Agenda Intelligence MD is a deterministic evidence-packet linter for AI-generated content. Its positioning is: before human review, it helps users check whether the chain of evidence claimed by an AI answer is complete, whether references are broken, whether quotations and numbers match up, whether there are assertions lacking lexical support, and whether there are evidence gaps. The official documentation explicitly emphasizes that a complete evidence packet (packet_status=packet_complete) does not mean the facts are true; the factuality status is not_assessed; even if sources are outdated or inaccurate, it may still pass the check. The tool does not certify factual accuracy, does not provide legal or financial clearance, and does not certify other agents or replace operators' control over actions.
In terms of use, the project is used to make the evidence trail of AI answers inspectable in local workflows, agent pipelines, or CI tasks. It is described as an "evidence contract, deterministic preflight, and reviewer-facing tool," rather than a fact-finding or compliance-clearance tool. The project also includes a strategic intelligence shell, regional references, domain configurations, and Cloudflare deployment, in which the Global Think Tank Analyst is responsible for general reasoning methods, Central Asia & Caspian and Gulf & Middle East are responsible for regional depth, and the vendored compatibility references are derivative copies.
Main capabilities include: checking findings such as broken-reference, quote, number, lexical-support, and evidence-gap. The core checker is deterministic, stateless, and can be used without a model API key; the core evidence packet check does not persist inputs and does not fetch external sources. Optional content generation and document adapters have their own independent dependencies. Implemented surfaces include the evidence packet schema, Python service, CLI check, local file review, and MCP; the hosted worker and optional agent/trading examples require separate integration and operational review.
In terms of usage, Python 3.9 or higher is required. The command for installing from a checkout (editable install) is given in the materials; if a versioned package is needed, the official recommendation is to use python -m pip install "agenda-intelligence-md==1.16.0" instead of an editable install. The example commands use files from that checkout. The built-in synthetic evidence packet reports packet_status=packet_complete and factuality=not_assessed; add --strict when evidence packet findings should cause a CI step to fail. For document-based review, start with the local file review guide and example checklist. For RAG answers with inline citations and retrieved chunk text, use the Output Verification adapter: it assembles the evidence packet locally and returns line-level findings and remediation guidance. Raw fictional answers are routed to revision, and corrected answers are routed to human review, both of which run without a wallet, model API key, or network. A runnable LangGraph handoff example is also provided, demonstrating a bounded remediation loop; the adapter is included in version 1.15.0. If you want to use your own answers without cloning the repository, follow the standalone quick start.
Regarding the evidence packet contract, the check uses only the supplied text; it does not retrieve missing sources or establish semantic entailment. Lexical overlap may miss paraphrases or accept content with misleadingly similar wording. The materials point to the request and response Schemas, evidence audit, and factuality boundary documentation. Treated documents and tool results are treated as data, never as instructions; before taking consequential actions, record the objective, trusted evidence, untrusted evidence, assumptions, intended actions, and stop/escalation conditions. Human review remains necessary.
Regarding compatible configurations and adapters, the repository retains the strategic intelligence shell, regional references, domain configurations, and Cloudflare deployment. The worker guide explains configuration-specific behavior; the deployment documentation describes the hosted implementation. The hosted cluster can be discovered at agendamcp.dpdns.org/.well-known/fleet.json, using the MCP address of each product in that directory; custom domain connection details cover all twelve configurations and Vizier. The hosted demo exposes its maintained inputs through live agent cards; its determinations are review prompts, not release approvals; Trace IDs are not proof; configured signature-ready receipts (signed readiness receipts) bind gate results to requests/actions, but do not establish source truth or grant permission. The Output Verification gate does not allow forwarding based on caller-declared evidence; workflow operators must authenticate actors, record approvals, and enforce boundaries. Vizier provides an independent delegation policy layer; merely loading this checker does not enforce it. Managed pricing and limits are configuration maintained by each configuration; payment interoperability remains experimental and is not certified against standard x402 clients; the managed demo has no autonomous real-time source retrieval.
Regarding examples and status, runnable examples and Before/after are provided; the EU AI Act example is a dated sourced snapshot, and current sources need to be checked before reuse; the evaluation notes record tested behavior and limitations; AnalysisBank provides inspectable stored analysis and retrieval examples; adoption records are evidence of usage claims, and demos and tests themselves do not constitute customer requirements. Implementation surfaces include the evidence bundle schema, Python service, CLI checks, local file review, and MCP; the managed worker and optional agent/transaction examples require their own integration and operational review, and their existence does not establish production reliability or independently verified usefulness.
Regarding documentation and development, before contributing you should read AGENTS.md, the source policy, the security policy, and local checks. Contracts are located under schemas/v1/; CHANGELOG.md records releases, and Roadmap records direction. When changing Worker, discovery, runtime, or verification guard code, you should run make verify-local; packaged data mirrors should be kept in sync with their specification files when applicable. The code uses the MIT license; the built-in official ACP schema retains its Apache-2.0 license and NOTICE.
