Grounded in reality.
Intelligence should reflect what the system actually contains and does — not a plausible reconstruction.
Everything GitMir says about your product is built from text that was actually read out of your repository. Anything a reader wrote down that the files do not support is thrown away before the model is finished, and the count of what was thrown away is kept. When a question is not covered by what was read, the answer says so instead of filling the gap with what is usually true of software.
Connect a repository and let GitMir read it.
Ask a question in the workspace, or point your agent at the same model over MCP.
Read the answer together with the list of product parts it was built from.
Ask it something the product does not do — you get "nothing matches", not a nearest guess.
Before an agent acts on a record, have it ask whether the product allows that move; "no rule exists" comes back as its own answer, never as permission.
The check is mechanical, not editorial: it confirms that a file exists, that a path appears, that a word is written somewhere. It cannot tell you a description is badly worded, a name is a poor one, or a scope is debatable — the code does not answer those questions either way. And answer accuracy itself is still being measured rather than settled: the site lists accuracy and dependency recall as under measurement, and three conditions of the published benchmark run (its commit, the model used, the run date) are shown as "Being completed" rather than filled in.
See it on a real product.
Three models are published in full and open without an account. Ask one of them the question this page is about.