kev-9B alert-triage — interactive probe

This static build emulates the severity-distribution interface of kev-9B-LoRA-ranker — 72-day production record, 90% vs 10% in 40-case human arbitration. The production model runs a LoRA + 256-d rank head on Qwen3.5-9B; this page runs its rule-lexicon stub for UI exploration.

PodCrashLooping back-off 14× restarts Flyte orchestration core down — workflows stuck Instance auto-recovered after restart (complete) OutOfMemoryError killed trainer pod

–info

–warning

–critical

verdict:

With none-injection (p=0.1) the head can abstain; here the stub always picks the argmax.

kev: severity is an ordered scale — trained with an RPS ordinal objective, calibrated via LBFGS temperature fit. Rule-stub logic shown here mirrors the downgrading of high-frequency auto-recovering alerts and escalation of orchestration-core outages (the behaviors that won 90% vs 10% in arbitration). Full receipts on the model card.