Update eval_semantic_stability.md

This commit is contained in:
PSBigBig × MiniPS 2026-02-26 15:47:04 +08:00 committed by GitHub
parent bdeab6d983
commit f8f3946e84
No known key found for this signature in database
GPG key ID: B5690EEEBB952194

View file

@ -16,6 +16,11 @@
> If you need the full triage and all prescriptions, return to the Emergency Room lobby.
</details>
> **Evaluation disclaimer (semantic stability)**
> Stability scores in this page are heuristic signals about how outputs move under small changes.
> They do not prove global robustness or safety and should be combined with other checks.
---
**Goal**
Quantify how **stable** your pipeline is under small, *non-semantic* perturbations: different seeds, low temperature noise, and benign **prompt jitters** (punctuation/whitespace/synonym swaps). A robust system should keep claims, citations, refusals, and constraint echos **invariant** (or nearly so).