PORTFOLIO RESEARCH SERIES · STUDY 016

RobustnessLab AI

Does a classifier still behave the same when the wording changes? Study multilingual prediction stability, uncertainty, data perturbations and research release gates with fully traceable evidence.

EN · ES · PT4 lexical baselinesPaired stress testsNo paid API

EXPERIMENT SNAPSHOT v1.0

—Evaluation records

6Test conditions

3Languages

2Held-out splits

Source: Academic Rhetorical Intent Mini · synthetic-curated · no real-world validation

iSynthetic research warning. Transformation labels are inherited for exploratory diagnostics only. Some changes may alter meaning. A flipped prediction is not automatically evidence of real-world model failure.

01 / COMMAND CENTER

Robustness overview

Paired baseline vs. perturbed evaluation on identical example IDs.

Loading reproducible evidence…