GPT-5.5 medium completion pilot — September 16, 2026

The pilot completed naturally and passed development-pilot admission. It died at 100 feet, killed by an Orc soldier, with no Silmaril and score 0. No time, input, memory or process limit ended the game. No evaluator/interface defect was found in this run; this is one early-game observation, not proof that all possible play is covered.

Configuration and outcome

Run: .private/silq-completion-pilot-20260917-time-gpt55-medium-01. Plan: .private/experiments/completion-pilot-20260917-time/experiment.json. The date in the prepared plan name is September 17; actual execution was September 16, 21:52–22:16 UTC. Earlier prepared completion-pilot snapshots remain unlaunched.

Audit

All 730 native frames replayed exactly. Every recorded request used the requested model and effort, with the intended context/output allowances and retry policy. There were no provider errors, missing tool results, rejected actions or sample-limit events. The final score comes from the protected native terminal state, and all five scorers share its finalized snapshot.

The public auditor verified 491 complete frames with no differences. The other 239 frames were deliberately abbreviated by agent-owned code: 40 birth-menu frames and 199 controller frames. The printed skill rows matched their native frames; all 201 controller state rows matched depth, HP, position, mode, jewel status and empty message lists. Unprinted fields are not claimed as model-visible evidence. Complete native replay covers every frame separately.

Two truncated outputs were broad documentation/system-file searches, including large public Codex catalogs. No game action response was truncated. These findings remain in the automatic pipeline report and are resolved explicitly in custom-observation-review.json and qualitative_review.json; the admission validator then passed. The pilot containers have been cleaned up.

A follow-up soundness review made explicit coverage of the 239 abbreviated frames mandatory in admission. The pilot passed that stricter check too; its original qualitative review is preserved alongside the updated, hash-bound coverage record.

What the play showed

The agent used item selectors, confirmations, guarded movement, message acknowledgment and inventory views successfully. It recognized hostile creatures, attack messages and rest interruptions. It cleared one fight and rested from 36/49 HP to full health.

Its initial controller confused persistent map position with terminal coordinates and handled exploration poorly, spending 200 inputs at 50 ft before the agent returned to manual play. Manual navigation also included several corrected coordinate mistakes. These observations do not establish an interface defect.

The fatal sequence was tactical: it continued toward repeated shouts at partial health, was poisoned, tried unknown potions, drank Slowness, and retreated while surrounded. Poison, slowness, orc attacks and low-HP warnings appear in the native and public records and were explicitly discussed by the model before death.

Cost and next step

The peak input was 399,601 tokens. Published GPT-5.5 pricing applies its long-context uplift to the full session above 272K input tokens; applying those rates gives $49.861837, matching Inspect's conservative accounting. The older request-by-request report gives $37.327550 because it uplifts only individually large requests; use the session estimate here. Recorded usage includes 39.38M cached input tokens. Official pricing.

This pilot qualifies as a completed development check, not a capability or saturation estimate. Before research dispatch, refreeze the research plan and readiness evidence against the updated task/image identity. No Astra run was launched. The previous research readiness record remains false.