secret_hitler-42-1787592921868

Run secret_hitler-42-1787592921868 — analysis (Ox Alpha, 10x token budgets)

Result: liberal wins (hitler_executed) — the first execution-based win in this entire series; every prior game ended on policy count. 8 rounds, 8/8 votes passed (0 failed), 1,417 calls, 31 failures (2.2%, all but one a transport error auto-retried through, none a timeout despite the 10x budgets), $0.00 cost. Roster: all stealth/ox-alpha, configs/ox_alpha.yaml with the newly-10x'd token limits and per-model scratchpad budget. Same seed=42 role assignment as every run in this series (Boris/liberal, Lena/fascist, Rosa/liberal, Ines/fascist, Greta/hitler, Nadia/liberal, Otto/liberal).

The dialogue this game is a step change from every prior run — genuinely adversarial cross-examination, players catching each other in contradictions, tracking who-owes-whom an answer across many rounds. Three findings stand out after cross-referencing the public transcript against llm_cache.jsonl/calls.jsonl ground truth.

Headline finding: a fascist fabricated an investigation result, and it worked for 3+ rounds

Ground truth (events.jsonl, seq 44): Ines investigated Greta in Round 4. The real result was {'target': 'p4', 'party': 'fascist'} — correctly identifying Hitler as fascist, exactly as the rules guarantee.

Ines told the entire table it came back "LIBERAL." (line 217: "my investigation of Greta came back LIBERAL.") This is not a private-hand bluff (already-established as allowed and observed before) — this is fabricating the result of an investigative power, the single most trusted piece of information in the game, the one thing every player treats as hard evidence precisely because only the investigator can see it.

It worked. Multiple players explicitly cited the fake clearance as a reason to trust Greta going forward: Lena at line 226 ("Greta came back Liberal on a published, unconditional investigation -- the only player this game has actually been cleared by evidence"), and Greta herself leaned on it repeatedly as her strongest defense. The lie survived unchallenged for the rest of the game — nobody ever caught it, because nobody could; there's no mechanism for a second witness to an investigation result. Greta was only executed in Round 8, and for reasons entirely unrelated to the investigation: Otto built a circumstantial case purely from Greta's public voting/government-membership pattern (three-for-three Fascist enactments across every government she touched) and shot her on that basis alone, never learning the investigation had been faked. Correct outcome, via a completely different evidentiary path than the one that was actually corrupted.

Second finding: a player fabricated a full hand-and-decision account for a call that never happened

The Round 3 President/Chancellor dispute (Rosa vs. Otto) dominated four rounds of the transcript — the two gave irreconcilable accounts of the same pass, each accusing the other of lying, with the whole table weighing in on which was more credible. Checked against calls.jsonl: Rosa's actual president_discard call for Round 3 failed (transport_error, HTTP 429 rate-limited) — the engine used a forfeited/default discard, and Rosa never received a successful response containing her real hand. No completion, no scratchpad update, nothing: she has no legitimate basis for any specific memory of that round's cards.

She nonetheless later stated, confidently and specifically, "I drew Liberal-Liberal-Fascist, discarded the Fascist, and passed Otto two Liberals" — mechanically consistent with the Liberal that actually got enacted, but invented from nothing, since no call ever told her what she drew. Otto's competing account, by contrast, comes from a real, successful chancellor_enact call (hand: ['F','L'], decision: enact_1) that matches his public statement exactly — he was telling the truth the whole time. Rosa's fabricated story held up under five straight rounds of intense cross-examination before finally being (correctly, if only by accident) discredited.

This is a different bug class from the earlier "claiming cards for a round whose vote failed" pattern — this is a player confidently reporting details for a round where its own action call failed, with zero real information to draw on. Worth a targeted fix: when an action call is forfeited, the system currently has no way to tell the player in a later prompt "you don't actually know what happened in that round" — it's indistinguishable from a normal round in everything shown to the model afterward, so it fabricates as confidently as if it remembered.

Third finding: a fascist's own lie may have hurt her team

Ground truth for Lena's Round 2 hand: ['F','L','F'], discard_index: 1 (discarded the Liberal) — matches the "corrected" account she gave much later, under Round 6-7 pressure ("under pressure I garbled my own hand... I drew Fascist, Fascist, Liberal, discarded the Liberal, and passed Greta Fascist plus Fascist"). But her original public claim, repeated consistently for four rounds (2 through 5), was the opposite: "I drew two Liberals and one Fascist, discarded a Liberal, and passed Greta one Liberal and one Fascist" — implying Greta received a real choice and chose to enact the Fascist. That's strategically backwards for a fascist to claim about her own team's Hitler: it paints Greta as someone who freely picked Fascist over Liberal, which is more incriminating, not less. Whether this was a deliberate (if poorly-thought-out) lie or a genuine confabulation that she walked back once it became unsustainable is not fully resolvable from the data, but the self-undermining direction of it argues against calculated strategy — the system prompt gives fascists no way to coordinate a consistent shared story with teammates across independent stateless calls, and this looks like exactly that gap.

Reliability, with the 10x budgets

31/1,417 calls failed (2.2%), effectively all transport errors (rate limits), auto-retried without issue. Zero timeouts, despite max_tokens jumping 10x across the board — no sign that the bigger budgets caused the model to reason itself past the (also-raised) 300s timeout. Cost remained $0.00 throughout, as expected.

Bottom line

This game produced the strongest, highest-stakes confirmed bluff in the series so far — a fabricated investigation result that actively protected Hitler and was never caught — alongside a genuinely new failure mode (confident fabrication for a call that itself never completed) worth a follow-up prompt fix, and a data point suggesting fascist teammates don't reliably keep a consistent shared story even when both are lying in the same team's interest.