Wizhut Research

Wizhut Research

Some ideas only give up their answer once you build them and put them in front of people. This is where we do that: working products, free to use, each one testing something we were not sure about — and each one publishing what it found.

The research projectsPublications
A question, not a roadmap
Each experiment starts from something we genuinely do not know the answer to. If the answer were obvious, building it would be product work, not research.
Free while it is an experiment
Experiments are how we learn, so charging for one would be charging you for our education. They stay free, and the ones that graduate into products say so loudly before anything changes.
The evidence ships with it
Every experiment has a page here with its method and its measurements, including the parts that did not go the way we expected. A claim without a number behind it does not go on these pages.

What is in the lab right now.

Each one opens onto its own page: the question it is testing, how it works, and every measurement we have taken. The headline figure is the single most useful thing that experiment has told us so far.

Kyori

Can a string score alone order an autocomplete list as well as the libraries applications already use?

The ranking cost inside our open-source JavaScript library: four ordered rules and a length tie-break, with no popularity data and no model. Kyori 0.7.0 is measured on held-out AmazonQAC searches against Fuse.js, fuzzysort, match-sorter, sorting by length and popularity — ties reported, every gap tested.

How it works, and what it found Use it
0.404

mean reciprocal rank on 11,607 held-out prefixes — level with Fuse.js, fuzzysort and sorting by length; popularity adds 0.018

Promptivo

Can prompt quality be measured without asking a model?

A deterministic scorer for AI prompts: seven research-backed merits, a regression calibrated once against an LLM judge and then frozen into coefficients. It runs with no model in the loop, so the same prompt always scores the same, and no model ever reads what you paste.

How it works, and what it found Use it
ρ = 0.53

against a held-out judge consensus — where two frontier judges only manage ρ = 0.61 with each other

Questions about the lab

What is Wizhut Research?
The research lab of Wizhut.tech, the company behind its Mac apps, developer tools and games. It builds small, working products to test ideas that have no obvious answer yet, and publishes what each one finds: the method, the measurements and the limits of the evidence.
What are the research projects?
Two, at the moment. Kyori asks whether a string score alone can order an autocomplete list as well as the libraries applications already use. Promptivo asks whether the quality of an AI prompt can be measured without asking a model.
Are they free to use?
Yes. Kyori is an open-source JavaScript library under the MIT licence, on npm and GitHub. Promptivo is free, permanently: there is no quota, no tier and nothing to buy, and an account unlocks the full report.
How is each project measured?
Against what people already use, on data held out from the design. Kyori’s evaluation tests every gap between methods for significance before calling it a gap; Promptivo’s sets its agreement with an LLM judge beside how well two judges agree with each other. Each project’s page says what its evidence does not cover.
Where are the results published?
On each project’s page here: the question, the method, the measurements and the limits of the evidence, tied to the version it measured. Papers from the lab are listed under Publications once they are published, with BibTeX to cite the lab’s software.
What happens to a prompt I score in Promptivo?
It is scored on Promptivo’s servers with no model call, so no model reads it. Each scored prompt is kept with its score as the experiment’s research data and never shared with a third party, so leave secrets and personal data out.

Real research on real problems.

Try PromptivoKyori on npm