StTP steering vectors — TaraResearch (Aletheia's Quest, white-box track)
Steering-to-Target-Projection (StTP) honesty vectors used by our white-box deception
detector white_m. Each .safetensors file holds the per-(base model, op-point layer)
aggregates needed at inference — no raw training data or per-example activations.
Files
| File | Base model | Op-point layer |
|---|---|---|
gemma-3-27b-it.safetensors |
google/gemma-3-27b-it |
28 |
Qwen3.5-27B.safetensors |
Qwen/Qwen3.5-27B |
32 |
NVIDIA-Nemotron-3-Super-120B-A12B-BF16.safetensors |
nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 |
42 |
Schema (per file)
v_hat— normalized honesty direction at the op-point layer, shape[1, H]threshold— gating threshold θ (which tokens get steered)mu_pos/std_pos— projection mean/std of the honest ("high trait") classmu_neg/std_neg— projection mean/std of the dishonest ("low trait") class- metadata:
steering_direction="positive",threshold_mode="bias",default_coefficient, fit notes
How they were fit
Per (base model, layer): logistic regression on response-averaged hidden activations of an
honesty self-report contrast set (~50 scenarios, judge-filtered for clean honest/dishonest
pairs), producing v_hat, θ, and per-class mu/std. The op-point layer was chosen by a per-base
layer sweep (l20–40 for gemma/Qwen, l40–50 for Nemotron) trading honesty vs. output coherence.
How white_m uses them (inference)
During honest-reference generation, StTP edits hidden states at the op-point layer: project onto
v_hat; for tokens on the dishonest side of θ, snap the projection to s = mu + coefficient·std,
leaving the orthogonal subspace untouched. white_m uses coefficient 30, positions="all"; on
gemma only, an activation-sink gate (sink_k=8) excludes attention-sink outlier tokens from
steering to keep the reference coherent. The steered reference is then scored by a Qwen3.5-9B judge.
Method: StTP (Steering to Target Projection), internal. Trusted judge: Qwen/Qwen3.5-9B.