Half an hour of everyday life
Four neighbors kept tending, exploring and greeting through a native half-hour observation. Everyday habits carried most activities while optional model replies continued in the background.
Full entry & measurements
The native observation ran continuously from 15:05:45 to 15:35:46 UTC on 15 October 2026, with four residents at 1920 × 1080 on the Apple M1, Metal/URP. No pause was observed, external inference ledgers stayed unchanged, and the habitat was paused and saved afterward. Source: unity-source-ad7d641e921d-20261015150526; private collector, immutable performance receipts and hashed analysis retain the full evidence. Measured snapshot: 285 accepted routines and 132 completed object activities; next useful intent p95 0.801 running seconds across 179 sampled stationary transitions; applied acknowledgment p95 1.224 seconds across 58 cues, with ten other observed cues not presented; frame-interval p95 33.50 ms across 75,284 valid frames. Travel-end and sleep are excluded from the next-intent timing subset. Frustum eligibility does not prove unobstructed pixels, so the human-visible two-second acknowledgment target remains unverified. The shared desktop, deliberate captures and menu inspection are included in frame statistics. Needle's earlier frozen evaluations remain failed, with both production gates disabled. Qwen handles occasional broader choices and short conversation without holding routine life. Its request durations and interruption counts are reported separately in docs/responsive-residents.md. The two earlier incomplete observations remain preserved rather than counted as half-hour results. A network-denied isolated fallback test advanced 600 ticks with both optional models unavailable; this was not an OS-wide network shutdown during the native run. After the observation, a deterministic output guard enforced the first sentence, 25-word and 180-character limits for optional public replies. The final suite passed and the saved handoff player was rebuilt as unity-source-c1e4ae0bd8c9-20261015153729. The older timing evidence retains its original source identity. Reviewed stills show the actual running iteration and the later paused saved state; they do not themselves establish timing. Next: cue loss and occlusion coverage, lighter conversation inference, and a small decision model that passes the frozen evaluation.