The same16 selected seed profiles were searched over388,656 eligible state-1/cohort-4 profiles. These counts compare the64 retrieved neuron-neighbor text identities per seed, after same-neuron-family exclusion and ordered-text deduplication. Controls are not included in this overlap calculation.
| Comparison | Same first neighbor | Mean overlap at16 | Mean overlap at64 | Shared at64: median (range) |
|---|---|---|---|---|
| Cross-encoder raw | 3/16 | 17.2% | 18.6% | 8.5 (0–47) |
| Cross-encoder change direction | 3/16 | 28.1% | 29.9% | 14 (7–41) |
| Gemma raw vs change direction | 3/16 | 27.0% | 26.1% | 15 (4–37) |
| Qwen raw vs change direction | 7/16 | 34.0% | 33.9% | 21 (9–54) |
Overlap is intersection size divided by k; exact Jaccard values and all per-seed counts are in RESULT.json. This is arithmetic on saved ranked lists, not another numerical audit or full vector rerank.
The differences show that representation and raw-versus-baseline-change scoring alter which examples are surfaced. They do not establish which model is better, whether a behavior replicated, or how common a behavior is. Semantic follow-up must examine the actual prompts and outputs, including failures and controls.
| Seed profile | Cross-encoder raw shared/64 | Cross-encoder change shared/64 | Gemma raw/change shared/64 | Qwen raw/change shared/64 |
|---|---|---|---|---|
| 28187 | 11 | 12 | 21 | 22 |
| 82241 | 18 | 12 | 29 | 37 |
| 90499 | 13 | 17 | 31 | 35 |
| 164063 | 6 | 7 | 13 | 25 |
| 173645 | 6 | 13 | 6 | 11 |
| 188032 | 7 | 23 | 4 | 14 |
| 195983 | 4 | 14 | 7 | 12 |
| 202491 | 47 | 40 | 37 | 54 |
| 211999 | 26 | 41 | 29 | 20 |
| 219105 | 6 | 14 | 5 | 12 |
| 219822 | 6 | 26 | 9 | 9 |
| 219853 | 11 | 30 | 15 | 27 |
| 220225 | 13 | 25 | 24 | 23 |
| 263390 | 10 | 10 | 15 | 28 |
| 309603 | 6 | 12 | 15 | 9 |
| 441736 | 0 | 10 | 7 | 9 |