Automated hiring systems typically emit a score: a scalar that ranks a candidate against others. Such scores resist explanation, give a candidate nothing specific to contest, and carry growing regulatory exposure. We present CogniHire, a technical screening system designed to audit the claims a candidate makes rather than score the candidate. A deterministic grounding gate constrains a language model to select claim text from the resume and forbids it to author any. Across roughly 100,000 synthetic trials the gate rejected every meaning-preserving paraphrase and negation trap while admitting every verbatim claim, and under a worst-case prompt-injection model it admitted none of 1,800 fabricated claims. For machine learning we compare nine classifiers under grouped cross-validation with significance testing and introduce a within-query diagnostic that separates matching ability from candidate-level quality. For deep learning we show that a plausible authored face-similarity threshold of 0.50 would have falsely rejected 41.4% of genuine identity matches, against 3.4% at a threshold calibrated on Labeled Faces in the Wild.