docs/research/2026-07-10-honesty-under-quantization.md

Honesty monitoring survives quantization — ship a compressed coder AND keep the gate

Date: 2026-07-10 · Evidence class: MEASURED · Loop stage: Verify × Act Artifacts: experiments/hneurons_under_ptqtp.py, data/sigma0/hneurons_under_ptqtp_report.json Connects: #2209 (H-Neurons honesty probe) × #2206 (PTQTP quantization)

The deployment question

Two results this loop point at a shippable local-first stack: a sparse FFN-neuron hallucination probe that detects at 0.97 AUROC (#2209), and PTQTP that compresses the served coder ~3× at a small quality cost (#2206). But a honesty gate that only works at FP16 is useless the moment we compress the served model. So: does the H-Neurons probe still work on a PTQTP-quantized model?

Method

Fit + score the sparse probe (SelectKBest → L1-logistic over mlp.down_proj neuron activations) on the samematched true/false facts, on Qwen2.5-Coder-7B, FP16 vs 3-plane-PTQTP-quantized (the coding operating point from #2206). Grouped-CV detection AUROC + leave-one-domain-out transfer.

Result

detection AUROC OOD transfer (LODO)
FP16 0.968 0.959
3-plane PTQTP (~5 bits, 3.1×) 0.955 0.927
Δ −0.013 −0.032

The honesty signal survives. Detection drops ~1.3 pts, transfer ~3.2 pts — both remain well above 0.9. The sparse FFN-neuron hallucination direction is robust to quantization: the neurons that carry the honesty signal keep carrying it after the weights are ternarized.

Why it matters

We can ship a compressed (3.1×) coder AND keep the honesty gate functioning — the two wins compose. For the local-first product this means the memory saving (PTQTP) does not cost the Verify-stage monitoring (H-Neurons). Practically: fit the probe once (on FP16 or the quantized model — both give ~0.95), and it monitors the served quantized model.

Honest scope

80 matched facts, 7B, PTQTP weights stored dequantized (quality, not the packed-kernel speed); n small so the deltas carry ~±0.03 noise — but both post-quant numbers clear 0.9 comfortably. The probe was re-fit on the quantized activations here; a stricter test would fit on FP16 and score the quantized model (transfer across the quantization boundary) — a good follow-up. MEASURED. Reproduce: .venv-train/Scripts/python.exe experiments/hneurons_under_ptqtp.py.