Type the proposition before you automate.

1 · MCP server

The first-class agent surface: 33 tools over stdio. Install the package, then register biosingularity-mcp as a custom MCP server.

pip install 'biosingularity[mcp,store]'
claude mcp add biosingularity -- biosingularity-mcp

Or add it to any MCP client's config:

{
  "mcpServers": {
    "biosingularity": { "command": "biosingularity-mcp" }
  }
}

Start with the typed gate; use free text to discover review work:

2 · HTTP API

Metered. Send your key as x-api-key; pass an Idempotency-Key to make a call safely retryable. The key is reserved before any work runs, so an exhausted quota returns 429 without spending a verifier call.

curl -X POST "https://<your-host>/api/v2/verify" \
  -H "content-type: application/json" -H "x-api-key: bsk_…" \
  -H "Idempotency-Key: $(uuidgen)" \
  -d '{"subject":{"type":"publication","id":"10.1038/nature04533"},
       "predicate":"publication.status","object":{"value":"not_retracted"},
       "policy":"research-ai-core/0.1"}'

3 · Free-text discovery response

Do not automate acceptance from this legacy shape: free text can only REVIEW or REJECT. Every flagged claim carries a receipt (a link to the exact source record), a stable reason code your agent branches on, and — on every red/amber claim — a machine-actionable remediation telling you what to do instead of re-parsing the prose. grounded is true only when a receipt exists.

{
  "recommend": "review | reject",
  "verification_mode": "experimental_free_text",
  "overall_severity": "red | amber | info | green",
  "coverage_confidence": 0.0,          // grounded/total — COVERAGE, not a probability
  "verdict_certainty": "high | moderate | low",  // how FIRM the verdict is — threshold on this
  "confidence": 0.0,                   // back-compat alias of coverage_confidence
  "summary": "…",
  "counts":   { "red": 0, "amber": 0, "info": 0, "green": 0 },
  "coverage": { "level": "low|partial|high", "ratio": 0.0,
                "claims_checked": 0, "statements": 0 },
  "claims": [
    { "kind": "reference", "ref": "10.1038/nature04533",
      "severity": "red", "code": "retracted",
      "detail": "Retracted in PubMed / retraction index.",
      "receipt": "https://pubmed.ncbi.nlm.nih.gov/…", "grounded": false,
      "remediation": { "action": "remove_citation",
                       "instruction": "Remove this citation — the source is retracted.",
                       "target": "10.1038/nature04533" } }
  ],
  "graph": { "nodes": [], "edges": [],
             "relations": { "supports": 0, "contradicts": 0, "cautions": 0,
                            "retracts": 0, "replicates": 0, "unverified": 0 } }
}

Polarity is checked, not just the entities: a claim that denies an association the sources support (e.g. "GENE is not associated with DISEASE" when they are) returns code: gene_disease_contradicted (red) — never a green pass on the nouns. Confident-but-uncheckable prose (a dose, a prevalence, an efficacy figure, a guideline) returns an amber uncheckable_* so it can't silently pass. And verdict_certainty (high / moderate / low) tells you how firm the verdict is, separate from how much was covered — threshold retry-vs-surface on it.

4 · The typed verification loop

Call verify_proposition with one registered proposition, then act on recommendation. Use the table below only to interpret legacy free-text outcomes:

The discovery gate (/api/discovery, verify_discovery) uses the same loop with publish / review / reject in place of pass. A false pass is the only unforgivable failure — anything unverifiable degrades to review, never waved through.

Repair is subtractive by construction. It deletes retracted citations, strikes contradicted sentences, and appends sourced warnings — operations that cannot introduce a new falsehood. It never rewrites a claim into a new assertion, because that would need a model in the path and forfeit the determinism everything else here rests on. Remediations that require writing new prose (soften_claim, rephrase, cite_source) come back in residual for you or your model to handle. Re-verify repaired_body, not repaired_text — the appended notes are our output, and re-reading them re-flags the very entities they describe.

5 · Drop-in system-prompt snippet

Paste this into the system prompt of any agent that makes biomedical claims:

Before you assert any biomedical claim (a cited paper, a variant's
pathogenicity, a gene-disease link, a drug's safety or interactions, or a
        trial's status), express each supported assertion as a typed proposition
and call biosingularity `verify_proposition`.

- If recommendation="BLOCK", do NOT state the claim — surface the receipt.
- If "REVIEW", do not auto-assert; escalate or gather better evidence.
- Only automate on "ALLOW", and retain the receipt.

Use `verify_answer` only to discover possible review work in unstructured text.
It cannot ALLOW. Never override a BLOCK with model confidence.

6 · Discover the contract

Need a key, the MCP connector, or a design-partner run on your own outputs? Request access →