Daecore/comprehension-classifier

Comprehension Classifier (fp16 ONNX)

An NLI cross-encoder that reads a span of prose and stamps the content-type facet set it asserts — trap, decision, rule, design — with full per-facet posteriors (observation is the residual when nothing clears its threshold). Used for meaning-based routing of memory and document chunks in the daecore runtime.

• Format: fp16 ONNX (keep_io_types), ~370 MB. Served by onnxruntime + tokenizers (no PyTorch dependency).
• Architecture: DebertaV2 sequence classification (2-label entailment); pairwise premise / per-facet hypothesis.
• Parity: matches the fp32 source exactly on a 300-chunk corpus sample — 0 stamp-set disagreements, max posterior |Δ| ≈ 0.02. (Dynamic int8 was rejected: DeBERTa-v2 disentangled attention collapses under weight-only int8.)
• Version: crossenc4_ship_v2.

Files

model.onnx · tokenizer.json · config.json · tokenizer_config.json