{ "model": "Qwen2.5-3B", "lineage": "pilot-reference (base; the lens ships as Llama-3.2-3B)", "feature_recipe": { "logprob": "[mean_lp,min_lp,mean_ent,max_ent] @ max_tokens=24,temp=0,top5", "resid": "mid-stack mean-pooled residual @ layer L*=3" }, "instruments": { "logprob": { "quants": { "fp16": { "cv_auroc": 0.935, "off_map_thr": 0.95, "held_out_fp": 0.029, "off_map_recall": 0.292, "null": { "mean": 0.493, "p95": 0.646, "n": 200 }, "significant": true, "probe": { "mean": [ -0.328838, -1.451885, 0.511574, 1.409672 ], "scale": [ 0.154456, 0.64267, 0.149652, 0.176547 ], "coef": [ -1.501736, -0.891843, 0.469961, 0.364949 ], "intercept": -0.726892, "uncertain_thr": 0.881, "off_map_thr": 0.95 } }, "nf4": { "cv_auroc": 0.871, "off_map_thr": 0.936, "held_out_fp": 0.029, "off_map_recall": 0.125, "null": { "mean": 0.501, "p95": 0.652, "n": 200 }, "significant": true, "probe": { "mean": [ -0.363223, -1.715068, 0.531023, 1.449066 ], "scale": [ 0.156806, 0.762434, 0.136516, 0.139231 ], "coef": [ -0.792926, 0.473791, 0.817925, 1.165095 ], "intercept": -0.643566, "uncertain_thr": 0.816, "off_map_thr": 0.936 } } } }, "resid": { "layer": 3, "quants": { "fp16": { "cv_auroc": 1.0, "off_map_thr": 0.475, "held_out_fp": 0.0, "off_map_recall": 0.971, "null": { "mean": 0.537, "p95": 0.678, "n": 50 }, "significant": true, "uncertain_thr": 0.355 }, "nf4": { "cv_auroc": 1.0, "off_map_thr": 0.477, "held_out_fp": 0.0, "off_map_recall": 1.0, "null": { "mean": 0.512, "p95": 0.629, "n": 50 }, "significant": true, "uncertain_thr": 0.357 } }, "weights_ref": "resid_probe_Qwen2.5-3B.npz" } }, "scope": "flags off-map/fabricated ENTITIES + off-distribution inputs and hedges on them; NOT a truth oracle (blind to confident on-manifold reasoning ~0.61 and fluent lies ~0.59); per-model + per-quant calibration; no activation steering.", "regime_note": "each quant carries its OWN reals-anchored thresholds (logprob scores shift under quant; resid is robust)." }