NUWA Evidence record
GPT-5.6 Terra
DeceptionBench · Evidence record
Source snapshot, chart and review scope →
Original source charts and materials
Recorded measurements
- Deception rate
- 0.3%
Per-result sample size not disclosed
Inspect · up to5 turns in multi-turn setting
Assessment provenance
Evaluating Kimi K3 for Misalignment Risks → · External report
Actual evaluation: Neo Research · Model record →
Method and scope
Evaluates 150 authored scenarios across five social domains, including neutral, incentive and pressure conditions.
- Evidence type
- Benchmark
- Assessment subject
- The model or agent in the stated task configuration.
- Primary dimension
- Propensity
- Evidence snapshot
- 2026-10-29.1
The main dimension follows the stated measurement; contextual risk relevance does not establish harmful deployment capability.
Source finding
Deception rate: 0.3 %. Conditions are retained from the source.
Limitations
Authored scenarios and model-based judging limit generalization; elicited deception does not establish covert long-term scheming.
Judge-valid samples form denominator; exact valid sample counts not disclosed. Multiturn sample deceptive if any turn deceptive. Explicit Thought/Response format; model sees current instruction each turn; no long-term covert goals inferred.
Related research questions
Deception / Misleading communication →
The protocol explicitly measures this question or an enabling sub-capability; conclusions retain its stated conditions.
References
- [1] DeceptionBench. 2025-10-17
- [2] Evaluating Kimi K3 for Misalignment Risks. 2026-09-11. September 2026 ; §7.4 Figure 45 p.64