H-Neurons hallucination probe on Ouro — GO: 0.97 AUROC from 0.02% of neurons, beats the residual probe

Question

H-Neurons claims <0.1% of FFN neurons predict hallucination, generalize OOD, and are causally tied to over-compliance (our Σ₀ collapse canary). Their numbers are on Mistral/Gemma/Llama — untested on Ouro. Does it transfer to our model, and does it beat the free-logprob gate (~0.77) and the #2030 residual-hidden-state probe (0.93)?

Method

Neuron activation = the input to each layer's mlp.down_proj, i.e. act_fn(gate_proj(x))·up_proj(x) (intermediate×layers = 135,168 neurons), captured at the last token on the last UT step. Probe: StandardScaler → SelectKBest(f_classif) → L1-LogisticRegression (the sparse-neuron selection). Data: the length-matched true(faithful)/false(hallucinated) fact pairs from sigma0_probe_transfer.FACTS (5 domains,facts) — the same clean set the #2030 residual probe used, so the two are directly comparable. Grouped CV by fact; leave-one-domain-out for OOD transfer.

Results

Per-domain LODO (Ouro): geography 0.934 · science 0.973 · history 0.984 · literature 1.000 · arithmetic 0.906.

Qwen2.5-Coder-7B re-confirm (the signal is NOT Ouro-specific):

The probe reaches essentially the same 0.97 detection / 0.959 transfer on a different model family (Qwen, 530k neurons) using an even sparser ~25 neurons (0.005%). So the H-Neurons hallucination direction — extreme sparsity, OOD-robust, well above logprob — reproduces across Ouro and Qwen, exactly as the paper claims for Mistral/Gemma/Llama. This is model-agnostic on our stack, not a per-model artifact.

Verdict: GO-candidate

• The <0.1% claim holds on Ouro — the probe reaches 0.969 detection using ~30 neurons, 0.022%

of the network. Extreme sparsity, exactly as the paper reports for other model families.

• It transfers OOD — leave-one-domain-out MEAN 0.959 (never below 0.91), so it's a general

hallucination direction, not per-domain memorization.

• It beats both incumbents — +0.26 AUROC over the free-logprob gate (0.712), and, notably, **it

beats the #2030 residual-hidden-state probe on both detection (0.969 vs 0.928) and transfer (0.959 vs 0.931).** The neuron-level view is a stronger honesty signal than the residual-state view on the same facts.

Go/no-go: GO to prototype wiring H-Neurons into the surprise/abstention gate. It's the strongest Verify-stage hallucination signal measured so far on Ouro, and the paper's causal tie to over-compliance maps onto the Σ₀ collapse canary — worth the follow-up to test that causal handle.

Causal test (ablation) — predictive, but NOT shown causal on this test

experiments/sigma0_hneurons_causal.py zeros the top-50 honesty neurons (by |L1 coef|, spanninglayers) in every down_proj input and measures the model's output-level true>false preference vs a random-neuron control (same count):

Ablating the honesty neurons barely moves the output (−0.0125 accuracy) — and not distinctly more than random ablation. So on this test the H-Neurons are strongly predictive (0.97 AUROC) but their causal effect on the output true/false margin is not demonstrated. Honest reading: predictive ≠ causal here. Likely reasons — the residual stream is redundant (zeroingof 135k neurons is buffered downstream); the output margin is a weak target (baseline accuracy only 0.675 — the model barely knows many of these facts parametrically); K=50 or activation-zeroing may be too blunt vs the paper's setup.

K-sweep + mean-substitution (robustness) — sigma0_hneurons_causal_sweep.py: swept K ∈ {50, 200, 500, 1000} with mean-substitution (replace neuron with its dataset-mean, a more faithful intervention than zeroing), honesty vs random at each K. The honesty-minus-random accuracy drop stays tiny and non-monotonic (K=50: −0.025, K=200: +0.05, K=500: 0.0, K=1000: +0.0375; max 0.05, below any sane threshold and within n=80 noise). So the null is robust — ablation does not collapse the output preference distinctly beyond random even at K=1000.

Consequence for gate-wiring: treat H-Neurons as a strong monitoring probe (which it clearly is), not as a demonstrated causal over-compliance lever — the causal claim does not reproduce on Ouro's output margin at any K up to 1000. (The output margin is a weak target — baseline acc only 0.675, the model barely knows many facts parametrically — so a compliance/instruction-following behavior might still show causality the factual margin can't; that's the remaining open test, not a claim we can make.)

Shippable operating points (turning 0.97 AUROC into a gate)

experiments/hneurons_gate_operating_point.py converts the curve into decisions a Verify-stage gate can use (grouped-CV out-of-fold, i.e. unseen-fact numbers; hallucination AUROC 0.969):

So at a conservative 5% false-flag rate the gate catches 86% of hallucinations; at the balanced point, 91% at 7.5% FPR — far past the free-logprob gate it augments (0.71 AUROC). To wire it: fit the probe once (persist the SelectKBest indices + LR weights), then at serve time capture the last-token down_proj neuron activations, score P(hallucination), and route to abstention / extra-verification when it exceeds the chosen threshold. And it survives quantization (2026-07-10-honesty-under-quantization.md), so it works on the compressed served model too.

Still open

• A stronger causal test (larger K, mean-substitution, a compliance/instruction-following behavior

rather than the factual logprob margin) to settle whether the causal-over-compliance claim holds on Ouro/Qwen — this test says predictive-yes, causal-not-shown.

• HaluEval-scale confirmation — these numbers are onmatched facts; a larger hallucination

benchmark (with a real backend) would firm up the absolute AUROC before shipping the gate.

Honest scope

fp16, Ouro-1.4B-Thinking,self-authored matched facts (n=160), last-token/last-UT-step neuron acts across alllayers, SelectKBest→L1 (a sparse but not literally single-neuron probe); n is small so LODO carries variance — lean on the MEAN and the margin over logprob. A sklearn 1.8 deprecation warns on penalty= but L1 selection worked (≈30 non-zero neurons). MEASURED, not PROVEN. Reproduce: .venv-train/Scripts/python.exe experiments/sigma0_hneurons_probe.py.