dvar

Module 5 · Model upgrades and inference

Upgrade the call model on proof, not hope

A model swap is safe only when an eval on your own calls says so.

The upgrade loop

  1. 1

    Pick one workflow

    Outbound sales call QA.

  2. 2

    One labelled slice

    200 to 500 calls your QA team has reviewed.

  3. 3

    Build the eval

    Gold, checks and a calibrated judge.

  4. 4

    Score every candidate

    Same calls, same checks.

  5. 5

    Gate

    No critical miss, no quality drop. Or it doesn't ship.

  6. 6

    Tune on the misses

    Misses become training data.

  7. 7

    Ship and watch

    Shadow first, then live, checked on every call.

What it costs at your volume

Quality decides which models are allowed; cost decides between them.

7×

closed-source cost vs open model + examples

14 vs 11

calls fully right: open model + examples vs today, at 1.1× the cost

0

critical misses for open model + examples

Today's misses become training data

Today's 7 misses, once reviewed, become training examples.

The same comparison on 180 calls →