v4: same recipe trained for 48,000 steps on 480k requests (Decision Index 0.3 public 30.90); v3 kept at tag v3

Update README.md

v3: same recipe as v2 trained for 24,000 steps (Decision Index 0.3 public 29.05); v2 kept at tag v2

v2: shared-state joint cross-encoder (ettin-reranker-150m), Decision Index 0.3 run; v1 kept at tag v1

Model card: replace copied LateOn card with the actual decision-model card

Update README.md

Upload folder using huggingface_hub

Upload README.md with huggingface_hub

Upload README.md with huggingface_hub

Upload README.md with huggingface_hub

Upload README.md with huggingface_hub

Upload README.md with huggingface_hub

docs: Update model card with official 150k Decision Index 0.2.1 ground truth (30.65 raw, 9.58 skill)

docs: update model card with official Decision Index results, 2k-case benchmarks, and calibrated temperature

docs: update benchmark presentation (JevBench 68.11, Classifier-Bench 57.1%, honest OOD breakdown)

docs: refresh model card with SOTA 44.21 Decision Index and 68.11 JevBench composite

docs: refresh model card with clean uncontaminated Decision Index SOTA results and modern quickstarts

card: present as v0, no internal versioning

v3 weights: noulhalved-1M + 10% procedural (cbv2 60.5/60.0, b77 47.8)

docs: consolidate v0 benchmark results

correct Tasksource rows to heterogeneous-split probes

plain-language card: public-only example, sub-200M table

v0: 512k natural typed sq_o (dev_selected step 4000) + card