The standard recipe for LLM-as-judge -- pick a frontier model, or average several -- is actively unsafe for grading the psychological safety of conversational AI. Using aipsy-bench, an open frozen safety instrument, we run a fully-crossed competence study: three frontier models (gpt-5.4-mini, claude-sonnet-4-6, gemini-2.5-flash) serve as both generators and judges of 3,000 mental-health, companion, and coaching messages against a psychologist's ratings. The disagreement is not noise: it is structured, concentrated on the safety-critical metrics, and one judge (Gemini) is an outlier -- the most lenient, carrying a +0.99 self-preference premium, flagging far fewer tail failures, and scoring a means-in-hand self-harm response "exemplary." Inter-judge agreement on empathy, where sycophancy hides, is the lowest in the battery (alpha 0.24). One axis stands apart: the binary crisis-detection flag is the one safety-critical signal judges agree on (alpha 0.80), erring toward over-flagging, the safe direction for a triage screen. Equal-weight averaging, the canonical fix, blends that leniency and tail-blindness into the safety score. Off-the-shelf open-weight judges are worse for a dispositional, not capability, reason -- and disposition is fine-tunable. We therefore distill a per-metric, psychologist-corrected target into a small, frozen, local model, aipsy-judge-1.0, an Apache-2.0 fine-tune of Gemma-4-26B-A4B. aipsy-judge-1.0 tracks the corrected target better than its base on the composite (ICC 0.64 to 0.75) and crisis detection (kappa 0.65 to 0.82), catches 92% of crises with a false-positive lean, and grades more faithfully than any single frontier judge, while every transcript stays on the machine. These are directional readings against a single-expert-informed target, not validated multi-rater agreement. A safety grader that shares a vendor's post-training shares its blind spots.
Real-world crisis intervention is inherently conversational, yet existing research largely focuses on static texts.Real-world crisis intervention is inherently conversational, yet existing research largely focuses on static texts. When applied to multi-turn dialogues, current models exhibit significant performance degradation, struggling to track risk signals that emerge as context evolves. To address this gap, we introduce CRADLE-Dialogue, a clinician-annotated benchmark for turn-level crisis detection in conversational settings. The dataset features 600 dialogues with multi-label annotations across clinically grounded risks, including suicide ideation, self-harm, and child abuse, distinguishing past from ongoing risk. We further propose an Alert-Confirm evaluation protocol that distinguishes early warning signals (Alert) from turns where a specific crisis becomes explicitly identifiable (Confirm), reflecting the clinical need to intervene before risk becomes explicit. Experiments show that identifying when risk emerges is much harder than recognizing that it exists: models achieve only mid-40% to high-60% Micro F1. Additionally, we release a synthetic training corpus and a 32B-parameter model that substantially outperforms existing open-source models and achieves competitive or superior results against proprietary models across turn-level, dialogue-level, and confirm-only evaluation settings.