RFD 1132: Priority list of converted models
Problem
RFD 1016 inventories fifteen deep learning models. Nothing says which to convert for the accelerator first, so the order is decided by whoever picks one up.
Ten ids are excluded rather than ranked, and RFD 1133 says what excludes each: three by the agreements, four by an abandoned scope, three by nothing found. One model has been measured. An order built from guesses reads the same as one built from evidence.
Decision
Rank by evidence, and say which rank rests on what.
RF-DETR keypoint is first because it is the only model with a measured export: 825 nodes over 22 operators, every one inside the allowlist. It is also the only one with a measured obstacle, which is worth more than an untested candidate.
Excluded models are not ranked here. RFD 1133 carries them, one group per kind of exclusion, because a blocklist row and an abandoned scope reopen on different events. Everything else is unranked pending a census, and the gate that produces one exists.
DETAILS.md gives the table, the measurement behind the first entry, and what each remaining model needs before it can be ordered.