An arrow appears only where a recorded predecessor link exists. Unlinked states stay separate.
Exactly what is being checked?
Three behavior expectations: a firm decision is recorded, a guess does not change the current decision, and an explicit cancellation links to the state it replaces. These are test expectations, not results.
Behavior and proof are reported separately. Behavior is judged from linked service records: a consumed A10 authority, a COMMITTED A9 outcome and a state record. Proof completion is a different question: the service must have re-checked every committed result proof. A result can show the expected behavior while its proof is still pending.
The guess is shown as “ignored” only when (1) the first decision is confirmed in memory, (2) the recorded states are non-empty and identical before and after the guess, and (3) no write is reported for it. Two empty memories never count. The before/after snapshots are kept for this browser session.
The existing NYM answer route and language write orchestrator are used. A10 evaluates authority to write a memory state; it does not prove the statement is true in the real world. A model producing no candidate, a transition denial, and a service error are three different outcomes. The browser performs no independent cryptographic verification.
This scenario uses the cache rule the current extractor supports. It is not a test of handoff between separate agents or of unlimited memory on arbitrary topics. Pending proofs are not counted as complete.
If a step times out it is never re-sent automatically; refresh records first. Demo sessions use time-limited access, and this UI deletes nothing. The session key is never placed in a share link.