The neighbors plant their first seeds

Native acceptance, 15 October 2026 Zurich. A private backup and coordinated service restart preserved the exact residents, objects, settings, time and provider configuration. Qwen3.5-2B passed a fresh native eye-camera check, then all four local residents took part in a bounded live session. Moss received one optional invitation to an existing empty planter; the simulator validated every subsequent model choice. Moss and Pebble each completed timed planting at an existing plot. Pip visited the market and created a new planted plot with the separate plant action. Ember explored the village and market. Moss later chose the market; Pebble chose a bench and was still approaching it when the session paused, so no resting reward is claimed. Two existing plots changed from empty to planted, plus Pip's new construction. There was no watering or harvest in this run. Authentic earned changes remain saved. Measured snapshot: 8 admitted local calls produced 8 accepted decisions; 2 timed planting activities completed; the run lasted 117.58 simulation seconds; per-call model latency was 13.28–16.50 seconds. These measurements come from the private bounded-run receipts at 23:31–23:33 UTC on 14 October (01:31–01:33 Zurich on 15 October), using native build unity-source-50c17e52bc6a-20261014232030 and the unchanged tested Python handoff. They are a short acceptance check, not sustained autonomy or a hardware benchmark. The earlier 344-test result was not rerun for this documentation update. Private image receipts cover all four residents with hash-verified 512×384 Unity eye views from that build. Raw resident images, state and model records stay private. The public visuals are separately filtered actual native gameplay frames, produced and individually approved by the game owner; the model-support session also inspected their pixels and sidecars. The running overview shows the real session, while the closer paused garden shows saved stages. Completion is established by simulator receipts alongside these stills, not by an image alone. Final read-back confirmed a paused world, no pending turns, the original local limit restored, unchanged settings and persisted state matching the end snapshot. Cloud and Codex usage ledgers were unchanged. No paid service or internet publication was used. Next: verify that residents return to crops for watering and harvesting, and assess a longer native session with bounded local calls. See docs/local-first-refinement.md for scope and provenance.