LOCAL-FIRST: engines run in your browser · zero AI tokens for quant questions · evidence packs when you do ask a model← terminalsectorsmap
1 · ENGINE RUNNER deterministic · whitelisted · no eval
Answer it locally before spending a single model token: stretch, volatility, stationarity, out-of-sample edge, touch odds — the same parity-tested engines the terminal runs.
Instead of re-sending files to Claude/ChatGPT, send this: every core conclusion for the loaded name, computed locally, hash-stamped, a few thousand characters. The 30× token saving from the report, made concrete.
Save a conclusion once (with its evidence hash) and reuse it — the ledger is the memory that stops you re-deriving, and re-paying for, the same answer.