vassiliylakhonin/agenda-intelligence-md

A deterministic evidence-packet linter for claim-backed AI output. Supply claims, source references, quotations, and source text; receive broken-reference, quote, number, lexical-support, and evidence-gap findings 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.**
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, from a checkout:
For the versioned package, use python -m pip install "agenda-intelligence-md==1.16.2" 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:
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:
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
| Input | Check |
|---|---|
| Claims and source IDs | References resolve within the supplied packet |
| Declared quotes | Quotations match the supplied source text |
| Claim wording and numbers | Deterministic support heuristics, unmatched numbers, and negation checks |
| Findings | Packet status, evidence gaps, and reviewer actions |
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
| Interface | Start here |
|---|---|
| CLI and Python service | Quickstart, check_evidence_packet in services.py |
| Local MCP (32 tools) | MCP.md; launch agenda-intelligence-mcp from the installed environment |
| Hosted MCP (per-product catalog) | Hosted quickstart; native free_trial_status and free_trial |
| CI evidence linting | GitHub Action, with text, JSON, or SARIF output |
| Agent repair loops | Integration guides |
| Protected agent actions | Interaction Trust + Vizier gate |
| MCP pre-release checks | MCP Integration Check CLI / Action |
| Agent checkout experiments | ACP sandbox readiness check |
| HTTP and A2A | HTTP shell and A2A adapter |
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.