funnel-compressionlisted

# Funnel Compression (denoising) Input: `original_text` (the long text the user pasted) + `{{use_case — e.g. Task_Delegation}}` Output: strict JSON — `{ "action": "compressed", "compressed_text": "...", "word_count_original": N, "word_count_compressed": N, "loss_rate": N.NN, "verification": {...} }` ## Procedure 1. **Load the compression prompt** — Read the `{{compression prompt location — the section or file where the model's compression prompt and acceptance rules live, e.g. a "Model reference" section in the agent instructions or a config file}}` and extract the compression prompt. Done when the denoising prompt and acceptance rules are in context; if the section is missing → stop, report `{"action": "error", "reason": "No compression prompt defined"}`. 2. **Count the original** — record `word_count_original` (whitespace-separated tokens). Done when the length is known; the compression target is `word_count_original × {{target ratio — e.g. 0.2}}`. 3. **Extract the core** — apply the compression prompt: extract the absolute core — **Action items, Data, and Conclusions** — and compress to less than {{target ratio}} of the original length. Preserve action items and deadlines **verbatim** — never paraphrase a deadline, owner, or required action. Done when the compressed text contains every action item and deadline from the original. 4. **Hallucination gate (Skill-12) — run BEFORE the compressed text can reach the user** — the original pasted text is the ground truth, so `--