The headline: a 0.40-nat improvement at constant compute.
REFINERY is a 200-line C kernel that runs alongside any language model's forward pass.
Its vibe_paradigm configuration reaches a regime the canonical scaling laws say should require 10× more training compute.
Same model. Same forward pass. Zero added FLOPs.
−0.40 nats
Loss saving at constant compute
≈ 10×
Compute-equivalent quality gain
83.7%
Win rate · 251 / 300 benchmark
235 KB
Working set · sub-ms overhead
The data, directly.
Public-safe visual outputs from the refinery proof run. The CSV below contains aggregate category metrics — per-row sequences are withheld.
Distinct-2 vs External-base PPL · the breach
Rep-2 vs External-base PPL
Self-BLEU vs External-base PPL
Entropy · before vs after
Effective temperature histogram
Sample controller trace
Download the public results CSV
Aggregate category metrics · safe for distribution · 19.7 KB