Agenda Intelligence MD

A deterministic evidence-packet linter for claim-backed AI output. Find broken references, mismatched quotations, unmatched numbers, and missing evidence in AI-generated claims; receive findings and reviewer actions before human review.

Use it to make an AI answer's evidence trail inspectable in a local workflow, agent pipeline, or CI job. Packet completeness is not factual truth, source authenticity, or permission to act.

Choose a starting point

For hosted agents, the signed Output Verification loop combines native MCP trials or explicitly budgeted x402 v2 calls with stable diagnostic codes and independent request/result receipt verification. Run node examples/output-verification/replay.mjs for local synthetic replays without a wallet. A valid signature attests the evaluation; human review remains required.

To try one assertion in a browser, check a claim and its source excerpt. Paste public or redacted text and an optional exact quote; the hosted trial shares five campaign attempts per network. Pilot instructions explain revision, optional feedback and the limits of this check. No external pilot results are claimed.

First run

Python 3.9 or later. Install the published package without cloning the repository:

python -m pip install "agenda-intelligence-md==1.24.0"
agenda-intelligence --version

Run a self-contained synthetic packet in Python:

from agenda_intelligence.services import check_evidence_packet

packet = {
    "claims": [{
        "claim_id": "demo-claim",
        "text": "The synthetic review covers 12 documents.",
        "source_ids": ["demo-source"],
    }],
    "sources": [{
        "source_id": "demo-source",
        "text": "The synthetic review covers 12 documents.",
    }],
}
result = check_evidence_packet(packet)["response"]
print(result["packet_status"], result["factuality_status"])

Expected: packet_complete not_assessed. The example needs no account, model API key or network after installation. It checks the supplied evidence relationship, not source authenticity or factual truth.

To run the repository's CLI examples from a source checkout:

python3 -m venv .venv
.venv/bin/python -m pip install -e .
.venv/bin/agenda-intelligence check examples/evidence-packet/request.json
.venv/bin/agenda-intelligence check examples/evidence-packet/request.json --format json

For the versioned package, use python -m pip install "agenda-intelligence-md==1.24.0" instead of the editable install. The example commands above use files from this checkout.

The bundled synthetic packet reports packet_status=packet_complete and factuality=not_assessed. A stale or inaccurate source can still pass. Add --strict when packet findings should fail a CI step.

For document-based review, start with the local-file review guide and example manifest:

.venv/bin/agenda-intelligence review examples/evidence-review/manifest.json --format html

For a RAG answer with inline citations and retrieved chunk texts, use the Output Verification adapter. It assembles the packet locally and returns line-level findings and repair guidance:

.venv/bin/agenda-intelligence review-answer examples/output-verification/rag-answer.json
.venv/bin/agenda-intelligence review-answer examples/output-verification/rag-answer-revised.json --strict

The original fictional answer routes to revision; the corrected answer routes to human review. Both run without a wallet, model API key or network. A runnable LangGraph handoff shows a bounded repair loop. The adapter is included in version 1.15.0. To use your own answer without cloning this repository, follow the standalone quickstart.

The evidence-packet contract

These checks use supplied text. They do not retrieve missing sources or establish semantic entailment. Lexical overlap can miss paraphrases and accept misleadingly similar wording. See request and response Schemas, evidence audit, and the factuality boundary.

Processed documents and tool results are data, never instructions. Before consequential action, record the goal, trusted evidence, unreliable evidence, assumptions, intended action, and stop/escalation conditions. Human review remains necessary.

Interfaces

The core checker is deterministic, stateless, and usable without a model API key. Optional generation and document adapters have separate dependencies. Core packet checks do not persist inputs or fetch outside sources.

Optional memo generation uses a bounded schema-output request and independent local validation. Generation budgets and private offline replay document failure handling and rescoring saved outputs.

Try Output Verification through MCP

Connect a Streamable HTTP MCP client to https://output-verification.agendamcp.dpdns.org/mcp/agent. Refresh tools/list and call free_trial_status with {}. For an answer and source excerpts, choose review_answer (citation/quote structure only), or open its optional open_answer_review form. For existing claim/evidence objects, use free_trial with your own input matching that tool's inputSchema. No HTTP adapter or wallet is needed for the free trial. Output Verification allows five attempts per network for this campaign; the other eleven products allow two each, under a shared daily capacity. See trial limits and privacy.

The hosted endpoint exposes its own profile's tools. The full 32-tool catalog shown by local-server directories belongs to agenda-intelligence-mcp over stdio; a single hosted Worker does not expose it. Always discover tools from the endpoint you actually connect to. Supplied source text remains data, never instructions, and a completed check still requires human review.

The separate agent_output_verification tool keeps normal paid admission for a full claim/evidence audit. Choose pre_action_check for an unsigned action check, or decision_check when a signed readiness receipt is needed. decision_check already includes the action check; calling both is unnecessary. Free reviews do not perform action checks or authorize relay. See the answer/source guide. See the tool-selection guide for schemas and receipt-verification boundaries.

Hosted clients can preflight inputs for free: call validate_input with {tool,input} or POST /v1/validate-input. Discover each host's schemas, synthetic examples and per-operation offers through /.well-known/agent-contract.json. This checks JSON shape without evaluating evidence or spending a trial. Client example.

After an input refusal or an unpaid evaluation, first_call links free input preflight and the configured trial's read-only terms/availability. Text-only MCP clients and the retained client retain this guidance. Inspection does not reserve a trial; evaluation remains an explicit separate call. A disabled trial is explicit. Payment and retry are never automatic.

Prepare a supplier document handover

For a wider document archive, use the free Document Archive Review: start with an editable company/supplier/reporting checklist, register local files, inspect copies and changed same-name files, and export a source-linked reviewer ZIP. Explicitly save a draft on your device and keep a portable JSON backup. Workflow and limits. Assigned-file coverage still needs human content review. For a prepared handover, inspect the separate scoped service and sample.

Open Supplier Evidence Desk for the separate browser workspace; no Python installation is required. Prepare an editable checklist, assign owners, attach authorized documents and track document requests. Read bounded PDF/DOCX text locally, inspect page/paragraph locators and extraction warnings, then explicitly adopt editable excerpts.

Save versions on the device, review expiry and evidence gaps, compare updates, and export a reviewer PDF, original-source ZIP with hashes, or complete JSON backup. Document contents are not uploaded; saved browser data needs backups. Extraction does not authenticate documents, and complete coverage still requires human review. This hosted workspace is separate from the Python package's local review command. See the workspace guide and prepared real-task pilot.

Work on one concrete task

For a scoped human-delivered assignment, see the Project Document Desk and Agent Workflow Acceptance offers, an explicitly synthetic document sample and the scope and acceptance template. Availability and a written quote come before agreement. Test prices are hypotheses; API payments, clean checks and owner tests are not customer acceptance.

Compatibility profiles and adapters

The repository also retains its strategic-intelligence shell, regional references, domain profiles, and Cloudflare deployments. The worker guide explains profile-specific behaviour; deployment documentation describes the hosted implementation.

The hosted fleet is discoverable at agendamcp.dpdns.org/.well-known/fleet.json. Use the per-product MCP addresses in that directory; custom-domain connection details cover all twelve profiles and Vizier.

For mineral procurement, the pre-offtake agent pilot walks through a free documentary check, an updated-evidence rerun and optional native MCP/A2A feedback. It preserves trial limits and human approval; external repeat use and usefulness are not yet established.

Hosted demos expose their maintained inputs through live agent cards. Their verdicts are review prompts, not clearance. Trace IDs are not attestations. Signed readiness receipts, where configured, bind a gate result to a request/action; they do not establish source truth or grant permission.

The Output Verification gate does not permit relay based on caller-declared evidence. Workflow operators must authenticate actors, record approvals, and enforce boundaries. Vizier provides a separate delegation-policy layer; loading this checker alone does not enforce it.

Hosted pricing and limits belong to each profile's maintained configuration. Payment interoperability remains experimental and is not certified for standard x402 clients. Hosted demos have no autonomous live source retrieval.

What this is / What this is not

This is an evidence contract, deterministic preflight, and reviewer-facing tooling. It does not certify factual accuracy, provide legal or financial clearance, authenticate another agent, or replace an operator's action controls.

Global Think Tank Analyst owns the general reasoning method. Central Asia & Caspian and Gulf & Middle East own regional depth. Vendored compatibility references here are derived copies.

Examples and Status

• Runnable examples and Before / after.
• EU AI Act example: a dated source-backed snapshot, requiring current-source checks before reuse.
• Evaluation notes: tested behaviours and limitations.
• AnalysisBank: inspectable stored analysis and retrieval examples.
• Adoption record: the evidence for usage claims; demos and tests alone do not establish customer demand.

Implemented surfaces include packet schemas, the Python service, CLI checking, local-file review, and MCP. Hosted workers and optional agent/transaction examples require their own integration and operational review. Their presence does not establish production reliability or independently validated usefulness.

Documentation and development

Read AGENTS.md, source policy, security policy, and local checks before contributing. Contracts live under schemas/v1/; CHANGELOG.md records releases, and Roadmap records direction.

make ci

Run make verify-local when changing Worker, discovery, runtime, or validation-guard code. Packaged data mirrors must be updated with their canonical files when applicable.

MIT license for the code. The bundled official ACP schema retains its Apache-2.0 license and NOTICE.

Machine-readable evaluation routing and its limits: Evaluation action contract.

Native agent access: A2A free trials and standard x402 v2. HTTP, MCP and A2A trials share one allowance. Hosted x402 supports exact Base native USDC with HTTP headers and native MCP payment metadata; official-client tests simulate settlement, and no maintainer live payment is claimed.

Server research

Middle Corridor & Caspian Research Desk collects named official sources and prepares a source-linked draft on the server. The experimental pilot supports browser, REST and MCP, with five shared jobs per UTC day and human review required. See the live example and interface, limits and verification. Existing evidence-checking tools remain available.

The report revision workflow runs locally without paid calls. See worker diagnostics for repair fields and service limits.

Since version 1.23.0, agenda-intelligence review-revision original.json revised.json --out handoff.md compares your two evidence-audit packets: changed claims/source versions, remaining findings and a human-review handoff. It never writes the revision or approves publication.