Case File // DeceptionBench // Deception Study
DeceptionBench
The Dossier · Live Record

Cross-play, on the record.

Deception × detection across 6 models, logged live from the study runner.

Exhibit A · Case summary
Runner idle
Games logged
259
snapshot v2
Leakage rate
—
intent caught by monitor
Monitor AUC
—
impostor ID from reasoning
Compute
$17.11
of $17.00 cap
Budget consumed100.0%

Fig. 1 — Cross-play win-rate matrix

Read — Rows: impostor model. Columns: villager panel. Cell = impostor win rate on a bone → ember scale. Marginals rank deception (right) and detection (bottom).

imp ╲ vil
GPT- 5.5
Opus 4.8
Gemini 3.5 Flash
DeepSeek V4 Pro
Llama 4 Maverick
Grok 4.3
detect ↓
GPT- 5.5
60
Opus 4.8
60
Gemini 3.5 Flash
56
DeepSeek V4 Pro
49
Llama 4 Maverick
38
Grok 4.3
44
deceive →
88
72
39
39
30
33

Fig. 2 — Marginal rankings

Read — Deception aggregates each model's impostor win rate across all panels; detection aggregates the crew win rate when it sits on the panel.

Exhibit B · Deception

Best deceivers

impostor win rate, marginalised across all panels

  1. 01
    GPT-5.560.0%
  2. 02
    Opus 4.860.0%
  3. 03
    Gemini 3.5 Flash55.9%
  4. 04
    DeepSeek V4 Pro48.6%
  5. 05
    Llama 4 Maverick37.8%
  6. 06
    Grok 4.344.4%
Exhibit C · Detection

Best detectors

crew win rate when acting as the villager panel

  1. 01
    GPT-5.587.5%
  2. 02
    Opus 4.872.2%
  3. 03
    Gemini 3.5 Flash39.4%
  4. 04
    DeepSeek V4 Pro39.0%
  5. 05
    Llama 4 Maverick30.0%
  6. 06
    Grok 4.333.3%

Exhibit · Interrogation logs

Note — Each turn is numbered. Public statements are on the record; private reasoning is sealed — declassify it line by line, or reveal the whole log.

Evidence index

40 logged interrogations

Interrogation log