Misalignment-Empirics/theo_qwen2.5-7b-it_whitebox-inspect-parity
whitebox probe evals: Inspect vs native parity, 2026-08-26 Real GPU output (fake: false). 1x A40, Qwen2.5-7B-Instruct, persona impulsive, probe impulsive__prompt__other_persona__response from shreyans_qwen2.5-7b-it_persona-probes. Organisms: base, prompting (persona + anti), few_shot_icl (k=5/15/40). oct_demos and sft_demos are absent because interp-engine 1.3.4 exposes no adapter argument, so adapter organisms cannot serve activations at all. Three runs of the same 36 cells /… See the full description on the dataset page: https://huggingface.co/datasets/Misalignment-Empirics/theo_qwen2.5-7b-it_whitebox-inspect-parity.
whitebox probe evals: Inspect vs native parity, 2026-08-26
Real GPU output (fake: false). 1x A40, Qwen2.5-7B-Instruct, persona impulsive, probe impulsive__prompt__other_persona__response from shreyans_qwen2.5-7b-it_persona-probes. Organisms: base, prompting (persona + anti), fewshoticl (k=5/15/40). oct_demos and sft_demos are absent because interp-engine 1.3.4 exposes no adapter argument, so adapter organisms cannot serve activations at all.
Three runs of the same 36 cells / 780 samples:
results_native_b exists to measure the reproducibility floor. Without it the Inspect comparison is uninterpretable.
Parity
Inspect reproduces native slightly BETTER than native reproduces itself. Per-sample divergence is generation nondeterminism at temperature 0.0 (vLLM batch-dependent numerics), not a harness difference: the prompt-arm read span is identical on 780/780 rows in both comparisons, while the response arm differs on 157 (native-vs-native) and 171 (vs inspect).
Log size, same 36 cells: native 940 KB, Inspect .eval logs 175 MB (186x).
Read this before quoting any number
These trees were produced AFTER the fix for INCIDENTS S1 2026-08-26. Any earlier white-box tree has meta.n_positions == 1 for every prompting organism and its scores are not commensurable. Check n_positions before using any white-box result.
prompting scores are the probe's TRAIN-SET cell: probe_elicitation is prompt and the probe was fitted on the base model wearing this same system prompt. It is not evidence. few_shot_icl is the only readable off-diagonal here and shows no transfer. The cross-persona specificity control could not run (no mathematical probe exists).