The Sourdough Experiment: design of experiments, explained with a loaf of bread
Six settings, eight bakes, one best loaf. Find it by gut feel or with a designed experiment, and see why changing one thing at a time so often misses the best recipe.
How the experiment works
The simulation turns a sourdough loaf into a small R&D problem. You control six settings: hydration, starter, salt, wholemeal flour, bulk temperature and bulk time. Each bake returns five results: loaf volume, crumb openness, sourness, crust colour and how easy the dough is to shape. A loaf score out of 100 combines them, weighted by what you care about.
Six settings, five results
Behind the scenes, the settings drive three things bakers will recognise: how much gas the yeast makes, how much acid the bacteria make, and how strong the gluten stays. Warmth speeds both microbes up, but above about 28 °C the yeast slows while the acid-making bacteria keep going. Salt slows both, and slows the bacteria more. Wholemeal flour soaks up extra water and adds tang. Leave the dough too long and the gluten weakens, so the loaf spreads instead of rising.
Loaf score = weighted geometric mean of the five results, so a poor result on anything you care about pulls the whole score down
Every bake is a little different, just like a real kitchen. In the home kitchen setting each result varies by about 10 points from one bake to the next. The test kitchen setting is steadier.
Gut feel or a designed experiment
In gut-feel mode you change whatever you like between bakes. In designed mode you pick two or three settings to study. The tool plans five bakes around your starting recipe, fits a simple model to the results and suggests three more.
Bakes 1 to 4: the corners of a half-fraction design. Bake 5: the centre. Bakes 6 to 8: the model's suggestions.
In the half-fraction, each setting goes up in half the bakes and down in the other half, so every bake tells you something about every setting. The catch is that some effects look identical after only four corners. Bake 6 starts to tell them apart.
Why one-at-a-time fails
The obvious way to improve a recipe is to change one thing, keep it if the loaf gets better, then move on to the next. It feels careful, but it has two weaknesses. First, settings interact. The best bulk time depends on the temperature: about 8 hours at 20 °C, but under 5 hours at 30 °C. Changing one setting at a time moves you along straight lines, and when the best region is a diagonal ridge, those lines slide past it. Second, noise fools single comparisons. One lucky or unlucky bake can send you the wrong way, while a designed experiment uses every bake to estimate every effect.
In 2,000 simulated games from the same starting recipe in a home kitchen, a designed experiment on hydration, temperature and time reached 91% of the best possible loaf on average. Changing one of those three settings at a time reached 84%, and changing settings at random reached 60%. The bigger difference is in the unlucky games: the designed approach's worst 10% still reached 82%, against 66% for one-at-a-time.
Choosing what to study matters most of all. A design on the wrong three settings does worse than careful one-at-a-time on the right ones.
What the simulation does not do
This is a simplified simulation based on published baking science, built to teach experimental design, not a recipe authority. The direction of each effect comes from the research listed below, but every number in the model is illustrative and has not been fitted to real bakes. Final proof, shaping and the bake itself are held constant. Sourness is a single number: in real dough, warm ferments make more total acid, while cool and stiff doughs shift the balance to acetic acid, which tastes sharper. Real bakers also cap salt for taste, which this simulation leaves out.
Worked example: the warm-day over-proof
This is the loaf the challenge starts from. The figures are for the everyday flour with the default scoring.
- Starting recipe: 70% hydration, 20% starter, 2.0% salt, 10% wholemeal, 7 hours at 28 °C
- Result: a flat, very sour loaf with a tight crumb and a dough too sticky to shape, about 30% of the best possible loaf
- Best loaf for this flour: about 78% hydration, 18% starter, 2.2% salt and 15% wholemeal, with 5 hours 45 minutes at 25 °C
- Cut the time alone, to about 4 hours 15 minutes: 66% of the best. The loaf rises more, but the crumb stays tight.
- Cool the dough alone, to about 20.5 °C: 70% of the best.
- Add water alone, to 77% hydration: 42% of the best. On its own, extra water barely helps an over-proofed dough.
- Cool the dough and add water together: 97% of the best. The two changes add far more together than their separate gains. That is an interaction, and it is exactly what one-at-a-time testing struggles to find.
Each game uses one of 25 flour batches, so the best recipe moves a little every time you play.
Six settings made this tricky. Formulators juggle forty, and that's what Uncountable is built for.
FAQs
Design of experiments is a way of planning tests so that a small number of runs shows how several factors affect a result, including how they interact. Instead of changing one thing at a time, you change several together in a balanced pattern and then fit a model to the results.
It misses interactions, where the best setting of one factor depends on another, and it relies on single comparisons that noise can easily mislead. In this simulation the best bulk time falls from about 8 hours at 20 °C to under 5 hours at 30 °C, so testing time at the wrong temperature points you to the wrong answer.
A full factorial design tests every combination of high and low settings. A fractional factorial tests a carefully chosen subset. Here, four bakes cover three settings, and each setting is high in half the bakes and low in the other half. The price is that some effects are confounded: they cannot be told apart until you add more runs.
No. It is a simplified simulation. The direction of each effect comes from published research on sourdough fermentation, salt, bran and browning, but the numbers are illustrative. Use it to learn about experiments, not to choose a recipe.
The simulation adds random variation to every bake, as happens in a real kitchen with a livelier starter, a hotter oven or different shaping. The variation is the point: it shows why one good or bad bake can mislead you, and why a design that uses all your bakes is more reliable.
Only the amount of variation. In the home kitchen each result varies by about 10 points from bake to bake, and in the test kitchen by about 5. With steady results, careful one-at-a-time on the right settings does about as well as a design. As results get noisier, the designed experiment pulls ahead.
Spot at Uncountable reviewed the model and the statistics before release, against the sources listed on this page. Any change to the model goes back to Spot for review.
No. The simulation runs entirely in your browser and stores nothing. If you share a result, the link carries your score, recipe and flour batch in the link itself, so nothing is saved on a server.
References
- Gänzle, M.G., Ehmann, M. & Hammes, W.P. (1998). Modeling of growth of Lactobacillus sanfranciscensis and Candida milleri in response to process parameters of sourdough fermentation. Applied and Environmental Microbiology 64(7), 2616–2623. pmc.ncbi.nlm.nih.gov
- Gänzle, M.G., Loponen, J. & Gobbetti, M. (2008). Proteolysis in sourdough fermentations: mechanisms and potential for improved bread quality. Trends in Food Science & Technology 19(10), 513–521. doi.org
- Thiele, C., Gänzle, M.G. & Vogel, R.F. (2002). Contribution of sourdough lactobacilli, yeast, and cereal enzymes to the generation of amino acids in dough relevant for bread flavor. Cereal Chemistry 79(1), 45–51. doi.org
- Belz, M.C.E., Ryan, L.A.M. & Arendt, E.K. (2012). The impact of salt reduction in bread: a review. Critical Reviews in Food Science and Nutrition 52(6), 514–524.
- Miller, R.A. & Hoseney, R.C. (2008). Role of salt in baking. Cereal Foods World 53(1), 4–6.
- Noort, M.W.J. et al. (2010). The effect of particle size of wheat bran fractions on bread quality: evidence for fibre–protein interactions. Journal of Cereal Science 52(1), 59–64.
- Gómez, M. et al. (2003). Effect of dietary fibre on dough rheology and bread quality. European Food Research and Technology 216, 51–56.
- Kim, J.H., Maeda, T. & Morita, N. (2005). Application of polished-graded wheat grains for sourdough bread. Cereal Chemistry 82(2), 144–151.
- Arendt, E.K., Ryan, L.A.M. & Dal Bello, F. (2007). Impact of sourdough on the texture of bread. Food Microbiology 24(2), 165–174.
- Purlis, E. (2010). Browning development in bakery products: a review. Journal of Food Engineering 99(3), 239–249.
- Calvert, M.D. et al. (2021). A review of sourdough starters: ecology, practices, and sensory quality with applications for baking and recommendations for future research. PeerJ 9, e11389. peerj.com
- Scanlon, M.G. & Zghal, M.C. (2001). Bread properties and crumb structure. Food Research International 34(10), 841–864. Cauvain, S.P. (2015). Technology of Breadmaking, 3rd edition. Springer.
- Derringer, G. & Suich, R. (1980). Simultaneous optimization of several response variables. Journal of Quality Technology 12(4), 214–219. Box, G.E.P. & Wilson, K.B. (1951). On the experimental attainment of optimum conditions. Journal of the Royal Statistical Society B 13(1), 1–45. Czitrom, V. (1999). One-factor-at-a-time versus designed experiments. The American Statistician 53(2), 126–131.
Model reviewed on 13 October 2026. Model spec version 1.0.
Disclaimer
The Sourdough Experiment is a free teaching tool. It shows how designed experiments work using a simplified model of sourdough baking. The results are illustrations, not measurements, and they are not baking, nutritional or professional advice.
The numbers in the model are illustrative. Real flours, starters, kitchens and ovens behave differently, so do not treat the simulated best loaf as a recipe.
We have taken care over the model, but the tool is provided as is. To the extent the law allows, Uncountable gives no warranty that the results are accurate, complete or suitable for any purpose, and accepts no liability for any loss that results from using them.
