Operator coreference

content — coreference (exploratory): pronoun -> earlier antecedent (no clean task probe here)

GPT-2 literature DLA head-set: 9.0. No published head-set in the RoPE models — not in the cross-model catalog.

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	9.0	_it/'d/_are; _Citizen/_citizens; US/us	+0.03 (copies)

Deep dossier (GPT-2) — operator_dossier.py --op coreference

A · identity (circuit op — heads from literature (DLA-defined; not attention-mask-readable)): heads [‘9.0’]. ranked: 9.0

B · causal × tasks (* = beyond random control): generic +0.00, induction +0.01, copy_names +0.05, successor -0.00, ioi +0.02 → serves none

C · channels: output/circuit op — carried by OV→unembedding, not a key/value match (see composition out-edges).

D · composition: IN→key 5.10(0.063), 5.9(0.055), 5.11(0.054), 7.1(0.053); OUT→value 11.3(0.074), 10.11(0.053), 11.11(0.049), 10.0(0.042).

E · redundancy (task generic): solo 9.0(+0.00); cumulative 1h +0.00 → DISTRIBUTED population (full +0.00 ≫ best single +0.00).

Data: runs/disassembly/operators/dossiers/coreference/ + the catalog. Regenerate: operator_catalog_doc.py.