mirror of
https://github.com/hanzoai/ai.git
synced 2026-11-21 02:05:26 +00:00
The flywheel has no separate ML engine to route through — the trainer
retrains the prefer table instead — so `engine_share` counted consultations
of a decision service that was never built. It read 0 on every day of its
existence and told the dashboards the learned router always loses.
Replace it with what actually learns:
learned_share share of decisions whose served model is the pick of the
TUNED prefer table (trainer-published arms + admin pins,
above the static conf seed)
by_source the decision-provenance histogram behind that ratio, so a
strategy that NEVER fires is visible at a glance instead
of being inferred from a share stuck at zero
Both engine and judge swallowed every failure and returned a bare
false, which is how an unset ROUTER_ENDPOINT and a judge with no service
bearer stayed invisible while the reward ledger went empty and the
trainer's gate never cleared. Each failure path now routes through a
counter (router.Misses, controllers.JudgeMisses) and a rate-limited line
carrying the reason, and an armed judge with no token says so at boot.
Judge reasons stay content-free — the no-retention guarantee covers logs.
Co-authored-by: Hanzo Dev <[email protected]>