What didn't work¶

• A SigLIP projector into Laya's text encoder (siglip-projector-experiment): kept text answers bit-identical but never learned to use the image in 5 runs; accuracy with shuffled images matched accuracy with real ones.
• Annealing the cross-entropy weight to zero so training ends on the proper scoring rule alone: tied on accuracy and made the raw model more overconfident (ModernVBERT on The Cauldron).
• Balancing AVA's levels on the train split only: the head learned a flatter vote histogram than the val voters produce and lost to a prior-only baseline until the balancing was removed.
• A third epoch over the same rubric data: half a point on VLFeedback, nothing elsewhere. More passes are flat; the next gains need new rubric data (Score head results).
• An in-browser demo (ONNX Runtime Web, WebGPU): the model exported to ONNX and run on the visitor's device. It ran, and matched PyTorch closely at fp32, but the smaller exports it needed (fp16 on the GPU, 8- and 4-bit weights) and the browser's own image decoding moved the answers enough to misrepresent the model, so it was removed in favour of the Hugging Face Space. Dynamic int8 quantisation, tried for it, moved probabilities by up to 0.62 and flipped the top answer on 4 of 9 validation questions. The code is in the repository history.