Retraining Without Regression.

I What We Mean By Retraining

Retraining at Phazur is not training from scratch. The base model we build on retrains itself on its own cadence. We do not touch it. What we retrain, quarterly, is the corpus slice: the commercialization pattern library, the per-industry ontologies, and the context-composition rules.

A quarterly cadence is the floor, not the ceiling. If a material shift happens mid-quarter, we ship an interim update. In the last five months, one such interim update was required, after a regulatory shift in one of the seven verticals.

II The Four Gates Before Promotion

A new corpus candidate passes four gates before promotion to production installs.

Gate One: Diff review. Every change from the last quarter’s corpus is enumerated. A human reviewer, per vertical, signs off. If the diff exceeds a threshold, we slow down and re-scope.

Gate Two: Regression benchmark. We replay a library of historical install interactions against the new corpus. We compare outputs. Any output that changes class (for instance, a previously compliant recommendation now flagged non-compliant) gets a manual review.

Gate Three: Shadow run. The candidate corpus runs in shadow behind the production corpus on a sample of live traffic for seventy-two hours. Outputs are logged but not returned. We diff.

Gate Four: Canary install. The new corpus is promoted to two early-adopter installs for one week. If no regression fires, we promote to the full install base.

III The One Regression That Nearly Shipped

In the 2026.Q1 refresh, a change to the wellness ontology narrowed what the recommender was allowed to propose in segments where a customer had paused interaction. The narrowing was correct in principle. In practice, it collided with a pattern already in production that produced wellness reactivation prompts.

The gate-three shadow run caught it. Two hundred recommendations per day were getting suppressed that previously shipped successfully. The corpus candidate was rolled back, the ontology was re-scoped to preserve the reactivation path, and the refresh shipped one week later than planned.

We publish this because transparency on near-misses is how buyers learn to trust the cadence. A refresh that never fails is a refresh that is not checking carefully enough.

IV What Quarterly Actually Buys You

Quarterly cadence buys three things.

First, the January corpus is not the October corpus. New patterns, new regulatory facts, new vocabulary. The install gets sharper while the customer sleeps.

Second, a competitor cannot skip the cadence. They can work harder. They cannot work faster than the calendar. Every quarter we ship, their catch-up window narrows.

Third, the customer does not have to do anything. The quarterly refresh ships to their installation, passes the four gates inside their environment, and promotes itself. Operationally, it looks like a version number moving on the dashboard. That is the entire experience.