Operator discovered_7.6

content — ad-hoc head-set 7.6 (discovered candidate; provided, not behaviourally re-found)

Deep dossier (GPT-2) — operator_dossier.py --op discovered_7.6

A · identity (ad-hoc head-set (provided discovered candidate; signal = its mass on the op’s reference pattern)): heads [‘7.6’]. ranked: 7.6 (0.05)

B · causal × tasks (* = beyond random control): generic -0.00, induction +0.57, copy_names +1.05, successor -0.02, ioi +0.02 → serves [‘induction’, ‘copy_names’]

C · channels (reader 7.6): KEY/match top 1.9 collapse +22% (concentration -93.3×); VALUE/move top 5.9 ΔV-out 0.52 (median 0.10).

D · composition: IN→key 6.2(0.063), 5.9(0.054), 4.2(0.053), 3.5(0.052); OUT→value 9.3(0.061), 8.5(0.057), 8.7(0.055), 8.10(0.047).

E · redundancy (task induction): solo 7.6(+0.57); cumulative 1h +0.57 → BOTTLENECK (one head ≈ whole op).

G · SAE operands: NOT RUN — needs a SAE (sae_lens / Gemma Scope); the op’s feature-space read/write operands are the next layer.

Data: runs/disassembly/operators/dossiers/discovered_7.6/ + the catalog. Regenerate: operator_catalog_doc.py.