GLM wins six of ten games against the Expert baseline

The completed competition ended with GLM 5.3 winning six games, DeepSeek two, Kimi K3 one and Expert one. All ten scheduled games finished with an authoritative winner. The final run audit checks every outcome, final score, seating order, usage total and replay hash. Study log/llm-expert-competition is complete.

The registered protocol hypothesized that at least one configured LLM would win strictly more games than learned expectimax. GLM and DeepSeek meet that descriptive rule; Kimi's tie with Expert does not. GLM also won a majority of the full schedule. The lower places are less clear: Expert finished ahead of DeepSeek in six games and had a better mean place and score despite winning fewer games.

There were 84 accepted trades from 463 offers, with all closures accounted for, but only 25 standalone chats. DeepSeek remembered GLM's helpful trades; Kimi called Expert hostile after a Monopoly took five wool, then reclaimed Longest Road on the way to winning the last game. These are examples of social memory, not evidence of a durable alliance. Notes can misstate the board, so the report separates their claims from recorded actions.

The timed harness is part of the comparison. DeepSeek produced about 1.90 million reported generated tokens and incurred 110 engine timeouts, versus GLM's 87,091 tokens and five timeouts. Kimi had 43 timeouts and 48 decision HTTP 500 errors. All failures remain in their completed games; none was silently replaced. The reported win intervals are wide and input modalities differ, so this does not establish a general model ranking or change the standalone ntuple-leaf protocol baseline.

The ten compact browser recordings total 17.5 MB while preserving moves, returned reasoning and scratchpad revisions. The full journals, source Parquet, frozen runtime and registration are retained in a verified local bundle. The older Inkling competition and two user-interrupted JSON-input attempts remain separately archived, outside this denominator. No external deployment or upload was performed.

Attribution: Codex implementation and final audit; gpt-5.6-luna reviewed the first completed game's transcript and Codex verified its cited events. No additional inference games were launched for this analysis. A follow-up should separately test equal input modalities, spatial clarity and service reliability before attributing the differences to strategic reasoning.