Tiny Aya citation grounding probe
This artifact contains the standardized logistic regression coefficients, intercept, normalization parameters, and frozen threshold learned from Tiny Aya Global citation drafts. run_config.json records the model revision, activation layer, and runtime settings. The Hugging Face base model is loaded separately by the GitHub notebook and is governed by its upstream terms.
For a draft activation vector, standardize with the saved mean and scale. Compute the sigmoid of the dot product with the saved weights plus intercept, then abstain when the score is above thresholds.probe_risk.
The probe was evaluated on a controlled English bibliography benchmark. Its output is a decision score with no guarantee of citation existence. See the GitHub notebook and held-out report for the data split and limitations.
Measured holdout: 0/24 unsupported releases, 11/12 correct evidence-backed citations, and 23/24 AuditKIT exact match. Token confidence matched these aggregate scores.