
What should we measure before enforcing a policy?
A high agreement rate in shadow mode is not enough. Diverging requests, missing signals and rollback conditions need to be reviewed together.
Modelion / Engineering journal
Engineering notes on the policy engine, routing and guardrails. Not product announcements; the reasoning behind decisions we made, and the walls we hit along the way.

A high agreement rate in shadow mode is not enough. Diverging requests, missing signals and rollback conditions need to be reviewed together.
6 posts found

Reporting spend after the fact and deciding whether a call may start are different jobs. Estimates, concurrency and actual usage need separate treatment.

A model name describes the outcome, not the reason. Read a decision trace from input signals through the winning rule to enforcement obligations.

Once the first response chunk reaches the user, retrying means something different. The boundary between a partial answer, a failure and a new attempt.

PII arrives inside the variables, not inside the template. That single observation fixes the entire stage ordering in the gateway.

A fallback chain normally buys you resilience. While a residency rule is in force it does the opposite — and it does it silently.

When model selection, redaction and refusal logic is spread through application code, two things break together. What changes when the decision lives in one place.