Operator negative_mover
output — copy-suppression / negative name-mover: writes against the copied token
GPT-2 literature DLA head-set: 10.7, 11.10. The RoPE head-set is now found behaviourally (below).
Cross-model (found behaviourally — IOI dossier)
The literature head-set above is GPT-2. The unified ResidualVM locates this operator by the ablation sweep (end→IO heads whose removal raises the IO−S logit-diff = copy-suppression) in every model (cross-model IOI dossier) — so it is no longer GPT-2-only:
| model | heads (top) |
|---|---|
| gpt2 | 10.7, 9.6, 11.10, 10.1 |
| gpt2-medium | 18.9, 22.14, 16.15, 20.7 |
| gpt2-large | 32.0, 26.0, 29.17, 30.0 |
| gemma-2-2b | 22.4, 22.0, 23.7, 18.5 |
| Llama-3.2-1B | 12.15, 15.12, 12.29, 10.25 |
| Qwen2.5-1.5B | 23.8, 25.4, 22.7, 27.11 |
SAE-feature operands (section G)
What this operator reads/writes in feature space (monosemantic SAE latents), via the per-layer GPT-2 SAEs / Gemma Scope — see the full SAE-operand table. READ = dominant key-feature where the head attends (glossed by top tokens); copy-score = OV→unembed on those tokens (+ copies / − suppresses). Provisional, single corpus; for positional/addressing ops the read-feature is incidental.
| model | head | reads (SAE feature) | copy-score |
|---|---|---|---|
| gpt2 | 10.7 | 'd/_not/_and; And/That/_and; _Citizen/_citizens |
-0.01 (≈neutral) |
Deep dossier (GPT-2) — operator_dossier.py --op negative_mover
A · identity (circuit op — heads from literature (DLA-defined; not attention-mask-readable)): heads [‘10.7’, ‘11.10’]. ranked: 10.7, 11.10
B · causal × tasks (* = beyond random control): generic +0.00, induction -0.28, copy_names -0.46, successor -0.02, ioi -0.54 → serves none
C · channels: output/circuit op — carried by OV→unembedding, not a key/value match (see composition out-edges).
D · composition: IN→key 3.0(0.054), 1.9(0.050), 1.6(0.050), 5.11(0.049); OUT→value 11.3(0.044), 11.10(0.044), 11.8(0.032), 11.11(0.032).
E · redundancy (task ioi): solo 11.10(-0.16), 10.7(-0.21); cumulative 1h -0.16 → 2h -0.54 → DISTRIBUTED population (full -0.54 ≫ best single -0.16).
Data: runs/disassembly/operators/dossiers/negative_mover/ + the catalog. Regenerate: operator_catalog_doc.py.