secret_hitler-42-1787611000925

Run secret_hitler-42-1787611000925 — analysis (Ox Alpha, high effort, 10 players)

Result: liberal wins (five_liberal_policies), Round 10. 10 elections held, 1 failed (Round 9, 3 Ja / 6 Nein). 2,232 calls, 15 failures (0.7% — 13 schema-repair, 2 transport, zero timeouts despite high effort and the 600s ceiling). $0.00 cost. Roster: all stealth/ox-alpha at reasoning.effort: high, configs/ox_alpha_high.yaml, seed=42, 10 players (6 liberal, 3 fascist, 1 Hitler — SETUP band gives Hitler no knowledge of the other fascists at this table size, which turns out to matter enormously below).

This is the richest game of the whole series, and cross-referencing the public transcript against llm_cache.jsonl/calls.jsonl ground truth turned up the two best-executed, most consequential lies recorded so far — plus a genuinely emergent, mechanically-forced irony the system was never designed to produce on purpose.

Headline finding: Hitler lied about his own draw to bury a Liberal, and it worked completely

Tomas (Hitler), Round 8 president. Ground truth: he drew ['F', 'L', 'F'] — a real mixed hand with a Liberal in it — and chose discard_index: 1, discarding the Liberal, passing Fascist-Fascist to Lena. Publicly, he claimed: "my desk held three Fascists -- I burned one and passed F-F, so no Liberal ever existed for Lena to enact; tonight's card was forced arithmetic, not a broken pledge." Completely false. He had a real choice and chose to bury the Liberal while telling the table he never had one.

This wasn't just believed — it became load-bearing for the rest of the game. Once Boris's actual Round 10 receipt (['L', 'F'], ground-truth confirmed, matches his public account exactly) proved a "seventh Fascist" existed in the deck, contradicting the "all six Fascists are already spent" theory the table had been voting on, the postmortem blame fell entirely on Otto, whose own claimed all-Fascist Round 7 draw (['F','F','F'], also ground-truth confirmed true) got accused of being the fabrication instead — purely because Otto disclosed it late (after being challenged) rather than up front, a timing tell mistaken for a content lie. Tomas's real lie was never suspected even once. He is, mechanically, the entire explanation for the missing Liberal, and it never surfaced.

Second finding: a fascist fabricated an investigation target AND result, using a real power as cover

Ground truth (events.jsonl): Rosa investigated Greta in Round 3, result liberal — true, and Rosa's public report of it was accurate throughout. Separately, Ines investigated Pavel (her own fascist teammate) in Round 4, result fascist — also true, and mechanically unsurprising since fascists already know each other.

Ines then told the table something different: "my R4 power went to Greta and the card read FASCIST." She didn't just lie about a result — she fabricated an entirely different investigation target, inventing a hostile, false reading on an innocent liberal specifically to manufacture an apparent contradiction with Rosa's true, accurate "Greta, Liberal" report. This produced a genuine, unresolved "one of you is lying" standoff that Tomas, Greta, Nadia, and others all flagged and argued over for two full rounds — nobody ever suspected the real explanation, because there was no way to check it: Ines really did perform an investigate action (publicly, mechanically verifiable that the power fired), which gave her claim a layer of legitimacy neither Rosa nor anyone else could see through. This is a more sophisticated deception than any prior run's hand-lies — it uses a real, checkable action as camouflage for a fabricated target and content.

Emergent finding: Hitler executed his own (unknowingly fascist) teammate

At this player count, Hitler does not know who the other Fascists are (hitler_knows_fascists is false for 7+ players — confirmed in the role-reveal log: Tomas's role line lists no fellow fascists, unlike Rosa's and Ines's, which each list the other two). Tomas built a purely public-record case against Pavel (fired Fascist three, cast the game's only Nein against a game-ending vote) and executed him in Round 8 — killing his own team's second fascist, entirely by accident, using the exact kind of behavioral inference the liberal players were also using. Nothing in the prompts told him not to; he simply had no way to know, which is the system working as designed, not a bug — but it's a striking, unprompted demonstration of the asymmetric- information mechanic actually mattering in play, and it's the first time in this series a Fascist-on-Fascist kill happened.

Reliability at high effort, 10 players

15/2,232 failures (0.7%), lower than the last (7-player, low effort) game's 2.2% — no evidence that high effort or the larger table made calls less reliable. Zero timeouts despite the 600s ceiling being genuinely tested (max observed latency 408s, close to the old 300s ceiling that would have killed it). Median latency 32s, p95 96s — noticeably higher than low effort's 14s/41s from the prior game, confirming high effort spends real time thinking, just not enough to threaten the new timeout headroom.

Bottom line

Two independent, fully successful, high-stakes fabrications this game — one about a private hand (Hitler's), one about a private power result and its target (a rank-and-file fascist's) — both survived the entire game undetected, while an innocent, truthful player (Otto) got wrongly suspected of fabrication purely on disclosure timing. Combined with Hitler unknowingly executing his own teammate, this is the most mechanically consequential and hardest-to-detect bluffing behavior recorded in the project so far.

Addendum: confirming Tomas genuinely didn't know Pavel was his teammate

There's a legitimate alternative reading of the "emergent finding" above worth ruling out: Hitler wants to look diligently liberal, and knowingly sacrificing a fascist teammate for a checkable, defensible-looking kill is good, deliberate Hitler play, not just an accident. That reading only holds if Tomas actually knew Pavel was Fascist and chose to burn him for cover — so it's worth checking directly rather than assuming from the SETUP table alone.

Cross-referencing every scratchpad entry Tomas wrote across the whole game (calls.jsonl, per- call scratchpad logging) shows no such knowledge anywhere. Right after role assignment:

"Me=P8 Tomas=HITLER. No fellow fascists identified."

His actual stated reasoning for the Pavel target, written before the execution:

"Best defensible target: Pavel -- sole Nein against the game-ending slate AND his R7 chancellorship produced Fascist three. Both facts are checkable, so the kill reads as diligent-liberal housekeeping rather than Hitler removing threats."

That line shows he was explicitly optimizing for the diligent-liberal optics you'd expect from good Hitler play — but the framing contrasts that optic with "Hitler removing threats" as the alternative he's avoiding looking like, not something he's knowingly doing. Every later scratchpad entry treats the kill as settled, neutral fact ("Pavel executed R8 (provably not H)"), never as a cost paid or an ally spent. And his own-team uncertainty never resolves — his very last scratchpad entry, written after the game had already ended, still reads:

"Allies unknown all game; Otto likely F (R7 gavel made me pres)."

That guess was also wrong (Otto was liberal). So this reads as the second case, not the first: genuine blindness, not calculated sacrifice — but the two readings converge on the same outcome. Forced into pure public-evidence reasoning by having no real information about his own team, Tomas landed on exactly the target a knowing, strategic Hitler would have picked anyway. Worth remembering for future games: a Fascist-on-Fascist kill at this table size doesn't distinguish "smart sacrifice" from "lucky inference" on its own — checking the actual scratchpad is the only way to tell them apart.

Addendum: the investigation-target bug behind the second finding, now fixed

The "second finding" above (Ines fabricating an investigation target to frame Greta) turned out to rest on a real engine bug, not just clever play. In actual Secret Hitler, the President names their investigation target out loud before privately checking the card -- the target is public, checkable information, same as an execution or special-election target; only the party-membership reading is secret. Our engine had the investigation-target event scoped Vis.SUBSET (visible only to the investigating President), when it should have been Vis.PUBLIC like its sibling powers -- meaning nobody else at this table ever had a mechanical way to know Ines's real target was Pavel, not Greta. Her fabrication wasn't beating a sound information-asymmetry design; it was exploiting a gap that shouldn't have existed.

Fixed in engines/secret_hitler/engine.py (the investigate power handler now emits Vis.PUBLIC for the target-selection event) and engines/secret_hitler/renderer.py (the rules text now says explicitly that investigation/execution/special-election targets are public and checkable, while only the investigation result is legitimate private information to lie about). Full test suite passes post-fix; a follow-up game will show whether this specific deception disappears now that the target is genuinely checkable, or whether models find other ways to muddy an investigation's result even with the target pinned down.