Resource allowances for the September 2026 study

The v3 research study used 12 hours (43,200 seconds) and 100,000 keys per game, with a $1,200 conservative spend limit per session. These limits were fixed before Astra research and were unchanged by the later context and display fixes. The 8+2 learning condition and fresh 0+2 controls had the same per-game limits and session spend ceiling.

Rationale

At the initial freeze, twelve hours was about 26 times the longest valid development game, and 100,000 keys was about 20 times its largest key count. The upstream manual estimates roughly ten hours for a winning human game (assets/SILQ_MANUAL.txt, line 114). These comparisons motivated a broad initial allowance; shallow losing pilots and a human estimate cannot establish the time needed for this coding-agent setup to win.

Keys include menus and blocked moves as well as actions that advance game turns. The spend limit applies conservative long-context prices to every request, so recorded request-specific cost estimates can be lower. Those estimates exclude developer inference and hosting and are not invoices.

Context configuration

The early gameplay pilot, dev03, used a 258,400-token effective context and compacted twice. Both evaluated models had 1,050,000-token API metadata, but that metadata did not configure the client. Live telemetry exposed the same mismatch in the first Astra session, leading to exclusion of the entire v1 cohort. No game had finished and no compaction had occurred in that session.

Setting the client window alone also failed: the bundled catalog capped GPT-5.5 at 272,000 tokens, and the client clamped the override to that maximum. Dev04 confirmed the 258,400-token effective window and was stopped. Neither excluded run supplies evidence that the context mismatch caused a defeat.

The corrected image copies the pinned client's model catalog and changes only context_window, max_context_window, and auto_compact_token_limit for the two evaluated models. Other fields, including prompts and tools, are preserved. The client loads this catalog through model_catalog_json with:

An offline check runs the built client and verifies the catalog transformation. Admission also checks live context telemetry and every raw request's output allowance. The 24-row dev05 pilot passed both checks. Compaction remains part of the setup and does not guarantee retention of every earlier detail.

The configuration mechanism is described in the Codex reference. The clamp appears in the model configuration code and was reproduced with client 0.153.2. That link points to an upstream branch, not the pinned source used by the client.

Development measurements

No resource limit ended these games. Dev03 and dev05 used the superseded 24-row terminal; dev06 validated the corrected 27-row interface.

Dev05's 413 native frames replayed exactly and were all captured as public screens, with 560 matching views. Review covered 171 tool results, 103 visible assistant messages, and 64 available summaries. Peak input was 179,950 tokens, with no compaction, clipping, missing usage, or service error. Eleven coordinate outputs matched the screens. The only code exception was a repaired missing import in an agent's read-only script.

Dev06's 506 frames also replayed exactly and were all captured, with 744 matching views. Review covered 264 tool results, 244 visible messages, and 97 available reasoning summaries. Peak input was 318,161 tokens with no compaction. Every request allowed 128,000 output tokens and the client reported 997,500 effective context. There were no service, model, clipping, or execution errors. The game lasted 411 player turns. Pilot review.

Direct play on the 24-row image reached 100 feet and ended in deliberate quit; all 81 frames replayed exactly. Earlier checks reached 150 feet and revisited a pilot's claimed blocked pocket. The native binary was unchanged by the context-only revisions. The pre-v2 suite had 46 passing checks.

V2 was excluded when song redraws were found to erase or overwrite depth at 24 rows. Its first session had eight practice deaths and an unfinished scored game. We retained those records and the original methods, source, and readiness in .private/scaffold-v2-freeze; they did not change the resource allowance. The 27-row correction produced a new binary and 49 passing checks.

Direct play on the corrected build reached 100 feet, started and stopped Elbereth without losing depth fields, and quit at full 41 HP. It tested abilities, six Wolf kills, a Grimhawk, traps, bashing, stun recovery, rest, camera shifts, inventory, and the overhead map. All 270 frames replayed exactly and all 92 recorded public screens matched. It used 269 keys and 318 turns over 1063.07 seconds, including developer deliberation. The developer had prior source knowledge and read no hidden state before quit and cleanup. This was an interface check, not an independent human baseline. Direct-play record.

Admission rule

Every research session receives a complete review. Time, key, or spend limits, infrastructure failures, and incomplete schedules prevent a session from establishing an unsaturated capability result. If a limit materially constrains play, retain the evidence, document a revised allowance, and run a fresh cohort. Do not replace selected losses or change the saturation threshold after results.

Before v3 dispatch, analysis/readiness.json bound the corrected dev06 pilot, direct-play04, engineering checks, sanitizer and hidden-state fixtures, isolation probe, and Inspect integration test to their hashes. No Astra outcome was used to set these allowances. Later harness revisions require new readiness evidence; these measurements do not calibrate a new study.