Pramaan प्रमाण

What this system does not do

It cannot reliably tell an answerable question from an unanswerable one. Two attempts were built and measured: retrieval-score features reached ROC AUC 0.5595, cross-lingual entailment 0.5992 — against 0.5 for a coin flip. Both were rejected rather than shipped.

What runs instead is a conservative score floor. It catches the degenerate case where retrieval found nothing similar at all. It is not a calibrated abstention model, and it is not described as one anywhere in this interface or in the API response.