ResultGate

Local MCP tools for extraction-contract checks and literal citation integrity. Version 0.1.1 is a release candidate, not yet published to a public registry.

A pass is not a truth guarantee. ResultGate evaluates the evidence you supply; it does not fetch sources, detect every missing fact, verify source authenticity, or judge semantic support. Incomplete completion evidence returns unknown.

Install from GitHub (Python 3.11+ and Git)

python3 -m venv .venv
.venv/bin/python -m pip install git+https://github.com/AHEDMUNI/resultgate.git
.venv/bin/resultgate --help

For reproducible deployments, append @<reviewed-commit-sha> to the Git URL.

Install from a locally built wheel

On Linux/macOS, after reviewing the artifact and its SHA-256:

python3 -m venv .venv
.venv/bin/python -m pip install ./resultgate_pilot-0.1.1-py3-none-any.whl
.venv/bin/resultgate --help

On Windows use py -m venv .venv and .venv\Scripts\python.exe -m pip install <wheel>; the executable is .venv\Scripts\resultgate.exe. Windows/macOS execution has not yet been tested in this release review.

After PyPI publication is verified, the registry launch command will be uvx resultgate-pilot==0.1.1. That public package command is NOT usable until publication. The package exposes both resultgate and resultgate-pilot executables.

Connect an MCP client

Configure the absolute installed executable; it runs locally over stdio, without API credentials:

{"mcpServers":{"resultgate":{"command":"/ABSOLUTE/PATH/.venv/bin/resultgate"}}}

Discoverable tools: check_extraction and check_citations. Each accepts payload and an optional usage object. Tool input/output schemas are exposed by the official Python MCP SDK. The default makes no network requests and records no usage. This local server trusts the invoking OS user; it is not a sandbox for untrusted processes.

Extraction example

{"payload":{"records":[{"id":"a","name":"Example"}],"required_fields":["/name"],"unique_fields":["/id"],"expected_count":1,"job_status":"completed","next_cursor":null,"truncated":false}}

Expected status: pass, scoped to caller_supplied_contract_only.

Fields use RFC6901 JSON pointers; unique_fields forms a composite key. Missing/null/blank required fields fail; zero and false are valid. Counts refer to the entire supplied dataset, not pages or provider jobs. An expected total must come from an appropriate external contract, not simply len(records). Known failure, scraping-in-progress, pending cursor, repeated cursor, explicit truncation, count mismatch and duplicate keys fail. Missing expected count, completion state, next-cursor evidence or truncation evidence yields unknown when there is no definite failure. Null is allowed only for next_cursor; omit unknown expected_count/truncated instead.

Limits: 1 MiB serialized input, 100,000 JSON nodes, depth 32, 5,000 records and 64 pointers per field list. Up to 100 findings are returned, with total count.

Citation example

{"payload":{"sources":[{"id":"s1","text":"The launch was postponed."}],"citations":[{"source_id":"s1","quote":"The launch was postponed.","start":0,"end":25}]}}

Exact literal Unicode matching only. Optional offsets must be paired and count code points, not bytes or UTF-16 units. A matching quote may still be misleading or used to support a false claim. Empty citations produce unknown. Duplicate source IDs are rejected. Search work is limited to 20 million source characters across searches without offsets; split larger requests or supply exact offsets. No URLs are fetched or instructions in source text executed.

Optional authenticated local HTTP API

export RESULTGATE_API_TOKEN="$(python3 -c 'import secrets; print(secrets.token_urlsafe(32))')"
.venv/bin/resultgate --http
curl -H "Authorization: Bearer $RESULTGATE_API_TOKEN" -H 'Content-Type: application/json' --data-binary @examples/extraction.json http://127.0.0.1:8765/v1/extraction

POST /v1/extraction and /v1/citations accept the payload itself, without the MCP payload wrapper. GET /health is informational. HTTP mode binds to loopback, requires a configured secret for audits, rejects browser Origins/unrecognized Hosts, limits body size and read time, and disables access logs. It is not a public hosted API or a remote MCP endpoint. Do not publish it through a tunnel or reverse proxy. HTTP usage measurement is not implemented.

Privacy-preserving, opt-in local measurement

No analytics by default. To opt in for a voluntary pilot, add --usage-dir /YOUR/PRIVATE/DIRECTORY to the MCP command, then pass usage:{"task_id":"opaque-task-identifier","purpose":"real_task"} when a real task uses a check. Reuse task IDs on retries. Test/demo/retry purposes and calls without a usage context are excluded. See PRIVACY.md.

Export for your own review with .venv/bin/python -m resultgate.local_usage /YOUR/PRIVATE/DIRECTORY --export. Nothing uploads automatically. Shared summaries are self-reports; they are not automatically counted as independent adoption or revenue.

Development and limitations

python3 -m venv .venv
.venv/bin/python -m pip install -e '.[test]'
.venv/bin/python -m unittest discover -s tests -v
.venv/bin/python benchmark.py

Unit and clean-install tests are internal. Synthetic benchmark fixtures measure rule coverage, not production accuracy or superiority over complete competing products. Tests use no paid models, browsers or APIs. Python 3.12/Linux is the tested clean-install environment. There is no payment code, hosted endpoint, or external customer count.

License: MIT, included in LICENSE. See SECURITY.md for security scope and THIRD_PARTY.md for dependencies.