REPRESENT searches representations under behavioural contract

REPRESENT is the optimizer above quantization: it freezes a behavioural contract — bounds on KL divergence, displaced probability mass and routing consequence at a fixed evidence scale — then searches candidate physical representations per region and returns the cheapest topology whose COMPOSED behaviour passes at authority scale. Its first earned result, on Kimi Linear 48B, keeps layers 0–23 at BF16 and takes layers 24–26 to Q8_0: 8,192-position PASS, 5.1 GB removed, measured faster beside its own baseline. Quantization is one mechanism it uses; deciding which representation, where, in which combination, is the point.

SUPPORTED · kimi-logit-v3 PASS @ 8192 positions — KL p99 4.153e-4 vs 1e-3 limit