Every real system carries more state than anyone wrote down — the combinations of
payment, fulfillment, and dispute flags that nobody enumerated. A traditional state
machine handles exactly what you predicted; everything else lands in an incident
channel, a silent else branch, or corrupted data. SESM inverts that:
a Pydantic model bounds what's possible, you map only what you know, and every
unmapped-but-real situation becomes a first-class discovery — triaged by a cheap
AI analyst, then either taught to the machine or flagged as the bug it is.
Instead of surfacing as incidents, situations the system was never taught arrive as deduplicated discovery events — ranked by how often reality hits them.
A bounded space makes "how much of what's possible do we handle?" measurable. Watch it climb in the ticker above as the machine learns.
A mini model (reasoning off, ~a hundredth of a cent per discovery) drafts the judgment: legitimate new state, or impossible — so it's a bug. Its reasoning is recorded on every turn you're about to click through.
Rules you already know go in validate=. Then even a wrong analyst
cannot admit a rule-violating state — compare the blind run's red cells with
the guarded run's clean map.
Every expansion carries what triggered it, who judged it, and why — the state machine's history is its own audit trail. Lifecycles that fit: orders, billing, agent/VM sessions, document workflows — anywhere reality outruns the enum.