Snugi-AI-v2 @ eRisk 2026 Task 2: Early Depression Detection via a Learned Stopping Policy with Sustained Confidence Gate

Summary

This CLEF 2026 working-notes paper presents Snugi-AI-v2, a submission to eRisk 2026 Task 2 (early depression detection from Reddit threads). The pipeline encodes each round with frozen MentalRoBERTa-base, maintains an O(1) incremental mean-pool over target-user posts, and classifies depression with a small MLP; timing is delegated to a learned stopping policy (a 5-32-16-1 MLP over trajectory features: probability, round, word count, delta-p, max p) plus a sustained confidence gate requiring N=3 consecutive high-confidence rounds. The authors claim this is the first stopping policy trained to directly optimize ERDE50, though training actually uses BCE loss on heuristic labels (first round with p>0.5). Five runs ablate two encoders, four stopping strategies, and three gate values on 909 eRisk 2025 subjects; 5-fold CV gives ERDE50 0.0290 vs 0.0369 (policy vs fixed threshold) with heavily overlapping CIs and no significance test. On the official test set (523 users, 500 rounds), best F1 is 0.73 (Run 1) and F_latency 0.70 (Runs 0/3), median alert round 8, and the system finished fastest (1h26m) among 16 complete-submission teams, ranking 9th by F1 but 4th on NDCG@100 (0.85). Negative results (GRPO, MentalLongformer, DeBERTa ensembling, BDI-II filtering) are documented. Main limitations: overstated 'direct ERDE optimization' framing, recall dropping below the fixed-threshold baseline despite the abstract's claim, mixed CV/validation scales in the ablation table, and an undefined DepRoBERTa in the encoder-reversal analysis.