fix duplicated promotion line and stale cross-reference

v46 release notes under the current-champion section

release notes lead the card; distribution detail moved under Files

v46 release notes: one multimodal model at 44% of the download, engine swap removed, text unchanged, last surviving defect eliminated

champion is now smart-coach-vision (text+vision); smart-coach stays the stable text-only build for existing app versions; link smartsight.app/ai-coach

correct the body-fat claim: DEXA-verified reference shows the tower still reads a full band high

document smart-coach-vision.litertlm: one model instead of two, 6.29GB -> 2.80GB, trained body-fat tower

add smart-coach-vision.litertlm — consolidated coach+vision, body-fat-trained tower

model card: v46 champion — avoid-list defect eliminated on both seeds

model card: champion v46-seed123

champion v46-seed123 (avoid-list violations eliminated (0/36 vs 5/36) by matching the corpus to the prompt format production actually sends)

model card: v44 champion section, measured numbers and stated regressions

model card: champion v44-seed123

champion v44-seed123 (Danish macro vocabulary (kulhydrat/fedt) and Danish challenge titles; /nutrition line now emitted on every applicable case)

Restore champion history (v27 through v42) lost in the v42 card rewrite

Fix model tree: base_model is now gemma-4-E2B-it-qat-q4_0-unquantized (changed at v42)

README: v42-seed123 champion (QAT base + fenced ChallengeGen + b64fc blockwise export)

v42-seed123: QAT base + fenced ChallengeGen corpus + blockwise b64fc export

Update README to current champion v38-seed456 (was v31); add measured weaknesses

Upload smart-coach.litertlm with huggingface_hub

v34-seed456: fixes v33 regressions (face-pull confabulation, within-day duplicate exercises), keeps working dietary-compliance fix. Beats champion under both loose and strict 3-seed grading. lr 8e-5->6e-5.

Document custom-op Hadamard export: 88% faster, quality confirmed via 312-row multi-seed adjudication

Same v31-seed456-hr weights, faster export: custom-op Hadamard rotation (88% faster decode, 85MB smaller, quality confirmed statistically indistinguishable via 312-row multi-seed adjudication)

Replace spot-check figures with full systematic 52-row-per-model read: v27 had 12 defects/2 timeouts vs both v31 seeds' 2 each, baseline 0

Remove alternate-seed section: only the single best-of-round checkpoint stays in this repo

Remove seed123-hr: only the single best-of-round model should be kept, not every candidate

Document v31 win over v27: technique-cue correctness, structured-output reliability, quantization-recipe finding, honest remaining gaps

add v31-seed123-hr alternate: clean 5/5 technique-question phrasing but fails the no-fish dietary-rule test (documented in README)

smart-coach v31: full-catalog technique + variety-compliance corpus, gemma4_mixed48_hr quantization -- beats v27 on technique-cue correctness and structured-output reliability

Upload README.md with huggingface_hub

Upload smart-coach.litertlm with huggingface_hub

Upload merged bf16 checkpoint + patched chat template + int4 .litertlm (test run)

Upload coach.litertlm with huggingface_hub

Delete coach.litertlm

Upload coach.litertlm with huggingface_hub

initial commit