evidence-gate

Daily AI Agent Toolkit

Local, deterministic tools for checking AI-generated work before it is released.

The toolkit contains two Model Context Protocol (MCP) servers and six portable Agent Skills. They inspect retained artifacts, declared claim support, literal citations, completion contracts, blockers, and release receipts without calling a model, uploading files, or requiring a Daily AI Agents account.

Release status: 0.1.1 patch release candidate. PyPI 0.1.0 is public; the patch corrects the case-sensitive MCP Registry ownership marker.

Why this exists

AI systems can produce convincing output without proving that required artifacts exist, cited text is present, or a completion contract was actually satisfied. These tools make those narrow checks repeatable and preserve uncertainty instead of silently upgrading it to success.

They deliberately do not determine broad truth, judge semantic entailment, approve a release, validate a live service, or establish legal or security compliance.

Included tools

Component	Interface	Purpose
Evidence Gate	MCP server, four tools	Inspect artifacts, declared claim evidence, literal citation containment, and aggregate verification states.
Release Gate	MCP server, four tools	Check completion contracts, preserve requirement states, format blockers, and build retained release receipts.
Six Agent Skills	Portable skill directories	Reuse artifact, claim, contract, failure-state, and verification procedures with deterministic self-tests.

Evidence Gate

• verify_artifact
• audit_claims
• audit_citations
• summarize_verification

Release Gate

• check_contract
• evaluate_completion
• format_blockers
• build_release_receipt

Agent Skills

• artifact-verifier
• claim-truth-gate
• completion-contract
• contract-checker
• fail-loud
• verification-bench

Install and run

Python 3.11 or newer is required.

python -m pip install dailyaiagents-evidence-gate==0.1.1
python -m pip install dailyaiagents-release-gate==0.1.1

dailyai-evidence-gate --root /absolute/path/to/workspace
dailyai-release-gate --root /absolute/path/to/workspace

Until the packages are public, install from a clean checkout:

python -m pip install ./servers/evidence-gate ./servers/release-gate

Both servers use MCP stdio transport. A representative client configuration is:

{
  "mcpServers": {
    "dailyai-evidence-gate": {
      "command": "dailyai-evidence-gate",
      "args": ["--root", "/absolute/path/to/workspace"]
    },
    "dailyai-release-gate": {
      "command": "dailyai-release-gate",
      "args": ["--root", "/absolute/path/to/workspace"]
    }
  }
}

See client configuration for the supported transport boundary.

Reproduce the proof

Run the complete local gate:

bash scripts/test-all.sh

The release candidate includes:

• unit tests for both servers;
• deterministic self-tests for all six skills;
• 20 declared-outcome examples spanning all eight MCP tools;
• rooted-path and symlink-escape protections;
• clean-wheel installation and stdio tool-discovery checks;
• package-specific checksums, SBOMs, and build provenance.

See the examples, technical report, receipt schemas, release and recovery process, and local verification record. A passing local check proves only the condition and scope named in its receipt.

Architecture and security boundary

Live repository controls and current fail-closed publication gates are recorded in Repository controls.

MCP client
   |  stdio + structured inputs
   v
Evidence Gate / Release Gate
   |  canonical path resolution beneath --root
   v
Local retained files  --->  structured status + scoped receipt

The servers do not use network access, telemetry, model inference, arbitrary shell execution, or hosted Daily AI Agents infrastructure. Inputs are treated as data rather than shell fragments. URL-only evidence remains UNVERIFIED, and literal citation containment is not semantic support.

Read SECURITY.md and the threat model before using the toolkit on sensitive work.

Project leadership

The toolkit was conceived and led by Cooper Reed, founder of Daily AI Agents LLC, with implementation, testing, release engineering, and documentation developed as a public engineering project. Public claims about the toolkit should link to reproducible checks or retained release evidence in this repository.

Contributing and support

• Contributing guide
• Support policy
• Release process
• Changelog

Apache-2.0 licensed.