cortexa-write-scorer (v1 — worker stub, no weights)

Production scorer for Pleius Write. There are no model weights in this repo. Write's launch backend is a deterministic lexical/structural feature extractor implemented in TypeScript inside the cortexa-proxy Cloudflare Worker — version-tagged write-worker-stub-v1. It is a pure function of the input text (no network, no learned parameters), so identical text always yields the same scores.

Why no ONNX?

Write's iter54 model concept (LC=66.90% on the LM-arena benchmark) was implemented as a worker-side approximation for launch: TF-IDF word 8k + char 8k surrogates, per-family logic, meta-blend across pillars. A neural Write scorer is a planned follow-up but not yet trained.

Pillars produced

• read_likelihood — will a reader start AND keep going? Flesch reading-ease analog + type-token ratio + word length.
• hold — structural variety. Sentence-length coefficient of variation (monotone cadence loses attention).
• structure — paragraph density / transition-word density / discourse marker frequency.
• score (overall) — meta-blend: 0.4·read + 0.4·hold + 0.2·structure.

Where the code lives

Worker source: https://github.com/matthewf725/Perspecting/blob/main/research/cortexa_trm/unified/product/proxy/src/index.js (function scoreWriteStub)

Python port for offline distillation: https://github.com/matthewf725/Perspecting/blob/main/research/distill_students/write_scorer.py

Surface

POST /predict/write on cortexa-proxy Cloudflare Worker.

License

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