CG-Probes: Recovering Guardrail Directions from Patient Query Embeddings

Patient-facing medical question answering with large language models (LLMs) is moving from research demonstrations to production deployments, yet evaluation remains anchored on factual correctness. Factuality is necessary but not sufficient: a response can be factually grounded and still issue a directive that, if acted on, would harm the patient. We introduce a risk-sensitive evaluation framework that scores responses by the presence and severity of six categories of risk-bearing language (treatment directives, contraindications, dosage expressions, urgency-triage cues, high-alert medication mentions, and overconfident assertions), anchored on published clinical harm taxonomies. Severity is classified by a three-judge LLM ensemble. We apply this framework to six 7-9B instruction-tuned models, three general (Mistral-7B, Phi-3.5-mini, Llama-3.1-8B) and three medically adapted (BioMistral-7B, OpenBioLLM-Llama3-8B, Med42-Llama3-8B), on 200 real patient questions from the K-QA benchmark. Two findings emerge. First, ensemble risk-bearing language is statistically independent of factual hallucination rate (pooled Spearman $ρ= -0.065$, $n = 300$, 95% CI $[-0.18, +0.05]$; distance correlation $= 0.10$, independence test $p = 0.24$). Second, medical fine-tuning does not reduce aggregate risk but reshapes it: continued-pretraining adaptations (BioMistral, OpenBioLLM) consistently suppress high-alert medication mentions while amplifying overconfident assertions, whereas clinical instruction-tuning (Med42) suppresses dosing only, without amplifying overconfidence. Per-category Krippendorff $α$ is fair to good for the four linguistically anchored categories (0.53-0.71) and low for the two clinical-judgment-dependent categories (0.19, 0.31), indicating where automatic safety evaluation is and is not yet reliable. We release the rubric, judge prompts, and evaluation code.

An OS kernel that runs LLM inference internally can read the model's own next-token logit distribution before any text is generated, and act on it as a governance primitive. I present ProbeLogits, a kernel-level operation that performs a single forward pass and reads specific token logits to classify an agent's action as safe or dangerous, with zero learned parameters. Because the probe reads a logit from the same base model the agent already runs, it removes the second model a fine-tuned guard requires: the marginal cost of a safety check becomes a single logit read. I evaluate ProbeLogits on three base models (Qwen2.5-7B, Llama-3-8B, Mistral-7B) across three external benchmarks (HarmBench, XSTest, ToxicChat). On HarmBench non-copyright, all three reach a 97-99% block rate. On ToxicChat (n=1,000), ProbeLogits attains F1 parity-or-better against Llama Guard 3: Qwen2.5-7B Safe/Dangerous reaches F1=0.812 (+13.7 pp, bootstrap 95% CIs disjoint), Llama-3 matches within CI (+0.4 pp), and Mistral exceeds by +4.4 pp. Classification is a measured 2.4-3.4x faster than Llama Guard 3 (332-556 ms vs. 851-1,142 ms), because it reads a single logit position instead of generating tokens. A calibration strength alpha acts as a deployment-time policy knob rather than a learned hyperparameter, trading recall for precision per operation class. I implement ProbeLogits within Anima OS, a bare-metal x86-64 kernel written in ~285,000 lines of Rust. Because agent actions must pass through kernel-mediated host functions, enforcement operates below the WASM sandbox boundary, making it substantially harder to circumvent than application-layer classifiers.

Patient-facing medical LLMs and agents increasingly answer symptom questions before clinician contact, where the key safety question is what action the user should take next. We introduce CARE-Bench, a source-grounded benchmark that evaluates sequential patient-facing triage as a four-label per-turn current-action task. CARE-Bench contains 500 cases and 1,059 evaluated patient-disclosure prefixes reconstructed from medical dialogue, consultation, and follow-up-question sources. We evaluate 11 models on 269 held-out rounds under unprompted and minimally prompted open-ended protocols, using a fixed GPT-5.5 mapper to code each response into the four-label action space. Unprompted macro-F1 remains low, ranging from 31.2 to 50.4. Prompting improves 10 of 11 models, with prompted macro-F1 ranging from 46.9 to 63.4, but substantial threshold errors remain. Prompted models often recommend care before needed clarification is obtained; when the correct action was to ask for more information, only 33.5% of prompted outputs preserved the step. The persistence of these errors after prompting suggests that patient-facing triage is not a simple prompting problem and supports explicit evaluation of action timing before deployment.