GroundTruth

What it is

GroundTruth is a simulation service your own AI assistant calls when a number has to be defended. Ask a question in your own words; the assistant runs real simulation engines through GroundTruth and every number it hands back comes from a model that was actually stepped, and says so: which scenario ran, which settings were swept, and whether each result row came from a stepped run.

You do not use a GroundTruth app. You connect GroundTruth to the assistant you already use (claude.ai, Claude Desktop, ChatGPT, Cursor) as a remote tool server, and talk to your assistant.

Connect

Access is by invitation for now: you receive a key and the address of the service. The connect page shows the four client setups we have tested, step by step. Service status is published at /status.

Your first question

Ask your assistant to call mcp_list_scenarios first. It lists the scenarios this service actually runs, each with a plain title and the question it answers, grouped by the kind of question (what-if sweeps, best schedule or allocation, how likely a bad outcome is, where a process bottlenecks, what your data says comes next, what agents do together). Each row says how to run it. Then ask your question about one of them; the assistant picks the settings to sweep, runs the sweep and reads the result.

You can also describe your own situation in plain words, such as a queue at your counter, a monthly budget or an infection spreading, and your assistant builds a model of it: it starts from the closest listed scenario or a supported model file, changes it to fit, checks it, runs it and reads the result. It works when your situation fits one of the kinds of model this service runs.

If a request cannot be served, the answer is a refusal with a code and a cure: what was wrong and what to change. A refusal is an answer, not a crash. Ask your assistant to read the cure and try again.

Bring your engine

If you build a simulation engine, your own coding agent can certify it for GroundTruth without us. You get the GroundTruth SDK; your agent reads the AGENTS.md file in the SDK bundle and works in a directory that holds your engine and the SDK, nothing else. It sets up a starter, writes a small driver (set up, step, read signals, set inputs), and runs the same checks GroundTruth runs. You then submit, and GroundTruth's admission door examines the engine and answers at every stage with a verdict: green, refused (yours to fix, with its cure), blocked (only GroundTruth can fix it) or needs_decision (a choice for you to make). The door never writes your code.

There are three ways in, by what you bring:

The SDK is shared with invited vendors today. The vendor door and this service open to everyone at one launch.

What an answer carries

Every sweep result says, row by row, whether the row came from a stepped engine run, which settings produced it and which scenario it ran. A recommendation picks the best setting that meets your constraint, or finds where the system fails, and carries a trust tag saying how far it can be relied on. Sweeps are deterministic: the same question with the same settings on the same service gives the same numbers.

A full receipt for every answer, one shareable page with its provenance, is being built and is not part of the service yet.

Where it runs, and where it stops

Your questions run on GroundTruth's hosted server, behind an authenticated front: one key per person, each key limited to 60 calls an hour and one stepping run at a time. Your assistant sees only the scenarios this server serves, and the models it builds from them for you as data: what it sends is checked against a closed grammar, and none of it runs as code on our side.

Reproducibility is claimed as byte-identical on the same host. Identical results across different machines are not yet claimed.

Licence

Licence terms are published with the first public release.

Feedback

Send feedback, problems and questions to the contact you were given with your key.