We build from evidence. Not from hype, and not from willpower — every product decision has to trace back to something you can check.
Who builds to this standard →A day is 24 hours. That number is not negotiable for anyone.
If the problem is “I want to, but I have no time,” the only move left is not to find more hours but to raise the result you get per hour. Once you accept that the supply is fixed, the design question changes: not “how do we get more of your day” but “how much further does the same hour take you.”
So mibuki does not build ways to try harder. We build products that implement the science of efficiency. We do not try to increase how much willpower you have; we increase how far the willpower you already have gets you.
The three principles below are the criteria we actually use. They are not sentiments — they are the rules that decide whether a feature ships.
Anything we consider building has to pass all three.
Effect size over opinion
We do not ship a feature because it looks promising. Before adopting a learning method or a behavioural design, we check how large the effect is, whether it replicates, and where it applies. That means reading past the original paper into pre-registered studies, systematic reviews and meta-analyses, depending on what the question needs.
We adopt something because it has been shown to work, not because it is popular. The reverse holds too: if it has not been shown to work, it does not ship, however loud it is. What matters is not the name of the method but who it worked for, by how much, and under what conditions.
In practice — LangDict uses spaced repetition. Whether that was the right call is something we settle with prediction accuracy rather than opinion (principle 03). The method, the data and the scores are all public.
See the Lexa research report →Remove friction, not add willpower
People do not quit because their willpower is weak. They quit because the design has friction. The number of taps before you start, the number of settings, the number of decisions you are asked to make — all of it decides whether a habit survives.
Instead of telling you to try harder, we delete steps. Fewer settings, usable at defaults, fewer decisions. The goal is a product that needs no willpower at all.
Measure it, then publish it
If we claim an effect, we publish it in a form you can check. This is a constraint on us, not a feature: if we cannot produce a number, we do not claim a result.
And we do not publish only the conclusion. We publish the baseline, the data and the metric as well, because a number without its conditions can only ever flatter us. A result is ready to publish when a reader could follow the same steps and see the same thing.
In practice — LangDict's memory model Lexa is compared against the default parameters of FSRS-7 on three public benchmarks (Anki / Duolingo / maimemo). The baseline is the stock configuration, with no per-user optimisation applied. The metric, the split and the scores are all published.
Read the Lexa research report →These promises are stronger than the principles, because you can catch us breaking them.
No “learn 5x faster.” When we publish a number, the baseline and the measurement sit in the same sentence as the number.
We do not hide the cancel path and we do not place buttons to catch mis-taps. If you want to leave, leaving is easy to find.
No notifications that hold a streak hostage. The number we want to grow is not your time in the app — it is the time you get back.
No reward theatre that makes progress feel real without any learning behind it. If the only thing that moved was a counter, nothing moved.
The answer is different per product, so here it is per product, with the reason.
Nothing that interrupts focus or learning carries ads. Lofi Station sells focus; putting something designed to break focus on the same screen does not add up. LangDict is the same case — anything that pulls attention mid-session works against the point of the app.
kikeruka and Atom do carry ads. They are free, and we have not yet built enough of a difference to reasonably charge for a paid tier. Right now there is no way to pay to remove them in either app. When there is a paid tier worth charging for, we will either drop the ads or make them something you can turn off.
So this is not a principle — it is where the products and the funding actually are right now. When that changes, this page changes.
| Product | Ads | Why |
|---|---|---|
| Lofi Station | None | A tool for focus should not host things that break focus |
| LangDict | None | Nothing that pulls attention away mid-session |
| kikeruka | Yes (no way to remove) | To keep it free |
| Atom | Yes | To keep it free |
A principle that is not implemented is not a principle. Here is where each one lands today.
| Product | Principles | How it shows up |
|---|---|---|
| LangDict | 01 + 03 | Uses spaced repetition, and publishes how accurately it predicts on public benchmarks |
| Atom | 01 + 02 | Habits designed as minimal units and triggers, following behavioural science |
| Lofi Station | 02 | Focus the moment it opens — the friction of configuration was removed |
| kikeruka | 02 | Scan and wait; checking back is no longer something you have to do |
open-spaced-repetition. FSRS: Free Spaced Repetition Scheduler. GitHub.
open-spaced-repetition. srs-benchmark: Benchmark of Spaced Repetition Algorithms. GitHub.
open-spaced-repetition. anki-revlogs-10k: Anki review logs from 10,000 users. Hugging Face.
Settles, B., & Meeder, B. (2016). A Trainable Spaced Repetition Model for Language Learning. Proceedings of the 54th Annual Meeting of the ACL.
Ebbinghaus, H. (1885). Über das Gedächtnis: Untersuchungen zur experimentellen Psychologie. Leipzig: Duncker & Humblot.
We list only works whose author, year and venue we have checked against the original.
Last updated: July 2026