Four Tests Standard (4TS) - Verifiable AI Governance

Version: 1.0.3
Steward: FERZ LLC
License: CC BY-NC-ND 4.0 (specification), MIT (schemas/test vectors)

The TCP/IP layer for AI governance
Ensure consequential AI decisions are stop-capable, owned, replayable, and escalatable—by design.

📄 Academic Publications

📖 Foundational Paper

Verifiable AI Governance: The Four Tests Standard (4TS) and Proof-Carrying Decisions
Edward Meyman | October 2025

Available on:

• SSRN (Primary)
• ResearchGate

Establishes formal specification, theoretical foundations, and proof of necessary and sufficient conditions for verifiable AI governance.

---

How to Cite:

Meyman, E. (2025). Verifiable AI Governance: The Four Tests Standard (4TS) 
and Proof-Carrying Decisions. SSRN. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5688982

Overview

The Four Tests Standard (4TS) is a vendor-neutral technical specification for verifiable AI governance. It enables organizations to prove that AI systems in regulated industries meet compliance requirements through:

• Proof-Carrying Decisions (PCDs): Canonical JSON objects encoding all information needed to verify governance at decision boundaries
• Deterministic Verification: Mathematical acceptance criteria that produce consistent results independent of who verifies
• Two Replay Modes: State-Replay for byte-exact reproduction, Protocol-Replay for gate-based validation
• Fail-Closed Design: Actions blocked by default unless approval can be cryptographically proven

The Four Tests

Test	Requirement	Enforced Through
STOP	System can be halted before side-effects	Effect-token issuance gated by approval
OWNERSHIP	Identified authority signs policy before execution	Cryptographic signatures with timestamp ordering
REPLAY	Decision can be reproduced at boundary	State-Replay or Protocol-Replay modes
ESCALATION	Mandatory custody transfer on denial/thresholds	Explicit routing with human-in-loop paths

The Enforcement Triad

A governance system that only warns is a monitoring system. Deterministic governance requires the ability to stop actions—not just flag them. Every decision boundary must resolve to one of three verdicts:

Verdict	Meaning	Operational Effect
ALLOW	Action authorized under governing policy	Execution proceeds; effect-token issued
DENY	Policy violation identified	Execution halted; no side-effects permitted
ABSTAIN	Insufficient confidence to render a verdict	Authority explicitly returned; action blocked pending human resolution

Why ABSTAIN matters

ABSTAIN is the system's explicit acknowledgment of its own epistemic boundaries—a controlled handoff of authority rather than an uncertain guess. Systems that always answer are optimizing for completion. Systems that know when not to answer are optimizing for correctness.

Operational contract:

• ABSTAIN triggers mandatory escalation—routing to a human-in-the-loop queue, policy authority review, or other designated escalation path
• In regulated contexts, ABSTAIN is fail-closed: the action does not proceed unless and until an authorized party renders a definitive verdict
• ABSTAIN is not a soft "maybe"; it is a hard gate that transfers decision authority while preventing unauthorized execution
• Default behavior: ABSTAIN MUST be treated as DENY unless an authorized override occurs

Quick Start

Enable auditors to verify AI compliance mathematically rather than through sampling—deterministic verification at decision boundaries.

Installation

# Clone the repository
git clone https://github.com/edmeyman/4ts-standard.git
cd 4ts-standard

# Install dependencies (Python 3.10+)
pip install -r requirements.txt

Validate Your First PCD

# Validate against test vectors
python tools/validator/quickstart_validate.py --json examples/model-deployment-state-replay.json

# Expected output: PASS

Create Your First PCD

from tools import pcd_builder

pcd = pcd_builder.create_pcd(
    boundary="deploy",
    artifacts={"models": [{"id": "my-model-v1.0", "sha256": "..."}]},
    replay_strategy="state"
)

print(pcd.to_json())

Repository Structure

4ts-standard/
├── README.md                    # This file
├── SPECIFICATION.md             # Complete technical specification (§§0-11)
├── LICENSE.md                   # Dual license (CC BY-NC-ND 4.0 + MIT)
├── CHANGELOG.md                 # Version history
├── CONTRIBUTING.md              # How to contribute
│
├── schemas/                     # JSON Schema definitions
│   ├── pcd.schema.json         # PCD structure (draft 2020-12)
│   └── verifier.config.schema.json  # Verifier configuration
│
├── examples/                    # Concrete PCD examples
│   ├── model-deployment-state-replay.json
│   ├── rag-system-protocol-replay.json
│   ├── llm-tool-inference.json
│   └── etl-pipeline-batch.json
│
├── test-vectors/                # Conformance test suite
│   ├── positive/               # Must PASS (8 vectors)
│   │   ├── PCD-A1_state_auto_approve.json
│   │   ├── PCD-A2_protocol_with_gates.json
│   │   └── PCD-A3_fail_closed_denial.json
│   └── negative/               # Must FAIL with specific errors (5 vectors)
│       ├── NC-1_posthoc_signature.json
│       ├── NC-2_missing_custody.json
│       ├── NC-3_untyped_lineage.json
│       ├── NC-4_side_effect_on_denial.json
│       └── NC-5_protocol_gate_fail.json
│
├── tools/                       # Reference implementations
│   ├── validator/              # Python reference verifier
│   │   ├── quickstart_validate.py
│   │   └── verifier.py
│   └── canonicalizer/          # JSON canonicalization
│       └── canonicalize.py
│
└── docs/                        # Additional documentation
    ├── quickstart.md           # Getting started guide
    ├── implementation-guide.md # Detailed implementation patterns
    ├── faq.md                  # Frequently asked questions
    └── error-catalog.md        # Complete error code reference

Use Cases

4TS supports diverse AI deployment patterns:

Profile	PCD Emission	Replay Mode	Key Considerations
LLM Tools	Per tool action with external effects	State or Protocol	Typed lineage for tool I/O
RAG Systems	Per response triggering workflows	Protocol (frozen index)	Gates on answerability/attribution
Model Deployment	At deployment and policy changes	State or Protocol (eval gates)	Pre-exec policy signature required
BPMN/ETL	Per job with external writes	State or Protocol	Compensating actions for rollbacks
Agentic Systems	Per plan execution	Protocol with explicit gates	Sub-PCDs for high-risk steps

Conformance

To claim 4TS conformance, implementers must:

1. Pass all test vectors: 3 positive (PASS), 5 negative (expected failures with correct error codes)
2. Publish conformance claim:

   Tool@Version • PCD-1 • Bundle-1.0.2 • 8/8 • sha256:manifest_hash • logs_link
   

3. Implement core verification: PCD schema validation, signature verification, replay logic, fail-closed enforcement

See SPECIFICATION.md §7 for complete conformance requirements.

Implementation Profiles

Healthcare/Life Sciences

• Boundary: Model deployment for diagnostic/treatment decisions
• Replay: State-Replay with frozen training/test sets
• Gates: AUROC, calibration error, demographic parity
• Regulatory: FDA 510(k), EU MDR alignment

Financial Services

• Boundary: Inference-level for credit/trading decisions
• Replay: Protocol-Replay with deterministic risk metrics
• Gates: Accuracy, fairness (disparate impact), attribution
• Regulatory: SR 11-7, ECOA, MiFID II alignment

Government/Defense

• Boundary: Deploy and policy-change for mission-critical systems
• Replay: State-Replay with air-gapped verification
• Gates: Security clearance checks, operational safety thresholds
• Regulatory: NIST AI RMF, DoD AI principles

Documentation

• Complete Specification - Full technical standard (§§0-11)
• Quick Start Guide - 10-minute implementation tutorial
• Implementation Guide - Detailed patterns and best practices
• FAQ - Common questions and answers
• Error Catalog - All error codes with triggers and recovery

Community & Support

• Issues & Bugs: GitHub Issues
• Discussions: GitHub Discussions
• Email: info@ferz.ai
• Website: https://ferz.ai

Contributing

We welcome contributions! Please see CONTRIBUTING.md for:

• Code of conduct
• How to submit issues and pull requests
• Development workflow
• Testing requirements

Related Standards & Documents

• Deterministic AI Governance - Executive Guide - Business rationale and minimum governance bar
• JSON Schema draft 2020-12 - Schema specification standard
• RFC 3339 (ISO 8601) - Timestamp format
• SHA-256 (FIPS 180-4) - Cryptographic hashing
• EdDSA/ECDSA - Digital signature algorithms

Version History

Version	Date	Changes
1.0.0	2025-09	Initial release with core standard
1.0.1	2025-10	Added adoption profiles, clarified gate format, expanded error codes
1.0.2	2025-11	Added traceability matrix, glossary, concrete PCD examples
1.0.3	2026-02	Added Enforcement Triad (ALLOW/DENY/ABSTAIN) with ABSTAIN operational contract

See CHANGELOG.md for detailed changes.

License

• Specification Text: CC BY-NC-ND 4.0

  • Attribution required
  • Non-commercial use
  • No derivatives

• Schemas & Test Vectors: MIT License

  • Permissive use for implementation

See LICENSE.md for complete terms.

For commercial licensing inquiries: info@ferz.ai

---

© 2026 FERZ Inc. | Vendor-neutral open standard for verifiable AI governance