Skill Routing
What I built
Skill validation/loading, calibrated routing and abstention, paired task experiments and rubric evaluation.
Result
Implemented skill validation, routing and abstention with a 64-example calibration split, a 136-example test split and paired task evaluation.
Evaluation setup and selected results
Routing was evaluated with 64 calibration examples and 136 test examples, alongside a separate 45-task paired study. The implementation includes skill validation, calibrated selection, abstention and uncertainty estimates. This case highlights evaluation and routing infrastructure; the paired study did not establish an overall task-quality gain.
Compare the evaluation evidence →
About the evidence
Based on source review and saved project artifacts. Recorded measurements apply to the stated model, dataset and configuration; they are not claims of a current production deployment.