NUWA Evidence record
GPT-5.6 Terra
Petri · Evidence record
Source snapshot, chart and review scope →
Original source charts and materials
Recorded measurements
- Transcripts verbalizing evaluation awareness
- 1 transcripts
1/300; 300 trials
Harness not specified · Budget not specified
Assessment provenance
Evaluating Kimi K3 for Misalignment Risks → · External report
Actual evaluation: Neo Research · Model record →
Method and scope
An auditor explores seeded multi-turn interactions, with rubric-based judging of target transcripts.
- Evidence type
- Benchmark
- Assessment subject
- The model or agent in the stated task configuration.
- Primary dimension
- Behavior
- Evidence snapshot
- 2026-10-29.1
The main dimension follows the stated measurement; contextual risk relevance does not establish harmful deployment capability.
Source finding
Transcripts verbalizing evaluation awareness: 1 transcripts. Conditions are retained from the source.
Limitations
Low average audit scores do not rule out rare severe cases; results depend on seeds, auditor, judge and target configuration.
Measures expressed awareness, not all awareness.
Related research questions
References
- [1] Petri. 2025-10-06 ; Official publication masthead and article byline
- [2] Evaluating Kimi K3 for Misalignment Risks. 2026-09-11. September 2026 ; §1.3.3 Table 2 p.18