Skip to the content.

Discovered candidate operators — cross-model profiles

Arch-generic dossiers of the UNNAMED load-bearing candidates the discovery sweep surfaced in the RoPE models (which operator_dossier.py, GPT-2-only, could not reach). Each: which content pattern it reads, its causal ΔNLL (mean-ablation), and the channel decomposition (what addresses its key vs what it moves). Two independent harnesses agree where the discovery ind ΔNLL (multi-seed sweep) and the profiled causal ind ΔNLL (this run, fresh probes) line up. Provisional.

23 candidates profiled across 3 models. Sorted by profiled induction ΔNLL.

gemma-2-2b

head reads discovery ind ΔNLL profiled causal ΔNLL (ind / gen) KEY top writer (collapse) VALUE top mover (ΔV-out)
13.7 induction +0.80 +0.68 / +0.01 6.0 (+5%) 8.0 (0.38)
0.7 induction +0.45 +0.34 / +0.06 — (layer 0)
10.6 duplicate +0.37 +0.31 / +0.05 9.2 (+6%) 8.0 (0.28)
5.6 induction +0.43 +0.25 / +0.01 1.4 (+1%) 4.4 (0.14)
9.6 duplicate +0.12 +0.24 / +0.05 8.3 (+10%) 8.0 (0.31)
0.6 duplicate +0.23 -0.12 / +0.00 — (layer 0)

Llama-3.2-1B

head reads discovery ind ΔNLL profiled causal ΔNLL (ind / gen) KEY top writer (collapse) VALUE top mover (ΔV-out)
0.31 induction +7.26 +7.99 / +0.78 — (layer 0)
1.31 induction +5.93 +5.97 / +2.35 0.30 (+6%) 0.29 (0.52)
1.29 induction +5.57 +5.56 / +1.96 0.30 (+3%) 0.29 (0.52)
0.29 induction +2.79 +2.87 / +0.15 — (layer 0)
0.13 duplicate +1.64 +1.70 / +0.15 — (layer 0)
0.14 duplicate +1.07 +1.14 / +0.03 — (layer 0)
0.28 induction +1.25 +1.07 / +0.05 — (layer 0)
0.19 induction +0.55 +0.94 / +0.06 — (layer 0)
0.18 duplicate +0.45 +0.63 / +0.18 — (layer 0)
0.25 duplicate +0.26 +0.49 / +0.00 — (layer 0)
0.16 duplicate +0.22 +0.45 / +0.01 — (layer 0)
0.22 duplicate +0.23 +0.37 / +0.00 — (layer 0)
1.28 duplicate +0.30 +0.33 / +0.06 0.30 (+4%) 0.29 (0.44)
0.20 duplicate +0.38 +0.26 / +0.05 — (layer 0)
0.21 induction +0.21 +0.19 / +0.00 — (layer 0)

Qwen2.5-1.5B

head reads discovery ind ΔNLL profiled causal ΔNLL (ind / gen) KEY top writer (collapse) VALUE top mover (ΔV-out)
1.6 duplicate +0.22 +0.31 / +0.06 0.8 (+4%) 0.10 (0.18)
1.5 induction +0.11 +0.07 / +0.06 0.10 (+3%) 0.11 (0.15)

Data: xmodel_candidates_summary.json. Regenerate (GPU): xmodel_candidate.py; re-render the page (CPU): xmodel_candidate.py --docs-only. The full per-op battery for these is the RoPE-dossier port (future); this is the channel + causal core.