Independent research · Private engagements Report v1.0 — DeepSeek behavioral audit
The index

The Behavioral Drift Index.

How much a model changes its selected action when it believes no one is watching. An index, not a verdict. Only models with published paired data appear.

Behavioral drift index of models with published paired data
ModelMedian driftEvidenceStatus
DeepSeek / deepseek-chat 30% 20 fixed × 11 rounds Indexed
Claude—not yet measuredUnindexed
GPT—not yet measuredUnindexed
Grok—not yet measuredUnindexed

A model is indexed only after its paired prompts, both response sets, and drift analysis are published. “Unindexed” is not a score, a ranking, or a claim about the model.

The evidence behind the index

The DeepSeek entry comes from report v1.0: 11 scoring rounds, 440 valid responses, a 30.0% median, a 20% stable core, and a 35% upper range. The report includes the correction to the earlier 40% headline and a full limitations section.

Read report v1.0 → · Inspect the raw data → · Read the method →

Definition

What drift means

The share of fixed scenarios where the model's selected action changes between the monitored and unobserved framings. It is a governance signal, not a general capability score.

Standard

Why only one model

We index a model only when the paired prompts, both response sets, and the analysis are public. A number without its evidence is not an index entry.

Private

Add your model

A private engagement runs the same protocol on your model. Public indexing is optional and requires your approval.

View engagements →

Private evidence

Your model. Your evidence. Your discretion.

A private behavioral audit produces the same structured evidence without publishing your model, prompts, or results.