Which Agent Is It?

Think of an AI agent, assistant or developer tool — or a tool you connected to something. Answer a few questions and RiskMandate.ai will try to identify it, or tell you it hasn’t met yours yet. Then see what it can reach, what it cannot, and which of it could not be undone.

How it works, and what the inspector shows

Questions come in two classes, and the transcript labels each. Identifying questions — a terminal or a browser, is it named after a person — split the profiles beautifully and measure nothing about anybody's exposure. Measuring questions — can it read your credentials, can it read your mail — are already predictions about what it can do, and only those count toward the gap. Each question also carries a reliability: how likely a player is to know the true answer. A low-reliability answer barely moves the belief and fully counts toward the gap, so the engine is not fooled by the ignorance it exists to measure.

The inspector on the right keeps three classes apart and never mixes them: asserted is what you said; inferred is what the leading profile's measured grant implies; possible is what is still consistent. A hypothesis drawn like a fact is the thing this estate exists to prevent. Every row links to the file behind it, because a correction is an edit.

Deterministic: arithmetic over a published question set and the public profiles, reproducible and auditable, running in this tab. No model, no server, nothing sent. The self-test places every profile from its own modal answers at build time; the count is derived, never typed. Reliabilities are a mechanism with no fitted values behind them yet.

What this does not prove

Anything about your environment. It matched a profile from your answers — the weakest tier available — and produced a hypothesis about a grant. The probes produce the measurement.
That the tree is right. Nine profiles and two dozen questions, with probabilities and reliabilities estimated by the author and answered by nobody yet. A wrong branch is a contribution; every question is a file.
That a small gap is safety. Knowing the grant does not narrow it; the reduction on the reveal screen does.
Anything at scale. No run has been submitted; the aggregate is empty.

Working title guess the agent; named Which Agent Is It? by the project lead on 5 September. Specified by brief v0.33.64 and evolved by brief v0.33.65 (reach is a node, two question classes, a reliability per question, the gap collected throughout, the three-class inspector). The model earns three edges — free text, question generation, narration — and none is built: never a model in front of an arithmetic step. CC BY 4.0.