LabscoMCP early access ↗
smq9sn5jck-coderMCP SERVER · OFFICIAL PROJECT

FaultKey CausalLayer

Deterministic fault allocation for multi-party AI incidents — a signed, independently verifiable certificate, no LLM in the scoring.

by smq9sn5jck-coder ↗Regulatory Compliance & AI Governance✓ Official

LABSCO SUMMARY

Deterministic on purpose, because the buyers cannot accept anything else.

The design constraint is stated rather than implied: insurers and regulated entities cannot use a fault attribution that gives different answers to the same prompt on different days, so the scoring deliberately contains no model. What makes that credible is the verification path — a third party can check the signature, the Merkle integrity and the issuer registry with a separate open tool and no network call back to the vendor, which is the opposite of trusting a score because it came with a logo.

What it is

The MCP server for the CausalLayer engine, packaged as a Cloudflare Worker. When an AI causes harm and several parties dispute who pays, it returns a signed certificate of fault allocation computed by a closed-form causal scoring algorithm — graph-theoretic, version-pinned, and byte-identical across runs, so two adversarial parties get the same number.

What you get

  • submit_incident — submit an AI incident for liability attribution and get back a signed CausalCertificateV1 with per-agent fault allocation, evidence-chain completeness and Bitcoin-anchored proof
  • verify_certificate — check a certificate's signature, Merkle integrity and issuer status independently, without calling FaultKey at all
  • get_anchor_status — the index of anchor batches, or one batch's full JSON with its signed Merkle root and OpenTimestamps proof reference
  • query_issuer_registry — the trusted issuer public-key fingerprints, their status and validity windows
  • Verification that does not depend on the vendor: the certificate, issuer registry and Merkle anchor log can all be checked by a third party using Node's built-in crypto and the separate causallayer-verifier tool
  • Three guardrails enforced at the edge: payloads are scanned for personal identifiers and rejected unless pii_acknowledged is set; every request must carry deterministic_only: true as the caller's acknowledgement that the output is closed-form rather than probabilistic; and at least one identified agent plus one timestamped, described event must be supplied or the request is refused

Requirements

No account to try it: the public Worker at https://mcp.faultkey.com/mcp runs in standalone demo mode with deterministic, watermarked responses. Point a client at npx -y causallayer-mcp, or call the endpoint over HTTP directly — the response carries an Mcp-Session-Id header to reuse on later calls. Transport is Streamable HTTP. Demo limits are per IP per day and differ by tool: 5 incidents, 50 verifications, unlimited anchor and registry lookups. Paid use is credit-metered, with verification at 1 credit and anchor and registry lookups free. Self-hosting to your own Cloudflare account is documented, and setting a standalone-demo flag short-circuits the upstream engine for demos without it. Apache 2.0 licensed.

Setup effort

One command — npx -y causallayer-mcp

See installation options