Most classifiers either generate free text you must parse or output uncalibrated scores; this model avoids both by never generating text and by training explicitly to report honest probabilities. That design makes it practical to run high-throughput decision workloads (intent routing, email triage, phishing detection, moderation) with a single forward pass and mathematically calibrated confidences.
What Sets It Apart
- Non-autoregressive, typed outputs: answers are structured (choice/score/noul) rather than free text, eliminating parsing errors and hallucinations.
- Calibration-first training (RLCD): trained with Reinforcement Learning for Calibrated Decisions so the model reports probability distributions that align with proper scoring rules; fitted temperatures further reduce ECE in practice.
- Low-latency, multi-checkpoint router: English checkpoint (ModernBERT-large, 421M params, 512-token budget) answers a single question in ~33–40 ms on a T4; a Router dispatches to multilingual or typed-decision checkpoints as needed.
- Option-marker scoring and batch answering: each option is scored at its own marker token and multiple typed questions are evaluated in one forward pass, making batched throughput high.
- Benchmark-aware: published speed and calibration numbers, plus measured strengths on AG News and several decision benchmarks; the fine-tuned typed-decisions checkpoint substantially improves workflow accuracy.
Who it's for and trade-offs
Great fit if you need deterministic, high-throughput decision outputs with calibrated confidences for production routing, ticket triage, content moderation, or automated guardrails. The Router helps mixed-language fleets by selecting a multilingual checkpoint when needed.
Look elsewhere if you need generative text, broad zero-shot reasoning, long contextual windows beyond the 512-token English budget, or strong ordinal scoring out of the box (SST-5 is a weaker primitive). Also refit temperatures on your distribution: the shipped checkpoints can be overconfident until you calibrate per question-type/option-count. Keep choice questions under ~20 options to avoid head-length limits.
