mild-rgb/eli5-human-vs-ai-detector

Card: qualify OOD generalization — generalizes across domain, not model vintage (240-doc OOD probe)

Card: move the Pangram line to the top

Card: note this was trained from a real corpus, not distilled from Pangram

Add ModernBERT-large detector weights, tokenizer, config, and run-record provenance (seed 1, binary|norm)

Card: full 11.3%% FPR in the misuse warning; qualify in-domain 1.000

Correct card: genuine OOD detector (bal acc 0.863), calibration warning

Add model card (limitations foregrounded)

initial commit