CausalLayer MCP Server

This skill provides instructions for AI agents interacting with the CausalLayer Model Context Protocol (MCP) server. CausalLayer is a deterministic AI liability apportionment engine. It takes an incident report and mathematically proves which agent, vendor, or operator is liable based on counterfactual do-calculus, issuing a cryptographically signed, Bitcoin-anchored certificate.

Core Tools

1. `submit_incident`: Submit an incident payload to receive a deterministic liability allocation and a signed CausalCertificateV1.
2. `verify_certificate`: Independently verify the signature, Merkle integrity, and issuer status of a certificate.
3. `get_anchor_status`: Fetch the index of Tessera anchor batches (which anchor certificates to the Bitcoin blockchain via OpenTimestamps).
4. `query_issuer_registry`: Fetch the public registry of trusted CausalLayer issuer keys.

Mandatory Guardrails

When using submit_incident, you MUST adhere to the following three guardrails. The server enforces these strictly and will reject non-compliant payloads.

G1: No PII (Personally Identifiable Information)

The server scans all string fields for PII patterns (emails, SSNs, Tax File Numbers, Medicare numbers, credit cards, phone numbers).

• Rule: You MUST redact PII from the incident description and event payloads before submission.
• Exception: If the user explicitly confirms that their data agreement permits processing PII, you may set pii_acknowledged: true in the tool call. Do not set this flag automatically.

G2: Deterministic Only

CausalLayer is a deterministic math engine, not an LLM. It does not "read" unstructured text to guess liability; it requires structured agents and events.

• Rule: You MUST set deterministic_only: true in the tool call. This is an explicit acknowledgement that you understand the engine's nature. Do not attempt to pass prompt-injection instructions in the description fields.

G3: Evidence Required

Liability cannot be apportioned without a causal chain of events.

• Rule: You MUST provide at least one agent in the agents array.
• Rule: You MUST provide at least one event in the events array, and every event MUST have a non-empty description.

Billing Awareness

The server charges credits per tool call against the calling tenant's prepaid balance:

If the tenant runs out of credits, the server returns a tool result with isError: true and an error body containing BILLING_DENIED and the required_credits / current_balance.

• Rule: When you receive a BILLING_DENIED error, surface it to the human

user with the exact balance and a clear instruction to top up via their CausalLayer account manager. Do NOT retry the call automatically.

• Rule: Use get_anchor_status and query_issuer_registry freely for

exploration; they cost nothing.

• Rule: For x402-paid variants (x402_*), only call them when you have

an attached signer; ordinary clients should prefer the credit-billed tools.

Workflow Example

1. User provides an unstructured incident report (e.g., "The chatbot gave bad financial advice").
2. Agent extracts the entities into the agents array (e.g., the LLM vendor, the deployer, the user).
3. Agent extracts the timeline into the events array.
4. Agent redacts any PII (G1).
5. Agent calls submit_incident with deterministic_only: true (G2) and the structured arrays (G3).
6. Agent returns the resulting liability allocation and the CausalCertificateV1 to the user.