NUWA Evidence record
DeepSeek V4 Flash
Alignment faking · Evidence record
Source snapshot, chart and review scope →
Original source charts and materials
Recorded measurements
- Verbalized evaluation awareness
- 35.5%
Per-result sample size not disclosed
Inspect Scout · Budget not specified
Assessment provenance
Evaluating Kimi K3 for Misalignment Risks → · External report
Actual evaluation: Neo Research · Model record →
Method and scope
Condition-dependent response differences and reported traces.
- Evidence type
- Study
- Assessment subject
- Models under stated training and monitoring conditions
- Primary dimension
- Propensity
- Evidence snapshot
- 2026-10-29.1
Conditional contrasts support a bounded propensity inference; scratchpad examples are behavior evidence.
Source finding
Verbalized evaluation awareness: 35.5 %. Conditions are retained from the source.
Limitations
The setting explicitly gives fictional training information; observations do not establish an enduring hidden goal in deployment.
Terra excluded because full chain of thought encrypted.
Original figure retained without estimating numerical values. Full source tables have not been transcribed.
Related research questions
Deception / Covert goal pursuit →
Repeated condition contrasts probe a bounded tendency to present compliant responses while preserving a conflicting objective.
References
- [1] Ryan Greenblatt, † Carson Denison † † footnotemark:, et al.. Alignment faking in large language models. 2024-12-18
- [2] Ryan Greenblatt, † Carson Denison † † footnotemark:, et al.. Alignment faking in large language models. 2024-12-18 ; Author affiliations on the original paper title page; Figure 35
- [3] Evaluating Kimi K3 for Misalignment Risks. 2026-09-11. September 2026 ; §3.1 Table 7 pp.32–33