2 件 ・ 関連度順
概要と使いどころ
Build and run a labelled eval set for a System One model (Jev, Von, or any typed-decision config), then sweep criteria wordings and thresholds against it. Use when a Jev/Von classification is wrong or unreliable, when choosing between the hosted API and a local open model, when tuning noul thresholds, or before shipping any typed-decision feature. Produces an accuracy-by-wording matrix and a calibrated threshold.
日本語の概要は準備中です。原文の説明を表示しています。
Add an external probabilistic classifier to a decision path without letting it take the path over. Covers finding the externally-graded rows you are already logging, measuring the confidence separation between the overrides it got right and the ones it got wrong, choosing an operating point your data licenses, failing open at the call site, and proving the whole arrangement offline with no API key. Applies to any service returning a score you can order — TypeSafe AI's System One models (Jev) are the worked example. Most of the method transfers to a local model or a second heuristic, with the parts that do not called out where they arise. Use when a rule-based path is wrong often enough to hurt, when someone proposes replacing a heuristic with a model, when a confidence threshold needs a defensible value, or when an oracle already in production has never been graded.
日本語の概要は準備中です。原文の説明を表示しています。