Trim README to YAML front-matter; note CPU in short_description

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

README.md

@@ -7,17 +7,5 @@ sdk: docker
 app_port: 7860
 header: mini
 pinned: false
-short_description: Policy linting with LFM2.5 Encoder, server-side CPU
+short_description: Policy linting with LFM2.5 Encoder on CPU
 ---
-
-# Policy Linting
-
-GLiNER-style company-policy linter powered by [`LiquidAI/LFM2.5-Encoder-350M-bizlint-v03`](https://huggingface.co/LiquidAI/LFM2.5-Encoder-350M-bizlint-v03).
-
-Write customizable policy rules and lint a draft. The bidirectional encoder scores every token against every rule in one pass and highlights likely policy issues. It runs on the Space's CPU; the browser only sends `{text, rules}` to `/api/lint` and highlights the result, so nothing is downloaded to the visitor and no WebGPU is required.
-
-## How it works
-
-- `bizlint.py` — one bidirectional LFM2 encoder forward over a `Policy: … Text: …` prompt, then a two-tower projection head (`rulehead.chunk`): mean-pool each rule's tokens to a 256-d vector, project every draft token to 256-d, and score each (token, rule) pair by a scaled dot product + bias through a sigmoid. The encoder runs in **f32** (linting is one forward, and int8 noise drops borderline spans). Weights load from `bizlint.gguf` (Q4_0).
-- `app.py` — Flask server: loads the model once, serves the page, answers `POST /api/lint` with per-token, per-rule probabilities.
-- `index.html` / `style.css` — the UI (with the model status row + loading bar). Word grouping, the sensitivity threshold, and highlighting run in the browser, so moving the threshold is instant and never re-runs the model.