The method
How we know what we say.
A fermentation model is easy to write and hard to defend. Ours is anchored in a certified organic kombucha factory that has been turning out batches every week since 2005 — and where it isn't, it says so.
The anchor
There's a house recipe — the kombucha one, the most repeated — that has been made hundreds of times under the same conditions: 70 grams of sugar per litre, 6 of tea, a 10 % starter, at 23 degrees. It's ready in seven or eight days. That point holds up everything else: any other combination the model calculates is measured against it.
A model without an anchor is a spreadsheet with opinions inside.
What we measure in a batch
Temperature of the liquid, not the room. pH at the start and at the cut. Starting sugar and residual sugar. The real day of the cut, and an assessment of how it turned out. None of that is exotic; what changes things is having it written down always, in the same format — which is exactly what almost nobody does, ourselves included for quite a few years.
Three origins, and they don't weigh the same
Every coefficient carries a mark of where it comes from, and that mark decides how much it weighs:
Measured. It comes from our batches. It outranks the other two.
Derived. It comes from published literature or trade standards — the sugar-to-gas and sugar-to-alcohol conversions are the ones used in brewing and winemaking, checked over decades. It holds until a batch says otherwise.
Adjusted. Set by judgement so the whole thing adds up. It's the weakest, and the first to move.
One method, one ferment at a time
The procedure is always the same: a real anchor, coefficients with their origin marked, and a check against fermentations we already know before publishing anything. What changes is how much of each one we've measured.
| Ferment | State of the model |
|---|---|
| Kombucha | Anchored in our own production. The most measured |
| Water kefir | Model built, calibration under way |
| Sauerkraut, kimchi, pickles, vinegar | In the queue, in that order |
We publish an Afinador — a fermentation tuner — when the model holds up to the check, not when the interface is ready.
What we don't know yet
The coefficient that turns consumed sugar into acidity is waiting on a series of pH measurements at the cut. We have two hints that it falls short, and we haven't touched it: changing it on two hints is guessing in better clothes.
Kombucha's alcohol is an industry-wide problem, not ours. Measurement methods disagree with each other, and in our own analyses the same kind of batch has produced results that vary more than anyone would like. That's why the model estimates and says it estimates: if what you ferment is sold, it has to be measured, not calculated.
How it gets corrected
Before touching code, the model is checked against fermentations we already know. That doesn't tune anything — it can't calibrate a coefficient — but it can knock the whole model down, and that's its entire usefulness: if it predicts four days where we never go under six, the problem isn't the number, it's the approach.
The coefficients aren't published. The method is — which is what can be judged.