Security teams receive far more automated alerts than people can read, so the valuable step is closing obviously benign alerts automatically, with a guaranteed low rate of closing a real attack. Language models that write out their reasoning before giving a verdict triage alerts much better than models that output the label directly. They have one serious flaw, though: once a model has argued itself into a conclusion, the probability it puts on the final label is close to 100% whether the argument was right or wrong. That makes its confidence useless for deciding what to auto-close.
This note proposes using a property of the setting that most of the hallucination literature lacks. An alert is a structured record, so many statements a model makes while reasoning about it can be checked directly against the record. Examples are "the parent process is winword.exe", "the destination port is 445", and "this binary is rare across the fleet". We train a small, domain-specialized model (1.7B–8B parameters) to write its factual claims in a checkable form. A deterministic checker verifies those claims against the alert record and fleet-wide statistics. We then use the pattern of verified and contradicted claims as the model's confidence signal and as a reward during training. The result is a single compact model whose auto-closure threshold can be set with a statistical error guarantee, with no second large model acting as a judge.
Core Claim
Our hypothesis is that, for structured alert triage, a rationale that misstates facts in the record it is reasoning about is a strong and nearly free predictor of a wrong verdict. Using it gives two things. First, combining "fraction of checkable claims that are true" with the label probability recovers the calibration that reasoning destroys, closely enough to support conformal auto-closure at 98–99% precision. Second, rewarding claim fidelity during reinforcement learning reduces hallucinated rationales without lowering verdict accuracy. If both hold, the separate calibrator model used in prior work becomes unnecessary, and the rationale shown to analysts becomes trustworthy at the level of individual facts.
The direct motivation is Cybersecurity Detection Classification with Reasoning-enabled Language Models. It trains a 30B reasoning model on Windows endpoint detections and documents this "confidence paradox". It fixes the problem with a second 30B model that reads the alert, the reasoning and the verdict, and then outputs a correct/incorrect probability. The paper also states that it never checks whether the intermediate reasoning hallucinates technical facts. We propose to answer that open question and turn the answer into the calibration mechanism. The idea of grounding each reasoning step in an external checker comes from GroundedPRM, which verifies math reasoning steps with Wolfram Alpha and gains large data efficiency. Our difference is that the "tool" is the alert record itself, which makes checking a cheap lookup rather than a search, so it fits a high-volume pipeline. The statistical layer follows TRAC, which uses conformal prediction to bound error in retrieval-augmented question answering. We apply the same kind of guarantee to a single decision: whether to close an alert without a human looking at it.
Research Program
Checkable rationale format. The model writes its reasoning freely, but every factual assertion about the alert goes inside a lightweight tag that names the field or statistic and the claimed value. A tagged lookup claim names a field, such as the parent image path or the destination port. A tagged aggregate claim names a statistic, such as fleet prevalence of a hash or how often a source talks to a destination. Statistics are computed only from the training window to avoid leakage. The checker marks each claim verified, contradicted, or unresolvable. Untagged inference ("this looks like lateral movement") is allowed and not checked; the check is only on the factual base the inference rests on. The format is taught by supervised fine-tuning on traces from a larger open teacher model, keeping only traces with correct verdicts and no contradicted claims, as in the filtered distillation of Specializing Smaller Language Models towards Multi-Step Reasoning. That work shows models of this size can learn multi-step reasoning when distilled this way.
Fidelity-aware reinforcement learning. The second training stage uses Group Relative Policy Optimization (GRPO), a reinforcement-learning method that compares several sampled answers per input. This mirrors the recipes in DynaGuard and the 30B triage paper. Our reward adds a fidelity term to label correctness: each contradicted claim is penalized.
The obvious failure mode is reward hacking, where the model makes fewer claims, or only trivial ones. We guard against it with a coverage term. Each alert type has a small set of decisive fields, derived from the attack categories in the dataset and the fields the teacher's correct traces relied on. The model is rewarded for making verified claims about those fields. We report claim counts and coverage throughout so any collapse is visible.
Calibrated decision rule. On a held-out calibration split, we fit a two-parameter logistic score $s(x)$. It combines the label log-probability with the contradicted-claim count and the verified fraction of decisive-field claims. We then choose the auto-close threshold $\hat\lambda$ by conformal risk control:
$$\hat\lambda = \min\Big\{\lambda : \frac{n}{n+1}\,\hat R_n(\lambda) + \frac{1}{n+1} \le \alpha\Big\}$$
Here $\hat R_n(\lambda)$ is the fraction of the $n$ calibration alerts that are malicious yet would be auto-closed at threshold $\lambda$, and $\alpha$ is the tolerated rate of missed attacks. This bound gives an expected-risk guarantee when calibration and test alerts are exchangeable. Security telemetry drifts, which the 30B paper explicitly reports, so the guarantee is expected to weaken over time. We measure that directly rather than assuming it away.
Baselines
The naive reference point is a direct-label model: the same backbone fine-tuned to emit only the verdict, thresholded on its label probability. Its probabilities are meaningful, which is exactly what reasoning breaks, so it shows whether reasoning is worth its cost at the precision levels that matter. On the proprietary endpoint data, Cybersecurity Detection Classification with Reasoning-enabled Language Models reports this baseline 43.0 points below the reasoning model in benign recall at 98% precision, which implies roughly 21.8% against 64.8%. We rerun it on our data with the same backbone and the same training-data budget.
The strongest existing method is the reasoning model plus a separate calibrator model from that same paper. Its reported result is 64.8% benign recall at 98% precision and 53.0% malicious recall at 99% precision, on its Windows endpoint test set of 42,686 detections. That dataset is not described as public. We will request access. If access is refused, we reimplement the pipeline at matched total parameter count, pairing an 8B triage model with an 8B calibrator on our benchmarks, because the comparison that matters is extra parameters versus free checking. Beating it means equal or better recall at 98% and 99% precision while using half the parameters and no second forward pass.
On CSE-CIC-IDS2018, the closest reported result is from (EC)2. It is a multi-agent investigation system that queries a topology graph and packet store, and it raised recall on uncertain anomaly-detector alerts from 0.54 to 0.72 while cutting false positive rate from 0.43 to 0.26, at roughly 17 minutes per event. We compare on the same uncertain-zone protocol. A result counts as competitive if it matches that recall/false-positive point at a latency several orders of magnitude lower. We found no reported recall-at-fixed-precision result on Lacuna for CSE-CIC-IDS2018, UNSW-NB15 or HIKARI-2021, so for those datasets our own reimplemented baselines are the reference.
The key ablations separate the two uses of the checker. In one variant, checked-claim scores are used for calibration only, with the model trained without the fidelity reward. In the other, the model is trained with the fidelity reward but calibrated on label probability only. A further ablation drops the coverage term to measure reward hacking.
Why not just…?
Why not just train a small calibrator or a probe on hidden states? A linear probe on the reasoning model's final hidden state is cheap and could plausibly recover much of the calibration. This is the real threat to the proposal. We include it as a baseline at the same calibration-data budget. The proposal survives only if claim checking matches or beats the probe on the operating points that matter, or if it holds up better under temporal drift. The latter is plausible because a contradicted fact is wrong regardless of distribution, while a probe's learned direction may move with the data. If the probe wins on both counts, the paper becomes a negative result on fidelity-as-calibration, and the fidelity reward has to justify itself through rationale quality alone.
Why not just sample several rationales and use their agreement as confidence? Self-consistency is a known hallucination signal; the file cites work showing consistency collapses under cross-examination. It costs K forward passes per alert, which conflicts directly with the throughput problem that motivates alert-fatigue work. We run it with K = 5 and K = 10 and report recall at fixed precision against total generated tokens. If agreement wins only at several times the compute, the single-sample checker is the better choice for deployment. If it wins at comparable compute, we adopt it and drop the calibration claim.
Why not just retrieve the relevant record fields into the prompt and trust the model? The model already sees the full alert record, so retrieval adds nothing that is not already there. The failure we target is a model misreading context it has, such as swapping a parent and child process or a source and destination. The file documents this kind of role-swapping in structured-to-text generation. TRAC likewise argues that retrieval alone gives no correctness guarantee. Our measured contradicted-claim rate for the untrained model directly tests whether misreading happens on alerts.
Why not just use the direct-label classifier for confidence and the reasoning model for the verdict? This hybrid is cheap and gives honest probabilities. We run it. The prior result suggests its weakness: the direct model's confidence tracks its own, weaker decision boundary, so it may sit badly with the reasoning model's verdicts on hard benign alerts. The experiment settles the question either way.
Proposed Experiments
The primary testbed is alerts built from CSE-CIC-IDS2018, where each flow record is enriched with host and topology context following the (EC)2 setup, with a strictly temporal train/calibration/test split. UNSW-NB15 and HIKARI-2021, used in ContextBuddy, serve as cross-dataset shift tests. Backbones are Qwen3-1.7B and Qwen3-8B, matching DynaGuard's sizes. Metrics are benign recall at 98% precision, malicious recall at 99% precision, expected calibration error, realized missed-attack rate among auto-closed alerts against the target $\alpha$, contradicted-claim rate, decisive-field coverage, and generated tokens and wall-clock time per alert. As a secondary check that the method is not tied to security data, we apply the same claim tagging to policy clauses on DynaBench, where DynaGuard-8B reports 73.1% F1. The question there is whether fidelity training preserves F1 while reducing misquoted policy content.
The first decisive check comes before any training. On the untrained, teacher-distilled model, we measure whether contradicted claims predict wrong verdicts. If the AUROC of contradicted-claim count for detecting errors is near chance, the core hypothesis fails, and we stop and report it. The second check is whether combining checked claims with the label probability closes most of the gap to the separate calibrator at 98% precision on a held-out test window. The third check is whether the realized missed-attack rate stays within $\alpha$ for each successive test month. If it does not, we report how fast the guarantee degrades with drift and whether recalibrating on a small recent window restores it. Finally, the fidelity-reward model must reduce contradicted claims without lowering decisive-field coverage or verdict accuracy. A reduction in claims bought with silence counts as a failure.
References
- Cybersecurity Detection Classification with Reasoning-enabled Language Models
- GroundedPRM: Tree-Guided and Fidelity-Aware Process Reward Modeling for Step-Level Reasoning
- TRAC: Trustworthy Retrieval Augmented Chatbot
- Specializing Smaller Language Models towards Multi-Step Reasoning
- DynaGuard: A Dynamic Guardian Model With User-Defined Policies
- (EC)2: Event-Centric Explainability for Cybersecurity Through Multi-Agent LLM Investigations
- ContextBuddy: AI-Enhanced Contextual Insights for Security Alert Investigation (Applied to Intrusion Detection)