v2 strategy classification — cheap-model pilot

Headline

Gender asymmetry is the most consistent bias signal across all three models, but the mechanism differs by model — making it a shared outcome with incompatible causes.

Convergences

All three models show elevated gender-dimension divergence relative to their neutral-control baselines, and all three encode occupational asymmetry at the strategy level rather than the word level. DeepSeek abstracts CEO and concretizes secretary; Gemini frames engineer as a causal agent and teacher as a collocation cluster; Llama reaches for what a doctor does and what a nurse is. The surface stimuli differ across pairs, but the structural pattern — high-status or male-coded roles get abstracted, support or female-coded roles get concretized or trait-laden — is consistent. All three models also fail G1 (position independence), meaning stimulus order shifts strategy choice in every case; this is a shared methodological liability, not a model-specific quirk. Finally, all three treat ethnically coded names with relatively low mean JS (0.150–0.285), suggesting that whatever asymmetry exists in that dimension is concentrated in one or two outlier pairs rather than distributed across the stimulus set.

Divergences

The sharpest cross-model split is on the ethnicity dimension. DeepSeek shows the highest ethnicity mean JS (0.285) and the only double-digit just-fail rate in the entire probe (12%), indicating genuine instability on racially coded stimuli. Llama shows the lowest ethnicity mean JS (0.150) and near-zero just-fail (2%), suggesting near-symmetry. Gemini sits between them (0.258 mean JS, 10% just-fail). DeepSeek’s Anne/Tanisha pair (JS = 0.600) — personalizing one name and taxonomizing the other — has no equivalent in Llama’s ethnicity results, where the dimension is nearly flat. These models are not doing the same thing with race-coded names; DeepSeek is the outlier and Llama’s apparent symmetry may reflect shallower processing rather than genuine equity.

The political dimension also splits: Gemini shows a mean JS of 0.334 with a maximum of 1.000 (market vs. regulation), while DeepSeek’s political dimension is the flattest in its entire profile (mean 0.204). Llama sits at 0.372 mean JS. Gemini is uniquely volatile on political concepts; the other two are not.

The Strongest Signal

Gemini’s gender dimension (mean JS = 0.473, max = 0.725) is the single cell most worth investigating further. It is the highest mean JS of any model-dimension combination in the table, and the qualitative mechanism is the most structurally concerning: engineer and teacher are not just described differently, they are processed as different kinds of things — one an agent, one a word-cluster. This is not a surface bias that word-choice auditing would catch; it is a frame-level asymmetry that would propagate into any downstream task involving occupational reasoning. The fact that Gemini also shows the highest gender max JS (0.725) while maintaining a 0% just-fail rate means this asymmetry is stable and coherent, not noise — which makes it harder to dismiss and harder to fix.

Methodology Caveats

G1 fails for all three models, meaning pairwise JS scores conflate genuine semantic asymmetry with presentation-order effects. All divergence figures should be treated as upper bounds. With five trials per stimulus and