.json
│ └── runtime_continuity_record.json
├── schemas/
│ ├── delegation_receipt.schema.json
│ └── runtime_continuity_record.schema.json
├── verifier/
│ └── verify_receipt.py ← Standalone offline verifier (pypqc only)
├── tests/
│ └── test_atf_receipts.py ← Conformance tests (MAR, CES, tamper)
├── reference-implementation/
│ ├── README.md
│ ├── pyproject.toml
│ └── atf_core/
│ ├── __init__.py
│ ├── receipts.py ← DR + RCR creation with invariant enforcement
│ └── verifier.py ← Invariant verification
└── CONTRIBUTING.md
| Family | IDs | RFC | Description |
|---|---|---|---|
| ATF-INV | 001–006 | RFC-ATF-1 | Delegation, signing, MAR, verifiability |
| RGC-INV | 001–008 | RFC-ATF-2 | Continuity, CES, AFG, HALT propagation |
| GPIL-INV | 001–003 | RFC-ATF-3 | Policy interoperability taxonomy |
| ELR-INV | 001–004 | RFC-ATF-3 | Evidence lifecycle retention |
| EAP-INV | 001–007 | RFC-ATF-3 | Archive pipeline integrity |
| OEP-INV | 001–006 | RFC-ATF-3 | Evidence package completeness |
| FEA-INV | 001–005 | RFC-ATF-3 | Export authentication |
| FVP-INV | 007 | RFC-ATF-3 | Forensic verification key identity |
| TOTAL | 40 |
| Designation | Requirements | Layers |
|---|---|---|
| ATF-Compliant | RFC-ATF-1 (6 invariants) | L1–L3 |
| ATF-RGC-Compliant | RFC-ATF-1 + RFC-ATF-2 (14 invariants) | L1–L4 |
| ATF-GPI-Aligned | ATF-RGC-Compliant + signed CRGC with counterpart | L1–L4 + cross-runtime |
| ATF-FEI-Compliant | RFC-ATF-1 + RFC-ATF-2 + RFC-ATF-3 (40 invariants) | L1–L5 |
from atf_core import create_delegation_receipt, verify_receipt
# Human principal delegates authority to an AI agent
dr = create_delegation_receipt(
delegator_id="HUMAN-harold-nunes-001",
delegate_id="AID-UAE-20260516-AABBCCDDEEFF0011",
task_scope={"action": "governance_decision", "domain": "trading"},
budget_granted=0.75, # agent gets 75% of principal's budget
budget_delegator=1.0, # principal has full budget
)
# Verify the receipt independently (offline, no platform needed)
result = verify_receipt(dr)
assert result["verdict"] == "PASS"
print(f"DR {dr['delegation_id']}: {result['verdict']}")
# DR ATFDR-A1B2C3D4E5F60011: PASSfrom atf_core import create_runtime_continuity_record
# Sample runtime health mid-execution
rcr = create_runtime_continuity_record(
tar_id="ATFTAR-1F2E3D4C5B6A7890",
delegation_id=dr["delegation_id"],
agent_id="AID-UAE-20260516-AABBCCDDEEFF0011",
chain_root_id=dr["chain_root_id"],
ces_temporal=99.0, # 99% of time window remaining
ces_budget=100.0, # full budget intact
ces_context=80.0, # 20% context drift
ces_integrity=100.0, # chain integrity perfect
budget_at_admission=0.75,
budget_remaining=0.75,
context_drift_pct=20.0,
)
# CES = 99×0.30 + 100×0.30 + 80×0.20 + 100×0.20 = 94.9
print(f"CES: {rcr['ces_score']} — {rcr['continuity_status']}")
# CES: 94.9 — NOMINAL# Verify any ATF receipt from a JSON file
python verifier/verify_receipt.py receipt.json
# Output:
# ✓ Type: DR (Delegation Receipt)
# ✓ ATF-INV-001 (MAR): budget_granted 0.75 ≤ budget_delegator 1.0
# ✓ ATF-INV-005: content_hash SHA-256 verified
# ✓ ATF-INV-006: receipt is independently verifiable
# VERDICT: PASSPaste any ATF receipt at costenho19.github.io/atf-protocol-standard/verify/ — verifies MAR, CES formula, and content hash client-side. No data leaves your browser.
See CONTRIBUTING.md. We welcome language ports (Go, TypeScript, Rust), conformance test contributions, and feedback on invariant completeness via Issues.
| Version | RFC | Date | Invariants |
|---|---|---|---|
| v3.0.0 | RFC-ATF-3 | May 2026 | 40 total (26 new) · DOI: 10.5281/zenodo.20247342 |
| v2.0.0 | RFC-ATF-2 | Mar 2026 | 14 total (8 new) |
| v1.0.0 | RFC-ATF-1 | Jan 2026 | 6 |
See CHANGELOG.md for the full history.
| Tool | Path | Description |
|---|---|---|
| Receipt Verifier | verifier/verify_receipt.py |
DR + RCR offline verifier — ATF-INV-001–006, RGC-INV-001–004 |
| OEP Archive Verifier | verifier/verify_oep_package.py |
Forensic ZIP archive verifier — OEP-INV-001–006, EAP-INV-001–007 |
Both tools have zero dependency on the OMNIX platform (EAP-INV-005).
pip install pypqc
# Verify a delegation receipt
python verifier/verify_receipt.py examples/delegation_receipt.json
# Verify a full forensic evidence package
python verifier/verify_oep_package.py evidence_package.zip --public-key issuer.b64The ATF Conformance Program provides test vectors for each profile. See CONFORMANCE.md for the full program.
| Profile | Invariants | Test Vectors | Badge |
|---|---|---|---|
ATF-Compliant |
6 (L1–L3) | 15 (8 positive, 7 negative) | [] |
ATF-RGC-Compliant |
14 (L1–L4) | 26 (14 positive, 12 negative) | [] |
ATF-FEI-Compliant |
40 (L1–L5) | 34 (18 positive, 16 negative) | [] |
| Language | Package | Status | Invariants |
|---|---|---|---|
| Python (reference) | reference-implementation/ |
✅ Stable | ATF-Compliant + ATF-RGC-Compliant |
| TypeScript | ports/typescript/ |
✅ Stable | ATF-RGC-Compliant (11 invariants) |
| Go | — | ❌ Wanted | Contribute |
| Rust | — | ❌ Wanted | Contribute |
ATF specifies cryptographic and protocol mechanisms for three things:
- Delegating agent authority — from a human principal to an AI agent, with PQC-signed, chain-traceable receipts (RFC-ATF-1)
- Monitoring authority health at runtime — continuous CES scoring, HALT propagation, escalation (RFC-ATF-2)
- Archiving and verifying the resulting evidence — lifecycle classification, Merkle-chained archive, self-contained forensic packages (RFC-ATF-3)
ATF explicitly does NOT specify:
| Out of Scope | Notes |
|---|---|
| Pre-execution authority resolution | Whether an action should be authorized is outside ATF — ATF records what was authorized and that it was still valid |
| AI model selection or safety constraints | ATF is model-agnostic |
| Real-time revocation notification | HALT propagates within a session; inter-session revocation requires out-of-band coordination (see SECURITY.md) |
| Network transport or API protocols | ATF defines receipt formats, not wire protocols |
| Regulatory compliance | ATF is architecturally aligned with traceability requirements in EU AI Act, NIST AI RMF, and ISO/IEC 42001 — it does not constitute compliance with any of them |
Conformance designations are self-declared based on the test vectors and verification tools in this repository. No third-party certification body is currently designated. See GOVERNANCE.md for the conformance claim process.
standards@omnixquantum.com | https://omnixquantum.com
© 2026 OMNIX QUANTUM LTD. Licensed under CC BY 4.0.