BodulskiHal: Evidence-Based Claim Verification for AI Agents

Check the evidence before your agent says “done”.

BodulskiHal is a deterministic MCP server that evaluates structured claims against the evidence you provide. It returns a classification, a decision, a reason and a SHA-256 receipt so your agent can decide whether to proceed, gather more evidence or ask for human review.

Use it to check observations, completion claims and evidence-sensitive agent responses. The verifier contains no LLM. Your application must call the tool and enforce its decision.

Connect your MCP client

This Actor serves Streamable HTTP through Apify Standby mode:

https://bodul--bodulskihal-claim-verifier.apify.actor/mcp

Supply your Apify API token in the HTTP header:

Authorization: Bearer <YOUR_APIFY_TOKEN>

For clients that support this configuration format:

{
  "mcpServers": {
    "bodulskihal": {
      "url": "https://bodul--bodulskihal-claim-verifier.apify.actor/mcp",
      "headers": {
        "Authorization": "Bearer <YOUR_APIFY_TOKEN>"
      }
    }
  }
}

Client configuration varies. Choose a remote MCP connection with Streamable HTTP and custom header support. Keep your real token in your client's secret settings.

Standby mode and the Start button

Connect through the MCP URL above. A normal Start run launches a long-running HTTP server; it does not perform a verification or produce dataset rows on its own. Such a run can end with Timed out when its configured run limit is reached. Use the Server tab for the Standby endpoint and connection details.

Available tools

Tool	Input	Result
verify_claim	One payload with claim, evidence, policy and evaluation_time; optional human_approved	One classification, verdict, reason and receipt for valid input
verify_claims	One payload with claims, evidence_by_claim, policy and evaluation_time; optional human_approvals	Per-claim results and an overall verdict

Send structured evidence, not just a question or a URL. See JSON_CONTRACT.md in the Source tab for the full schema, allowed values and batch format.

Try a single claim

Call verify_claim with these MCP tool arguments:

{
  "payload": {
    "claim": {
      "id": "count-1",
      "text": "Conversation has 30 questions",
      "type": "observation",
      "risk": "low"
    },
    "evidence": [],
    "policy": {
      "version": "3.0",
      "max_age_seconds": 300
    },
    "evaluation_time": "2026-11-03T12:00:00+00:00"
  }
}

With no evidence, the expected result includes:

{
  "classification": "UNKNOWN",
  "verdict": "BLOCK",
  "reason": "insufficient_eligible_evidence"
}

This is an excerpt; valid requests also return a receipt. For a live workflow, set evaluation_time to the current evaluation time and provide actual evidence with its original observation timestamp. A recent supporting observation still needs to pass the applicable policy gates before approval.

Read the result

Classification describes the evidence assessment. Verdict is the decision your application should enforce.

Classification	Meaning
FACT	Direct eligible support; approval still depends on the remaining gates
INFERENCE	Derived or indirect support; blocked from being presented as a verified fact
ESTIMATE	Approximate claim; blocked from being presented as a verified fact
UNKNOWN	Insufficient eligible or authorized evidence
REFUTED	Eligible evidence contradicts the claim
CONFLICT	Contradictory input or conflicting eligible evidence; blocked

Verdict	What your agent should do
APPROVE	Proceed within the evaluated claim and policy scope
BLOCK	Stop the claim or dependent action and address the returned reason
ESCALATE	Obtain the required clarification, independent evidence or human approval

Evidence freshness follows max_age_seconds. Future evidence is rejected by default. Action-completion claims require direct post-action evidence under the default policy. Medium-risk claims require two independent sources by default; high-risk claims require human approval.

Billing

The wrapper charges one claim-verified event per evaluated claim. Batch calls charge per claim, not per request. A valid evaluation is billable even when its verdict is BLOCK or ESCALATE. Invalid input rejected by the verifier wrapper is not charged a claim-verification event.

See the Pricing tab for current rates and applicable platform costs.

Limits and receipts

BodulskiHal evaluates the evidence you submit. It does not browse sources, authenticate external evidence or establish real-world truth independently. A SHA-256 receipt supports integrity checking; it is not proof that the submitted evidence is true.

The MCP transport is stateless. The tool wrapper does not explicitly persist claim payloads; this is not a guarantee about platform logging or retention. Review your hosting and client settings before sending sensitive data.

For schema details, open JSON_CONTRACT.md in Source. For questions or custom policies, open an issue on this Actor.