Groundcheck

The grounding check agents run before they commit to an answer.

Groundcheck verifies a factual claim against live sources and returns a verdict, a confidence score, and citations. Any agent — Claude Code, Cursor, your own — can call it mid-task, before it states a fact it isn't sure of.

It is also a verification layer for agentic commerce: when an agent pays another service over x402, attest_delivery verifies what was delivered against what was advertised and issues a signed, offline-verifiable delivery receipt binding payment → delivery → grounded content — the neutral accountability trail the a2a-payments literature calls the missing layer (docs/delivery-attestation.md).

Architecture

Two parts, each in the language that fits it:

server/   TypeScript MCP server   — thin protocol layer (stdio). Holds no logic.
engine/   Python FastAPI service  — retrieval + stance classification + the verdict brain.

The MCP server is spawned by your client over stdio and talks to the engine over HTTP (GROUNDCHECK_ENGINE_URL, default http://127.0.0.1:8723). The engine is the single source of truth for how a verdict is reached, and it classifies source stance through the canonical Python free-llm-router (free-tier providers).

verify_claim ─▶ TS MCP server ─HTTP▶ Python engine
                                        ├─ retrieval  (Wikipedia, keyless; or your own search)
                                        ├─ stance     (free-llm-router → supports/refutes/neutral)
                                        └─ verdict    (refuses on conflict, saturating confidence)

Tools

verdict is one of supported · refuted · unverified. Each verdict also carries a sufficiency tag (sufficient · insufficient · no_sources · no_stance · conflict) so an agent can tell "I found nothing" from "sources exist but don't establish it" from "sources disagree" — the three ways an abstention happens carry different meaning and are no longer collapsed (SURE-RAG).

Compound claims are decomposed. A claim like "Marie Curie won two Nobel Prizes and was born in Paris" is split into atoms (Fact in Fragments), each verified on its own evidence and recombined weakest-link: one false part refutes the whole, one unproven part blocks a supported. The true half can no longer carry the false half past the check. The atom breakdown is returned in atoms. (Decomposition is rule-based and high-precision — it splits only on clean conjunction boundaries and otherwise leaves the claim whole; disable with GROUNDCHECK_DECOMPOSE=0.)

Remote MCP (no install): add https://groundcheck.seiche.info/mcp as a remote MCP server (Claude/ChatGPT/Cursor connectors, or a gateway like Smithery/Glama). Speaks streamable-HTTP JSON-RPC; verify_claim is free, the paid tools answer HTTP 402 with an x402 offer.

Quickstart

The MCP server auto-starts the Python engine if one isn't already running, so a single registration is enough — no separate process to babysit.

make install                      # deps for both halves (pip + npm)
npm --prefix server run build     # compile the server
export GROQ_API_KEY="gsk_..."     # one free key for stance classification (Groq: ~2 min, 14,400/day)

# register with your MCP client — the engine spawns on first use and stops with the server
claude mcp add groundcheck -- node "$PWD/server/dist/server.js"

Already running the engine yourself (make engine or docker compose up -d)? The server detects and reuses it — and won't touch an engine it didn't start. Set GROUNDCHECK_NO_SPAWN=1 to stop it from ever spawning one.

Once published to npm, registration becomes claude mcp add groundcheck -- npx -y groundcheck-mcp. Auto-spawn needs a local engine/ + Python deps; for an npx-only install, run the engine via docker compose up -d and the server connects to it over GROUNDCHECK_ENGINE_URL.

With no provider key the engine still runs — retrieval works, but every verdict is unverified. It degrades honestly: a disabled backend, a missing key, or conflicting sources all flow toward unverified. An unconfigured Groundcheck cannot return supported.

Note: OpenRouter's :free models are quota-throttled (HTTP 429) and make a poor sole provider. Prefer Groq or Cerebras for the fast classification tier.

Why grounded verdicts, not LLM-judgment

Asking an LLM to judge whether a claim is true is unreliable in a way that's easy to miss. In TraderBench (Yuan et al., 2026), the same candidate responses re-scored by three frontier LLM judges swung by ~29 points on the knowledge-retrieval section — while the performance-grounded section, whose scoring is anchored to verifiable computation, swung 0.3. The lesson: the more you constrain a judgment with external evidence, the less it varies.

Groundcheck is built on that principle. It never asks a model "is this true?" from parametric memory. Instead it:

• retrieves sources first, then asks only the narrow, evidence-anchored question — does this cited passage support, refute, or stay neutral on the claim (stance classification);
• refuses on conflict and saturates confidence, so disagreement flows to unverified rather than a confident guess;
• returns citations, so the verdict is checkable, not taken on the model's word.

That's the difference between an LLM judge and a grounding check: the judge's discretion is the product; here it's deliberately fenced in by retrieved evidence.

Calibrated verdicts: the "error ≤ α" guarantee

A confidence number without a promise attached is just vibes with decimals. When a calibration artifact is deployed, Groundcheck attaches a guarantee object to directional verdicts, built with split conformal prediction (adapted from Multi-LLM Adaptive Conformal Inference, arXiv:2602.01285):

• Stance classification runs as a panel: up to GROUNDCHECK_ENSEMBLE_MAX free providers judge the claim independently (different model families disagree on which claims they get wrong, so the ensemble beats any one of them). Per-source stances are majority-voted; each panelist also emits a probability the claim is true given only the snippets, combined into a weighted ensemble_score.
• scripts/calibrate.py runs the real pipeline over a labeled claim set and stores finite-sample thresholds per claim group (instrument / general, global fallback) in calibration/calibration.json.
• A verdict is certified (guarantee.certified: true) only when its score clears the threshold. The math guarantees that, for claims exchangeable with the calibration set, a false claim is certified supported with probability ≤ α (default 0.1), and symmetrically for refuted. No distributional assumptions, exact in finite samples.

Honest degradation, as everywhere else: no artifact → no guarantee is ever claimed; too little calibration data for a given α → the threshold is refused rather than extrapolated. The guarantee is only as good as the exchangeability assumption — recalibrate with domain claims before leaning on it in a new domain.

Configuration (engine)

Machine-payable hosting (x402)