petrol-fastwin-v7 (TTA-flip + agreement gate, single 19.4 MB model)
Same agreement-gate pattern that drove v5's +12% composite, but the "second opinion" comes from the same model running on a horizontally-flipped image instead of a different model.
- one ONNX (SuperBitDev/petrol2 weights, 19.4 MB FP16)
- repo total ≤ 30 MB ✓ (fits the petrol element rule)
- run pass 1 on original, pass 2 on flipped, un-flip pass-2 boxes
- WBF cluster boxes within IoU 0.55, emit only where BOTH passes contributed
- singletons = noise = dropped
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support