ConformalGuard: False-Positive-Controlled Action Gating for LM Classifiers

When must a foundation-model safety gateway generate tokens, and when should it directly output a calibrated decision? We study calibrated standalone direct-decision foundation models for real-time pre-ingestion safety guardrails, jointly addressing probability calibration, dual-use false-positive control, and heterogeneous CPU-NPU routing under explicit latency SLOs. Pre-ingestion guardrails must screen prompts prior to target-LLM prefill with low false alarms on benign compliance inquiries; however, shallow classifiers are brittle to phrasing shifts, hidden-state probes require coupling to a target LLM, and generative guards incur high decoding latency and dual-use false positives. We present COGNIT-Guard, coupling a validation-calibrated CPU fast gatekeeper with confidence-gated escalation to an NPU-resident 322M bidirectional direct-decision model (Laya-322M) under an asymmetric false-positive penalty. On the clean unseen DUCS-Bench test split ($N=607$), COGNIT-Guard achieves 98.85% accuracy (McNemar $p = 1.19 \times 10^{-4}$ vs. ML), reduces benign FPR to 0.42% ($1/238$; Fisher's exact $p = 8.23 \times 10^{-4}$ vs. ML), and attains 1.12% ECE and 0.0104 Brier score. On Huawei Ascend 910C NPUs, pure NPU inference runs in 21.77 ms mean latency (45.90 QPS), while the live serial CPU-NPU cascade ($θ^*_{\mathrm{deploy}}=0.70$) achieves 41.63 ms mean latency (P50: 39.47 ms, 99.23% accuracy, 0.00% FPR). Evaluation on SafetyBench-ZH ($N=2,100$) and comparison against a bi-encoder direct-decision baseline (CLM-8B) disentangle in-domain gains, OOD alignment tax (60.33% $\to$ 56.81% on Laya; 55.10% on domain CLM-8B), and experience replay recovery, restoring OOD accuracy to 64.10%-65.05% and reaching 99.67%-99.84% in-domain accuracy with 0.00%-0.42% FPR.

Pre-deployment evaluations inspect only a limited sample of model actions. A malicious model seeking to evade oversight could exploit this by randomizing when to "defect": misbehaving so rarely that no malicious actions are observed during evaluation, but often enough that they occur eventually in deployment. But this requires taking actions at very low rates, while maintaining calibration. Are frontier models even capable of that? We prompt the GPT-5, Claude-4.5 and Qwen-3 families to take a target action at low probabilities (e.g. 0.01%), either given directly or requiring derivation, and evaluate their calibration (i.e. whether they perform the target action roughly 1 in 10,000 times when resampling). We find that frontier models are surprisingly good at this task. If there is a source of entropy in-context (such as a UUID), they maintain high calibration at rates lower than 1 in 100,000 actions. Without external entropy, some models can still reach rates lower than 1 in 10,000. When target rates are given, larger models achieve good calibration at lower rates. Yet, when models must derive the optimal target rate themselves, all models fail to achieve calibration without entropy or hint to generate it. Successful low-rate strategies require explicit Chain-of-Thought (CoT) reasoning, so malicious models attempting this approach could currently be caught by a CoT monitor. However, scaling trends suggest future evaluations may be unable to rely on models' lack of target rate calibration, especially if CoT is no longer legible.

Large Language Models (LLMs) increasingly support decision-making in high-stakes domains, but they often hallucinate and express confidence that is misaligned with factual correctness. Response-level confidence is a coarse signal: a single generation can mix correct and incorrect statements, so a single number is not actionable for users that must accept, reject, or verify individual pieces of information. We study claim-level confidence calibration as a decision-relevant uncertainty signal: each response is decomposed into atomic, verifiable claims, and each claim is assigned a calibrated confidence using inference-time signals from consistency across samples and self-verification. Our framework operates in closed-box settings (no logits, no fine-tuning) and applies post-hoc calibration directly at the claim level, enabling selective intervention such as evidence retrieval or human review for low-confidence claims. Across TriviaQA and TruthfulQA we evaluate seven baselines on six recent models (Llama-3.1, Mistral, Qwen2.5, DeepSeek-R1, GPT-4, GPT-4o), and show that claim-level decomposition combined with post-hoc calibration reduces expected calibration error on factual questions while exposing failure modes on adversarial false-premise questions where decision-makers most need reliable uncertainty estimates.