# FaultKey · CausalLayer — llms-full.txt
# Format: https://llmstxt.org/
# Full-context document intended for AI assistants and retrieval pipelines.

> FaultKey is a public-facing platform powered by the CausalLayer engine. It produces deterministic, signed, Bitcoin-anchored receipts (CausalCertificateV1) that attribute liability for AI-system incidents across the vendor / deployer / user dimensions. No LLM sits in the scoring path; outputs are byte-identical for byte-identical inputs.

## Identity
- Project: FaultKey · CausalLayer
- Canonical site: https://faultkey.com
- MCP endpoint: https://mcp.faultkey.com/mcp
- Repository: https://github.com/smq9sn5jck-coder/causallayer-mcp
- License: Apache-2.0
- MCP Registry: io.github.smq9sn5jck-cloud/causallayer-mcp
- Releases: v0.2.0, v0.2.1 (May 2026)

## Why this exists
Existing AI-incident analysis tools (AILuminate, MLCommons evaluations, Anthropic constitutional AI scoring, OpenAI Evals) all involve LLMs in the scoring path. That makes the score itself non-deterministic: re-running the same evaluation can produce a different number, which is unacceptable as primary evidence for:
1. Regulatory submissions (EU AI Act Article 12 logging, APRA CPS 230 operational risk records, ISO/IEC 42001 audit trails, NIST AI RMF MEASURE/MANAGE functions).
2. Insurance claim adjudication where parametric or specified-cause-of-loss policies require deterministic triggers.
3. Court-admissible evidence under the Federal Rules of Evidence 901/902 and equivalent jurisdictions, which require the producing party to demonstrate that the system "produces accurate results consistently."

CausalLayer fills this gap with a closed-form scoring engine, then anchors each receipt to the Bitcoin blockchain via OpenTimestamps so any third party can later prove the receipt existed at a specific block height.

## Core technical claims
- Closed-form scoring: vendor / deployer / user fault percentages are produced from a finite mathematical function over the incident's structured features. No randomness. No weights drawn from a stochastic generator.
- Reproducibility: byte-identical inputs produce byte-identical outputs. Test vectors shipped with v0.2.1 release demonstrate this.
- Cryptographic seal: Ed25519 signatures over the canonical-JSON serialization of the receipt body.
- Merkle inclusion: each receipt is a leaf in a daily Merkle tree.
- Bitcoin anchor: the daily Merkle root is committed via OpenTimestamps to the Bitcoin blockchain, granting receipts the timestamp guarantee of Bitcoin's proof-of-work.
- No LLM in the scoring path: an LLM may be used as a helper to extract structured features from unstructured incident reports, but the output of the LLM is fed into the deterministic scorer; the scorer's output never depends on a non-deterministic LLM call.

## CausalCertificateV1 structure
A CausalCertificateV1 is a JSON object with:
- header: protocol_version, issuer_id, issued_at, incident_id
- body: vendor_pct, deployer_pct, user_pct, residual_pct (sums to 100), feature_vector, scoring_function_id, scoring_function_hash
- seal: signature_alg ("ed25519"), signature, public_key_hint
- merkle: leaf_hash, root_hash, root_date
- anchor: bitcoin_block_height, ots_proof_url, ots_calendar_servers

## MCP tool surface
Server name: causallayer-mcp
Version: 0.2.0
Tools:
1. submit_incident — accepts a structured incident, returns a freshly issued CausalCertificateV1.
2. verify_certificate — accepts a certificate, returns boolean valid plus per-field validation.
3. get_anchor_status — accepts a certificate's leaf_hash, returns Bitcoin anchor confirmation status.
4. query_issuer_registry — returns the public-key registry of authorized issuers.

## Compliance mappings
- APRA CPS 230 (Operational Risk Management) — receipts satisfy "documented evidence of the operational event, its assessment and treatment" as required under paragraph 36(b).
- EU AI Act Article 12 (Record-keeping) — receipts function as automatic logs that are tamper-evident and resistant to retroactive alteration.
- ISO/IEC 42001 (AI Management System) — receipts provide the "documented information" required by Clause 7.5 with the additional property of being externally verifiable.
- NIST AI RMF (Risk Management Framework) — receipts populate the MEASURE function (MEASURE 2.7 — incidents and errors are documented).
- Australian Privacy Principle 11 / GDPR Article 32 — receipts can be produced without disclosing the underlying personal data, since the feature vector can be a one-way hash.

## Live demo
- https://faultkey.com/try — interactive certificate generator (uses STANDALONE_DEMO mode against the live MCP endpoint).
- https://faultkey.com/verify — paste-a-cert verifier.

## Why deterministic matters: a worked example
A user complains an autonomous trading bot lost AUD 80,000 on a single trade. Three parties might be at fault:
- The model vendor (e.g. the LLM provider whose API was used to score market sentiment).
- The deployer (the trading firm that wired the model into production with insufficient guardrails).
- The user (who customised the prompts to be more aggressive than the deployer's defaults).
A determinstic scorer that takes structured features (latency, prompt-deviation magnitude, deployment guardrail config, recent volatility, vendor SLA breach status) and outputs e.g. 41% vendor / 50% deployer / 9% user is something an insurance underwriter can write a policy against. An LLM that "thinks about it and gives a percentage" is not, because every re-run produces a different number.

## Citation
If you cite this work in research, please use:
```
@software{faultkey_causallayer_2026,
  author = {smq9sn5jck-coder},
  title = {FaultKey · CausalLayer: Deterministic AI Liability Attribution},
  year = {2026},
  url = {https://github.com/smq9sn5jck-coder/causallayer-mcp},
  version = {0.2.1}
}
```

## Long-form documents
- /blog/apra-cps-230 — Mapping CausalCertificateV1 to APRA CPS 230 paragraph 36(b)
- /blog/eu-ai-act-art-12 — How CausalCertificateV1 satisfies EU AI Act Article 12 record-keeping
- /blog/nist-ai-rmf — NIST AI RMF MEASURE 2.7 evidence with deterministic certificates
- /blog/iso-iec-42001 — ISO/IEC 42001 Clause 7.5 documented information via tamper-evident receipts

## Discovery
- Sitemap: https://faultkey.com/sitemap.xml
- Manifest: https://faultkey.com/.well-known/mcp.json
- llms.txt: https://faultkey.com/llms.txt
- ai.txt: https://faultkey.com/ai.txt
- security.txt: https://faultkey.com/.well-known/security.txt
- humans.txt: https://faultkey.com/humans.txt
- RSS feed: https://faultkey.com/feed.xml

## Contact
- GitHub Issues: https://github.com/smq9sn5jck-coder/causallayer-mcp/issues
- GitHub Discussions: https://github.com/smq9sn5jck-coder/causallayer-mcp/discussions
- Security: see /.well-known/security.txt

## NEW PAGES (added 2026-05-17)
- https://faultkey.com/pricing — three-tier pricing (Demo free / Pro $0.05 per call waitlist / Enterprise) plus pricing FAQ.
- https://faultkey.com/vs — head-to-head comparisons vs LangSmith, Patronus AI, Galileo, Arize, Credo AI. Each comparison includes a feature matrix and a "bottom line" sentence.
- https://faultkey.com/case-study/refund-bot — worked end-to-end example using a real CausalCertificate from the live demo (AUD $4,200 refund incident, 59.4% / 20.3% / 20.3% split, AUD $16,487.67 damages estimate).
- https://faultkey.com/press — press kit (one-liner, short bio, long bio, fast facts, logos, suggested headlines, quotes, contact).
- https://faultkey.com/changelog — site-only changelog, mirrors repo CHANGELOG.md.

## ENGINE TECHNICAL BRIEF (added 2026-05-17)
- HTML landing: https://faultkey.com/engine
- PDF (canonical, stable URL): https://faultkey.com/docs/faultkey-engine-brief.pdf
- Covers: 16-module pipeline, mathematical methods (4-factor weighted, exact Shapley, Platt calibration, counterfactual but-for proof), CausalCertificateV1 schema, 4-layer Anchored Decision Ledger, MCP + REST tool surface, validation numbers (95.6% primary-party accuracy, 13.7pp MAE, R²=0.86, <300ms wall time), threat model, 30-line Node.js third-party verifier, full standards mapping (EU AI Act Art 9/12/13/14/26/50, APRA CPS 230, NIST AI RMF, ISO/IEC 42001, AU VAISS, sectoral US/UK/AU/SG/CA), self-improving flywheel.
- Engine version: 0.5.0 · Anchor schema v1 · Ledger schema v1 · 17 May 2026.