The request outlasted the work

The Fixture and What We Watched

We ran a bounded live observer trial with Alvin, Pedro and Beatrice: two separate hauling opportunities, with Beatrice initially anesthetized and downed. Native work priorities were disabled to prevent unrelated jobs and rescues; needs, idle behavior and social routines stayed native. No personality overrides were applied. The core remained scripted. Initial offers were paused, followed by five minutes of continuous observation.

What the Pawns Actually Did

Alvin and Pedro each accepted three ten-unit hauls. Alvin noticed Beatrice and requested rescue while retaining his hauling agreement. By the time the core handled that request, his hauling had finished. The replacement-only policy declined the request because its original context had expired—not because Alvin refused rescue. No rescue offer or rescue job followed.

At the later round, Alvin accepted two additional trips. His public reply correctly said his earlier work was complete, noted Beatrice was still down then, and welcomed a rescue proposal. Pedro had no currently observed further hauling option at that round. Eight trips delivered eighty steel. By the final game state Beatrice was standing and wandering without rescue; her recovery's exact time and cause are not established here.

What the Logs and Tests Show

Twenty-two public snapshots preserve eight addressed messages and sixteen work records. They support reconstructing the offers, request, decline and deliveries without private reflections. But the log omits Beatrice's native recovery. This was analyst inspection, not a blinded human usability study.

The window recorded 17,992 ticks and 316 samples, none paused. Nine Claude attempts and three Jev calls were used, below the twelve-attempt/twelve-appraisal ceilings. The six-reflection sub-limit stopped further attention turns early, so this was not five minutes of unrestricted deliberation. No rerolls.

All 147 automated checks passed. Two deadline races found in independent Codex review were fixed and re-reviewed clean; an over-strict rehearsal restore assertion was also corrected. Scripted and live paired/cold restore passed without further inference. Historical ledgers stayed unchanged, and staging is stopped.

What Comes Next

Next, separate a requested goal from the agreement that prompted it: once old work finishes, a still-relevant request might deserve a fresh standalone offer and fresh consent. Selected native state changes and more selective reflection pacing also need attention. These are findings for the next step, not features implemented by this trial.