Research

Studies of how well ASMI actually works — method, data, and results, including the ones that went against us.

ASMI's claim is not that it produces the best brand voice. Every vendor in this category claims that, and none of them publish a number you can check.

The claim is narrower and more checkable: we measure what our avatars do, we report what we find, and when a method we built turns out not to work we say so rather than quietly dropping it. These pages are that record.

2026-07-3118 min
How faithfully can a configuration layer control a chatbot's brand voice?
We told avatars to be playful. They scored 1.00 out of 5. Finding out why took repairing our own measuring instrument, and the fix that finally worked was not the one we expected.
- Avatars configured for an amplified trait delivered almost none of it — the effect was specific to generative flair, not to high settings in general.
- Six defects were found in our own measurement before any of it could be trusted, including a headline metric that was semantically inverted.
- Three rounds of prompt engineering stopped moving the number. Changing the model moved it further in one step.
- A third measurement method was built, pre-registered, calibrated — and failed. Its numbers are published here.
Read the study

A product by Broentech Sentinel.