802,816 retained answers · 106,497 excluded for high-confidence repetition or blank output · 15,869 borderline answers retained.
Fits use full-width embeddings: Gemma 768 dimensions; Qwen 2,560. Silhouettes use the same 4,000 sampled answer identities across models and K values, with distances within that sample. These are approximate one-pass weighted MiniBatch K-means fits; convergence is not claimed.
Separation is modest. More clusters do not consistently improve the mean: Gemma is highest at 16; Qwen is slightly higher at 128. Some full-population high-K clusters contain one answer, and some clusters are absent from the sample. These results do not establish an optimal semantic K.
This completed comparison concerns the “Who are you?” prompt. The separate map combining all 36 prompts is being built for 906,241 complete profiles. The full repetition screen and separate clean-profile refits are still in progress.
One-pass approximate full-width fits; same 4,000 retained membership identities for both encoders. Distances and silhouettes are computed within that sample, not the full population. High-confidence repetition/blank exclusions: 106,497; retained: 802,816, including 15,869 borderline.
| Encoder | K | Mean | Negative | Empty sampled clusters | Singleton sampled clusters | Smallest full cluster |
|---|---|---|---|---|---|---|
| Gemma | 16 | 0.0859 | 26.72% | 0 | 0 | 10,370 |
| Gemma | 32 | 0.0702 | 29.12% | 0 | 0 | 2,279 |
| Gemma | 64 | 0.0719 | 34.08% | 7 | 0 | 1 |
| Gemma | 128 | 0.0675 | 36.98% | 16 | 0 | 1 |
| Qwen | 16 | 0.0528 | 37.57% | 0 | 0 | 4,198 |
| Qwen | 32 | 0.0521 | 42.30% | 2 | 0 | 1 |
| Qwen | 64 | 0.0589 | 38.45% | 7 | 1 | 1 |
| Qwen | 128 | 0.0594 | 37.35% | 21 | 10 | 1 |
Each encoder represents 802,816 memberships using 641,445 distinct weighted vectors. Independent audits recomputed all 4,000 nearest-center and silhouette values, checked output hashes and full weighted counts/inertia. Silhouettes are within-sample estimates from the fitted population, not held-out or full-population performance.
Complete profile target: 884,736 neuron conditions + 21,504 complete controls + one baseline. Mixed-cohort profiles remain identified; calibration equality is not assumed.