RES — Nine candidate open-weights models were selected (1B-12B across four families)
Nine candidate open-weights models were selected (1B-12B across four families); eight were empirically evaluated after one reasoning-style model was excluded for parser incompatibility (emits internal reasoning blocks, ~0.8 parse-failure on a smoke test).
Source: PUMA project documentation · Traceability: corpus unit
MEM-025· Confidence: verified-at-primary-source