evaluation

Provide the candidates, a rubric with weighted criteria, and acceptance thresholds, then invoke the evaluation mode. It produces a scored table of candidate-by-dimension results, evidence citations per score, and a final pursue or pass decision.

It requires the candidates being compared, a weighted rubric defining each criterion, acceptance criteria such as floor scores, an evidence base to cite, and exclusions listing criteria not to score. Missing inputs are requested before any scoring begins.

Yes, the adversarial re-score extension is designed for high-stakes decisions like hiring, architecture calls, and vendor contracts. It attacks the weakest scores and most questionable citations, then revises only those before finalizing the verdict.

Skip it when you need open-ended exploration rather than a decision, since a verdict that restates the question signals the wrong mode. Also skip the adversarial pass on low-stakes scoring, where judge-disagreement noise dominates at small sample sizes.

The failure conditions require naming the specific deciding variable instead of hedging. Every score must cite evidence, weaknesses must be specific rather than categorical, and the output must always end with a concrete decision.