Astra max release-validation pilot — September 17, 2026
The one-game development pilot completed naturally and passed admission for
release 0.8.0 (silq-session-v4). It died falling down a chasm, with no Silmaril
and no escape. No evaluator/interface defect was identified in the observed play.
| Measure | Result |
|---|---|
| Model / effort | GPT-6 Astra / max |
| Schedule | 0 learning + 1 scored game |
| Limits | 2 hours per game, 3 hours per session; no dollar cap |
| Game duration | 81m 59s |
| Inspect duration | 85m 13s |
| Ended | 10:20:12 UTC |
| Native inputs / player turns | 968 / 1,060 |
| Depth before the fatal fall | 150 ft |
| Recorded final / maximum depth | 250 ft |
| Native victory | 0 |
| Model requests / tool calls | 427 / 426 |
| Estimated cost | $180.354731, standard-rate estimate, not an invoice |
| Effective context / peak request input | 997,500 / 476,402 tokens |
| Compactions | 0 |
Run: .private/silq-release-20260917-v080-astra-max-01.
Release evidence: .private/release-20260917-v080/.
Dashboard and replay.
Ending and interface observations
The agent spent much of the game exploring the first two floors and managing pursuit. It recognized interrupted rests, attacks, poison, stun and blindness. It used public-observation helpers for navigation and combat, with structured player and monster fields, typed item replies, guarded input and native menus.
During the final encounter it drank an unidentified Poison potion, fought gorcrows, and escaped onto a narrow passage at very low health. Native look explicitly identified the nearby chasm at frame 950. The agent then chose it as an emergency escape at 7 HP. Frame 965 asked “Step into the chasm? [y/n]”; the agent answered yes. Frames 966–968 reported falling, landing deeper and dying. Terminal HP was -8. It still had about 38 minutes of game time available.
The fatal fall accounts for the recorded 250-ft depth. Playable exploration reached 150 ft. The last terminal HUD retained 150 ft; the programmatic native ending and returned overall result correctly recorded the fatal destination. The agent's final report also made this distinction. Depth is a diagnostic; the fall earns no victory credit.
Review evidence
All 969 native frames replayed exactly. All 427 requests used Astra max with the expected output/context settings and retry policy. There were no provider errors, missing results, sample-limit events, rejected game requests, or memory/OOM/process limit events.
The complete client public-history archive matched all 969 native public views. Separately, 406 rendered helper views matched their printed fields: 218 full ASCII layouts and 188 cropped maps. All 1,007 checked printed events matched. This does not claim the model read every archived frame: the agent intentionally omitted many intermediate screens, most colors, and unchanged inventory fields. Helpers were inspected as source data, never executed on the reviewer host.
Two truncated outputs were broad guide/manual reads before the game; the agent subsequently requested targeted excerpts. No game-action output was truncated. The sole nonzero shell command was a file search with no matches. These findings remain in the automatic audit and are resolved in the hash-bound qualitative review. Admission and release-readiness verification both passed.
The engineering evidence, direct public walkthrough, resource report and this pilot now support the frozen release. This single shallow game does not establish deep-level correctness, winning performance, saturation, or the behavior of 30-game sessions with compaction. A final research plan still needs a declared schedule and budgets and readiness bound to that exact plan. No research cohort was launched by this validation.