Measured, not asserted.
Every core claim is measured on real, public, non-toy workloads and released with a one-command reproduction harness — flywheel run all-local — that reproduces the results with no database or cluster required.
Production-capable at reference scale.
On a real 251k-case procurement log, the continually-retrained flywheel beats a frozen policy (signed-rank p≈0.009). A label-permutation placebo collapses the gain 389× — the value is freshness, not volume. M
411× cheaper than full re-materialisation; 13–814× vs a strong incremental baseline on a real 251k-cell workload. M
Zero argmax mismatches between the C++/ONNX production engine and the Python reference, with ~20× tighter tail latency. M
Sub-layer latent activation steering recovers 98.5% of native FP16 reasoning at 4-bit (73.7% vs 74.8% FP16). M
A randomised trial + heterogeneity sweep establish a falsifiable rule: a learned policy beats a fixed rule only in the regime of pricing, discounting, and material re-sourcing. M
On OULAD, the policy shows no significant disparity by disability or deprivation — auditable via lineage, not asserted. M
HELIX v1.7 Build #58 performance is documented separately on the benchmarks page (official vLLM GuideLLM, AMD EPYC 9254).
Academic collaboration or a review packet?
Book a call for the full evidence dossier and reproduction access.