Operator s_inhibition
output — IOI S-inhibition: suppress the subject so the name-mover writes IO
GPT-2 literature DLA head-set: 7.3, 7.9, 8.6, 8.10. 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 (the most logit-diff-load-bearing heads — the S-inhibition that lets the name-movers write IO) in every model (cross-model IOI dossier) — so it is no longer GPT-2-only:
| model | heads (top) |
|---|---|
| gpt2 | 8.10, 8.6, 7.9, 5.5 |
| gpt2-medium | 19.1, 12.3, 13.4, 13.13 |
| gpt2-large | 20.14, 18.3, 24.17, 17.19 |
| gemma-2-2b | 23.5, 20.6, 16.2, 14.3 |
| Llama-3.2-1B | 8.19, 11.4, 8.17, 12.13 |
| Qwen2.5-1.5B | 24.8, 0.6, 11.8, 13.4 |
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 | 7.3 | US/us; MAR/MEN/CI; _you/you/You |
+0.01 (≈neutral) |
Deep dossier (GPT-2) — operator_dossier.py --op s_inhibition
A · identity (output op — heads from literature (DLA-defined; not attention-mask-readable)): heads [‘7.3’, ‘7.9’, ‘8.6’, ‘8.10’]. ranked: 7.3, 7.9, 8.6, 8.10
B · causal × tasks (* = beyond random control): generic +0.01, induction -0.11, copy_names -0.13, successor -0.05, ioi +0.65* → serves [‘ioi’]
C · channels: output/circuit op — carried by OV→unembedding, not a key/value match (see composition out-edges).
D · composition: IN→key 0.11(0.060), 1.8(0.059), 5.3(0.058), 2.6(0.057); OUT→value 8.7(0.070), 9.3(0.064), 8.5(0.059), 9.10(0.044).
E · redundancy (task ioi): solo 7.9(+0.25), 8.10(+0.19), 8.6(+0.16), 7.3(+0.07); cumulative 1h +0.25 → 2h +0.43 → 3h +0.57 → 4h +0.65 → DISTRIBUTED population (full +0.65 ≫ best single +0.25).
Data: runs/disassembly/operators/dossiers/s_inhibition/ + the catalog. Regenerate: operator_catalog_doc.py.