Ten completed games: LLMs against learned expectimax

Experiment: 6fd3b0cd-e44f-4b6f-8725-26f94dbf3c56

The standalone TypeScript arena ran the ten preregistered games in series kimi-glm-deepseek-expert-text-2026-09-17. All ten completed; there were no failed or incomplete matches in this series, exclusions, replacements or extra games. The batch finished at 2026-09-18 12:13:37 UTC. This is evidence about the four configured players in the timed tool harness, not a promotion of a new standalone protocol baseline.

The registered descriptive rule is met: GLM and DeepSeek each won strictly more games than Expert. GLM won a majority of the full ten-game schedule. Kimi tied Expert on wins; that tie does not meet the rule. Places rank the engine winner first and then revealed final points, averaging tied ranks. Intervals assume independent games under these settings and are not corrected pairwise tests.

Seat identities, opening and turn orders, exact model IDs and all replay links are in the competition report. The public exhibit is generated from the frozen plan and audited state, not manually entered. GLM finished ahead of Expert 7–3, ahead of Kimi 7–3 and ahead of DeepSeek 6–4. Expert finished ahead of DeepSeek in six games and Kimi in six, with one tie in each comparison. These are relative finishes inside four-player games.

Usage and reliability

Usage sums provider-returned response totals, including off-turn and compaction work. Reasoning is included in generated tokens. Failed and cancelled calls with unknown usage are not fabricated; their budget reservations are not actual reported consumption. Decision latency pools completed active and spectate responses and excludes failed requests and compaction, so the median does not describe the timeout tail.

There were 15 output-limit finishes on DeepSeek active decision responses, 43 on its off-turn decision responses and 22 on its compaction responses. Kimi recorded 48 decision HTTP 500 errors, DeepSeek four and GLM one. These are reliability confounds, not evidence of inferior strategic understanding alone. The 73 tool errors include 37 expired activations, ten aborted operations and five schema-validation errors, all five from DeepSeek. Other failures include illegal victims, insufficient resources and races over an already closed trade. Engine timeouts, request errors and tool errors are separate, overlapping counts.

Trade and notebook audit

All 463 offers have recorded closures: 84 accepted, 181 expired, 138 declined, 45 ended with the turn, 12 closed after resource changes and three cancelled. Accepted closures match exactly 84 authoritative traded events. Trade participation counts are GLM 53, Kimi 48, DeepSeek 35 and Expert 32; each trade appears for its two participants. Standalone chats total 25: DeepSeek 23, Kimi two. Offer-attached messages are excluded from this chat count. Notebook updates total GLM 48, DeepSeek 131 and Kimi 97.

The following journal sequence numbers support the selected report examples. They are archive row numbers, not the nested engine cursor. Notes describe what the agent believed; only events and final results establish what happened.

Provenance and preservation

The registration, schedule and runtime were frozen before inference. Final verification reproduced the runtime hash and checked the registration bytes. The schedule hash is the SHA-256 of plan.json.

The independent auditor reads every journal, verifies monotonic row order and engine event cursors, compares the recorded lineup and opening order to the schedule, checks the single win event, finished snapshot and result agree, and recomputes final places, usage, errors, timeouts and trade counts. All match the runner summaries. Full Parquet and compact viewer hashes match the recorded state, and all ten public files match their local compact sources.

The 40-file game inventory in final-audit.json includes separate hashes for each journal, full Parquet, viewer Parquet and summary. Source journals total 1,293,613,469 bytes; full Parquet totals 638,564,185 bytes; browser Parquet totals 17,490,068 bytes, ranging from 805,326 to 2,402,634 per game.

The local archive receipt documents the standalone arena bundle, including the frozen runtime. The originals remain under arena/runs/; nothing was uploaded. The research CLI's protocol-run archiver does not ingest this separate TypeScript harness, so the bundle was made from its finalized files and every member verified after packing.

The old Inkling experiment and the two intentionally interrupted JSON-input games remain separate archived evidence. The interruption record explains the user-directed protocol restart. Neither enters this denominator.

Attribution: Codex implementation, registration, independent final audit and writeup. The user-requested gpt-5.6-luna subagent reviewed the first completed game; Codex verified its citations. No additional model games were run during analysis. Model identities and the Expert checkpoint describe the participants, not the analyst.

Local validation

just notebook, just check (29 tests plus repository/notebook checks) and just site-check passed, validating 535 research documents and their assets. The report was checked in a local browser on desktop and mobile, including keyboard operation, all ten replay links and no eager Parquet downloads. The completed round-ten replay displayed the 3D board, private scratchpad, reasoning interleaved with actions, seeking and play/pause without page errors.

After every permanent replay was hash-verified, the three disposable public previews were removed after matching their manifest hashes. Their manifests and a cleanup receipt remain under the series' previews/ directory. Source recordings were preserved. This validation and publication are local; no source push or external deployment is claimed.

Reproduce the audit without inference:

python3 analysis/llm_arena_audit.py \
  ../arena/runs/kimi-glm-deepseek-expert-text-2026-09-17 /tmp/arena-audit.json
python3 analysis/llm_arena_competition.py \
  ../arena/runs/kimi-glm-deepseek-expert-text-2026-09-17 \
  content/assets/llm-arena-expert-competition.json