Pre-registering eval claims for Giskard scenarios (falsify-giskard): of interest? #2498
Replies: 1 comment
|
Quick follow-up with something concrete. I built Here it is locking a It reads Not selling anything. I'd genuinely value two lines: does "lock the eval claim before the run" resonate as something Giskard users would want, and am I reading the result model the right way? Repo: |
Uh oh!
There was an error while loading. Please reload this page.
Hi, and thanks for Giskard.
I just shipped
falsify-giskard(on PyPI), a small integration that lets you pre-register a Giskard eval claim (a metric and threshold, e.g.pass_rate >= 0.9or a named check's score) to a SHA-256 before the run, then verify theScenarioResultagainst it with a PASS / FAIL / TAMPERED verdict. It reads yourCheckResult/Metricobjects directly. PRML (the manifest format) is an open standard with byte-equivalent reference implementations in four languages.The idea: a scenario report shows what happened, not what was promised beforehand. Anchoring the claim to a pre-run hash means relaxing a threshold or swapping a model after seeing results breaks the hash.
pip install falsify-giskard # preregister(...) before scenario.run(), then verify_scenario_result(result, manifest) afterI built it because Giskard's check/metric model is a clean fit (I did the same for Inspect AI). Mostly curious for your read: is pre-registration something Giskard users have wanted, and would a one-line mention in the docs be welcome? Happy to PR it. No rush.
All reactions