Choosing things that exist
A resident copied an example object ID during the local trial. The corrected action schema now draws target IDs from that resident’s filtered observation, while the simulator keeps checking whether each choice is legal.
Full entry & measurements
The recovery life-controller trial stopped on a real rejected choice: a resident submitted the illustrative fake object ID from the system prompt. The simulator correctly rejected the nonexistent object and preserved its state. Successful choices and real rest/watering completions were retained, but the trial remains failed. The detailed counts and timing remain in docs/local-first-refinement.md and the private endurance-20261015-030718 records, from native build unity-source-975d21e7292f-20261015005815; they are not a general model-quality score or clean life acceptance. The offline fix removes the executable fake-ID example. Generated targets are limited to IDs in the same already filtered observation: visible objects/neighbors, public landmarks, own memory entries, participant birth proposals, actual menu choices, and blank. The bounded object list can differ from the nearest-first menu, so both contribute their real IDs. Text in old history and earlier actions cannot add targets. Existing history is preserved, including the rejected choice. No global-state retrieval was introduced. Target kind, reachability, reservations, materials and eligibility still pass through the unchanged simulator validator; the vocabulary alone grants no permission. Measured snapshot on 15 October Zurich: 390 Python tests passed in 19.46 seconds, including six new grounding checks. Private grounded-target-final-tests.xml and grounded-target-final-source.json retain the exact result and unchanged hashes. The cases cover privacy, current target roles, obsolete history, removed objects, unchanged semantic rejection and a feasible menu target outside the bounded object list. No live model call used the corrected source at this increment. The old native rendering collector continued independently to its original deadline, with life paused and no further inference after the failed controller trial. Any resulting frame statistics must disclose the intermittent bursts, rejected choice, paused tail and concurrent offline development checks. The later same-run rendering receipt is documented separately; the original life-controller failure is not relabeled as a pass. The remaining authorized allowance is reserved for a separately built corrected-source follow-up. No new image was produced for this nonvisual prompt/schema change. The reused approved bench frame is earlier actual activity evidence, not proof of new-schema model reliability. See docs/local-first-refinement.md.