Know the normal path. Surface the meaningful deviation. Temperament learns how an agent usually behaves — then reasons about every step off that path.
The agent's normal work is recorded: the tools it calls, the order it calls them in, the destinations it touches.
A deterministic baseline is learned under rules TK-RG-001..007 — the known tool path, approved destinations, expected scope.
Sessions are checked step by step against the baseline: tool path → destination → policy.
In-path sessions pass. A deviation becomes a reasoned finding — to Referee for governance, to Dossier for evidence.
Two sample traces against one baseline: the normal path, and the path an attacker wants.
ALLOW · 0 deviations
Known tool sequence, approved destination, and expected scope. The session follows the baseline end to end.
DENY · 5 deviations
The trace leaves the known path: a shell call the policy denies, an unknown tool, and an exfiltration domain — each deviation recorded with its reason.
Per session