Show HN: Jevstiller – Distill Jev into a local model, with a disagreement bound

Jevstiller shows how a local model can absorb a high-volume classification workload from a vendor model while preserving a measurable agreement guarantee: it answers requests in about 15 ms on CPU and routes only uncertain or out-of-distribution traffic back to the original service. For CIOs, the strategic significance is that this creates a practical pattern for reducing latency, vendor dependence, and per-request cost without losing control over error budgets—using a statistically bounded contract rather than a simple confidence threshold that can silently overshoot risk. IT organizations would need to treat this as a governed cascade architecture with continuous calibration, shadow testing, and versioned retraining if they want the SLA-style guarantee to hold in production.