Ainvest-Decision-27b

Text-only decision model on Qwen3.8-27B scoring options through a custom 255-way readout with last-token pooling and noncausal full-attention layers — evaluated as an M02 weight merge, 26B stored params in BF16. #15 on the Decision Index's Nov 6 draw (68.23); the card is precise about what its numbers are ("existing public-run scores, not a maintainer-assigned Full score"), ships publication validation (zero choice flips on a fixed 100-request sample), and credits tasksource for much of the gain. Prompts over 16,384 tokens per question are refused, never truncated. Apache 2.0 with the full inference package in-repo.