# FaultKey · CausalLayer

> Deterministic AI-liability attribution. Every AI incident → a signed, Bitcoin-anchored CausalCertificateV1 receipt with a vendor / deployer / user fault split. Closed-form scoring (Ed25519 + Merkle + OpenTimestamps), byte-identical reproducibility, no LLMs in the scoring path.

FaultKey/CausalLayer is the first production-ready, deterministic AI-liability attribution layer for Anthropic's Model Context Protocol (MCP). It targets AI-insurance underwriters, AI-incident response teams, and any party subject to APRA CPS 230, the EU AI Act (Article 12 logging), ISO/IEC 42001, or the NIST AI Risk Management Framework.

## Core endpoints

- Live MCP endpoint (Streamable HTTP): https://mcp.faultkey.com/mcp
- Health check: https://mcp.faultkey.com/health
- Public stats: https://mcp.faultkey.com/stats
- Server card: https://mcp.faultkey.com/.well-known/mcp/server-card.json
- Source (Apache-2.0): https://github.com/smq9sn5jck-coder/causallayer-mcp
- OpenAPI spec: https://github.com/smq9sn5jck-coder/causallayer-mcp/blob/main/openapi.yaml

## Verifiable predictions (Anchor Log)

FaultKey publishes a public, cryptographically sealed log of deterministic predictions on live, unresolved AI-liability cases. Each entry was committed *before* the case resolved, so the engine's calibration can be audited after the fact without us being able to revise history.

- Anchor Log page: https://faultkey.com/anchor-log
- Manifest (8 sealed predictions, SHA-256 of each artifact): https://faultkey.com/anchor-log/manifest.json
- Aggregate root payload (anchored to Sigstore Rekor + Bitcoin OpenTimestamps): https://faultkey.com/anchor-log/anchor-payload.json
- Aggregate root SHA-256: 2d003092…0a2f (anchored to Rekor log index 1592642535)
- Cases tracked: Moffatt v. Air Canada (calibration anchor, resolved), Mobley v. Workday, Raine v. OpenAI, Garcia v. Character.AI, PA AG v. Character.AI, Lokken v. UnitedHealth, Walters v. OpenAI, Tesla Autopilot MDL

## Public algorithm spec

The full factor-weight specification of the deterministic scoring engine — closed-form formula, event-type severity table, role-duty multipliers, but-for handling, and a worked example — is published inline at https://faultkey.com/engine#appendix-a. A byte-identical Node CLI of the same engine ships in the open repo at https://github.com/smq9sn5jck-coder/causallayer-mcp/blob/main/demo/score.js (`node demo/score.js --scenario loan`).

## Tools (MCP)

- submit_incident — submit AI incident, return signed CausalCertificateV1 with fault split (50 credits)
- verify_certificate — verify Ed25519 signature + Merkle path + OpenTimestamps anchor (1 credit)
- get_anchor_status — latest Bitcoin anchor batch status (free)
- query_issuer_registry — public Ed25519 issuer key lookup (free)

## Quick Start (Claude Desktop / Cursor / Cline)

```json
{
  "mcpServers": {
    "faultkey": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.faultkey.com/mcp"]
    }
  }
}
```

## Example Usage

Ask your AI assistant:
- "Submit an AI incident where a healthcare chatbot gave wrong dosage advice"
- "What's the liability split if an autonomous trading bot loses $4M?"
- "Verify this FaultKey certificate"

The engine returns a full CausalCertificateV1 with:
- Deterministic liability percentages (vendor / deployer / user)
- 16-module scoring pipeline (opacity, autonomy, guardrail_bypass, severity, foreseeability, etc.)
- Causal graph with but-for chain analysis
- Regulatory mapping (APRA CPS 230, EU AI Act, NIST AI RMF, ISO 42001)
- Damages estimation with actuarial pricing
- Stress test results across 8 adversarial scenarios
- Discovery/subpoena checklist for litigation

## Why deterministic matters

Every other AI-incident tool today uses LLMs in the scoring path, which makes the score itself non-deterministic and therefore inadmissible as primary evidence in audit, insurance, or court contexts. CausalLayer separates the deterministic scoring engine from any LLM helper, producing byte-identical outputs anyone with the same inputs can reproduce.

## Compliance fit

- APRA CPS 230 — operational risk evidence trail
- EU AI Act Article 12 — automatic logging requirement
- EU AI Act Article 26 — deployer obligation documentation
- ISO/IEC 42001 — AI management system audit
- NIST AI RMF — measure & manage functions
- UK AI Safety Framework — sector-specific compliance

## Supported Industries

Healthcare, Financial Services, Legal, Autonomous Vehicles, Cybersecurity, Insurance, Government, Education, HR/Recruitment

## Full documentation

See /llms-full.txt for complete API schema, all tool parameters, and response formats.

## Citation

If you cite this in research, see CITATION.cff in the repository root.

## Disclaimer (voluntary, non-binding, NIST AI RMF MAP 4 → MEASURE 2.7 → MANAGE 1.3)

Every FaultKey output — score, certificate, anchor entry, prediction artifact — is voluntary, non-binding, and does not constitute legal advice or a legal finding of liability, fault, negligence, or breach. Attribution percentages are deterministic mathematical estimates produced by a published, byte-reproducible methodology operating on the inputs supplied to it; they are not adjudicative findings and are not warranted as fit for any particular purpose, including but not limited to litigation, insurance underwriting, regulatory enforcement, or employment decisions. Any party affected by an output may submit a calibration challenge under the public procedure published at https://faultkey.com/working-group . Full disclaimer JSON: https://faultkey.com/anchor-log/disclaimer.json .

## Bifurcated governance (in formation)

Authoring entity (rule-sets, methodology, prediction artifacts, source register): FaultKey Working Group, Inc. — Delaware non-stock nonprofit in formation; intended IRC §501(c)(3) public charity; charter prohibits payment from any party named as an attributed party in any prediction artifact it has published. Commercial entity (hosted MCP API, integration, indemnity): FaultKey OpCo — for-profit, jurisdiction TBD; non-exclusive trademark license; cannot modify, override, or selectively apply any rule-set; cannot sell differential access to outputs. Doctrinal hooks: Anderson v. Stability AI (N.D. Cal. 2023); Abu Dhabi Commercial Bank v. Morgan Stanley (S.D.N.Y. 2009). Full bylaw skeleton at https://faultkey.com/working-group .