cortexa-write-feedback (distilled student)

A ~4.4M-parameter conditional decoder distilled from M725/cortexa-write-scorer (the worker-side TF-IDF/lexical stub). Takes MiniLM text features (384-d) + the 4 Write pillar scores and emits a creator-vernacular phrase chain about the draft:

"first line hooks | ending sticks"
"tight middle | shareable"
"wall of text | no reason to read"
"drags | no payoff"

Files

file	purpose
student_int8.onnx	TinyTransformer decoder, 4 layers / 256-dim / 4 heads, INT8 dynamic-quantized. 6.8 MB.
tokenizer.json	Whole-phrase tokenizer (vocab ~120; specials <pad>, <bos>, <eos>, <sep>).
config.json	Encoder dim (384), pillar names, vocab size, special-token ids.

Inference shape

inputs:
  encoder_feats   (1, 384)  float32   # sentence-transformers/all-MiniLM-L6-v2 mean-pooled, L2-normalized
  scores          (1, 4)    float32   # [read_likelihood, hold, structure, score] in [0,1]
  scores_present  (1,)      float32   # 1.0 anchored, 0.0 fast-mode
  input_ids       (1, T)    int64
outputs:
  logits          (1, T, V) float32

Training

See research/distill_students/train_write.py in the app repo. Teacher is score_write_for_rules() — the Python port of the cortexa-proxy worker's deterministic TF-IDF write scorer.

License

Pleius internal — see https://pleius.com. Not for redistribution.