Italianhype/Blum-Finance-4B-Challenger

BLUM Finance 4B Outcome-Aware Challenger

This repository contains a non-production challenger trained locally from the immutable BLUM snapshot snapshot-3e00c6ee6fdd.

It teaches BLUM to critique a historical thesis after a mature outcome is observed. The original decision remains point-in-time; outcome evidence is introduced only in a later conversation turn.

Lineage

• Champion: Italianhype/Blum at ef36cc90f8d4e6c7df20b40a82f9b0e41ba4c24c
• Dataset branch: snapshot-3e00c6ee6fdd
• Dataset commit: fc56f70
• Snapshot SHA-256: 3e00c6ee6fdd59a770293ed8eda227838a31199b2c4bf9544e5d948d73be0041
• Splits: 416 train / 52 validation / 53 temporal test
• Method: MLX LoRA, 16 layers, rank 8, 20 iterations, learning rate 2e-5
• Peak memory: 7.412 GB

Measured Result

• Validation loss: 5.685 -> 0.003
• Five-example outcome-reflection smoke test: champion 0/5, challenger 5/5

This is not enough to promote the model. The latest temporal test contains 53 confirmed outcomes and zero contradicted outcomes, so a model that always emits confirmed can look successful. Decision-contract preservation also requires a larger independent evaluation.

Promotion status: BLOCKED. The adapter is published for reproducibility and future champion/challenger evaluation only. It does not establish trading alpha, forecast accuracy or profitability and must not authorize trades.

See evaluation/evaluation.json for the complete blocker record.