A sourdough bake logevery loaf recorded, compared and improved

Recipe No. 26Kitchen and table

For: A home sourdough baker who wants to know why last Saturday’s loaf worked
You bring: A starter, a kitchen scale and thermometer, a notebook of past bakes, and the formulas you keep coming back to
You get: A document per formula with every bake as a numbered finding, the same table of numbers on each, crumb and crust notes, the best bakes marked Key, and your assistant’s comparisons of what changed
Time: An evening to set up, then five minutes during each bake and five after
Works in: know.sh on the web · the iOS app · your AI assistant, through MCP
Keep it: Private. This one never goes on a public link.

Sourdough rewards the baker who writes things down and punishes the one who remembers. The loaf that opened beautifully three weeks ago had a warmer dough, or a younger levain, or two more hours in the fridge, and by now nobody knows which. Notebooks fill with numbers that are never compared, because comparing twenty pages of pencil is a job for a wet Sunday that never comes.

This recipe makes the comparing easy. Each formula is a document whose overview holds the formula in baker’s percentages. Each bake is a numbered finding with the same table every time: flour, hydration, levain, dough and room temperatures, bulk and proof times, bake, and outcome, followed by notes on crumb, crust and taste. The best bakes are marked Key. The assistant you already use, Claude, ChatGPT or a small local model connected to know.sh, does the typing up: it turns your notebook and your bake-day scribbles into those tables. Then it reads the log and tells you what the good bakes have in common and what the dense ones do differently.

The loaves, the starter and the crumb shots stay in your kitchen and your camera roll. The log holds the numbers and what they taught you.

What you will use

Shelf: One shelf, Bread, with a line on your oven, your flour and your kitchen, since every number in the log depends on them.
Research document: One document per formula: Country loaf, 78%, Rye, 40%, Seeded whole wheat. The overview is the formula and the process as you bake it now.
Finding: One finding per bake, filed as Evidence; the bakes you would repeat marked Key; a Decision when the formula changes; a Question for the next experiment.
Your AI assistant: Types up the old notebook as a finding per bake, turns bake-day scribbles into the bake table, and compares bakes to find the variable that moved.
The editor: Where you check each table against your notes, fix the rows the assistant misread, and type notes into today’s bake as it happens.
The iOS app: The iOS app, now in TestFlight beta, on the counter during the bake: the formula open and today’s finding in Edit.
Revisions: Every earlier version of a formula, so the 75% loaf you baked in spring is still there after you move to 78%.
Look up: Look up (⌘K) finds every bake that mentions cold retard, levain or oven spring, in its title or anywhere in its notes.

Method

1
Write each formula as a document
Make a shelf called Bread and describe your setup in its line: “Dutch oven, gas oven that runs hot, unbleached bread flour at 12.7% protein, kitchen at 68–72°F.” Then make one document per formula. Its overview is the formula in baker’s percentages and grams, and the process in short steps: mix, autolyse, levain in, folds, bulk, shape, proof, bake.

A new bread gets a new document. A change to this bread, say 75% hydration to 78%, is a change to the overview, recorded in a Decision finding. Revisions keep the old formula.
2
Have your assistant type up the notebook
Photograph the notebook pages and give them to your assistant with the formula. Ask it to file each bake as a finding in that formula’s document, titled Bake 1, 3 August — 75%, dense, with the same table every time: flour, hydration, levain, dough temp, room temp, bulk, proof, bake, outcome. Thirty bakes become thirty findings in one go.

Then open each in the editor beside its page and fix what it misread. There are no buttons for making a table in know.sh; tables arrive from your assistant or from Markdown you paste into the editor. That suits a log, because once the first table is right, every later one is a copy.
3
Record every bake the same way
Title each bake with its number, date, the one thing you changed and the outcome: Bake 14, 21 September — 16-hour cold retard, open crumb. Titles are what Look up shows first, so cold retard and open crumb are easy to find.

Under the table, write three short headings: Crumb, Crust and Next time. Describe what you see in plain words: “even, glossy, a few large holes under the ear”. The Next time line is the most useful sentence in the log.
4
Keep the phone on the counter
The night before, ask your assistant to add an empty finding for tomorrow, Bake 18, 28 September. On bake day, open it in the iOS app. With flour on your hands you will not fill in a table, so tap Edit and type the numbers as they happen: “levain peaked 9:40”, “dough 77°F after mix”, “bulk ended 2:10, about 60% rise”.

After the bake, ask your assistant to turn that finding’s notes into the bake table in the same shape as the previous bake, and to leave anything it could not place under a heading of its own. Read the table against your notes; an assistant can misread a scribble as easily as you can.
5
Mark the bakes worth repeating
When a loaf is the one you want again, mark its finding Key and highlight the line in its notes that you think explains it. Highlighted bakes are the ones Open at returns you to, so the latest lesson is always one tap from the title page.

Be strict. If every decent loaf is Key, the log stops telling you anything. Three or four Key bakes per formula is a good sign you are learning what the formula wants.
6
Ask your assistant what changed
Once a formula has a dozen bakes, ask your assistant to compare the Key bakes with the ones that disappointed and to list every row of the table where they differ, with the size of the difference. The variable you suspected is often not the one that moved: the dough temperature crept up in August, or the levain went in past its peak.

Turn what you learn into a Question for the next bake, “Does a 76°F dough fix the tight crumb?”, and change one thing only.
7
Change the formula on purpose
When a question is answered, edit the formula in the overview and file a Decision saying what changed and which bakes settled it: “78% from bake 11 on; bakes 11, 14 and 15 at 78% all opened better than bakes 8 to 10 at 75%.” Link the bakes by pasting their addresses.

If a change turns out worse, open Revisions on the overview and restore the earlier formula. The bakes that tested the change stay in the log as evidence.

Specimen

You
Using know.sh, open “Bake 18, 28 September” in the document Country loaf, 78% on my Bread shelf. Turn my notes into the same bake table as bake 17, and keep my crumb and crust notes under their own headings.
searchFound Country loaf, 78%, 17 and 18
finding_getRead “Bake 17, 25 September” for the table rows
finding_getRead “Bake 18, 28 September”: 11 lines of notes
finding_updateRewrote “Bake 18, 28 September” as a 9-row table with Crumb, Crust and Unsorted
A local model through Ollama
Done. I could not tell whether “5 1/2” in your notes was the bulk time or the proof, so it is under Unsorted. Please check the table against your notes.

A small model running on the baker’s own computer does one modest job: turning bake-day notes into the table, and saying what it could not place.

BreadCountry loaf, 78%

14of 17

Bake 14, 21 September16-hour cold retard, open crumb

Evidence, key finding, 1 highlight, 386 words

Row	This bake
Flour	400 g bread, 100 g whole wheat
Hydration	78% (390 g)
Levain	100 g, at peak, 3 h 15 min after feeding
Dough temp	76°F (24°C) after mix
Room temp	71°F (22°C)
Bulk	4 h 40 min, 4 sets of folds, about 60% rise
Proof	16 h cold retard, 38°F (3°C)
Bake	500°F (260°C) lid on 20 min, 450°F (230°C) lid off 27 min
Outcome	Tall, open crumb, blistered crust

Crumb. Open and even, the best yet. The longer retard gave a thinner, blistered crust without tightening the crumb.

Crust. Deep brown, blistered, crackled for ten minutes on the rack.

Next time. Keep the 16 h retard. Try the same at 80% to see where the dough stops holding its shape.

A Key bake as a finding: the same table of rows every time, then the notes that say what the numbers meant.

You
Using know.sh, read the document Country loaf, 78% on my Bread shelf. Compare the Key bakes with bakes 12 and 13, which came out dense. List every row of the table where they differ and by how much.
ChatGPT
The Key bakes are 11, 14 and 15. Bakes 12 and 13 differ from them in three rows; flour, hydration, proof and bake are the same.
- Levain. The Key bakes used it at peak, 3 to 3¼ hours after feeding. Bakes 12 and 13 used it at 5 hours; your note on 12 says “levain smelled sharp, past peak”.
- Dough temperature. Key bakes: 76–77°F (24–25°C) after mixing. Bakes 12 and 13: 71°F (22°C), on the two coolest mornings in the log.
- Bulk. Both dense bakes ended bulk on the clock at 4½ hours with about 30% rise. The Key bakes ran to about 60% rise, whatever the time.
The pattern points to under-fermentation from a cool dough and a tired levain, not the flour. Shall I add a Question for bake 18: “Warm water to reach a 76°F dough, and end bulk by rise, not the clock”?
From your library
1. Country loaf, 78%, 11: Bake 11, 7 September — 78%, first open crumb
2. Country loaf, 78%, 12: Bake 12, 10 September — tight crumb, flat
3. Country loaf, 78%, 13: Bake 13, 14 September — dense, pale
4. Country loaf, 78%, 14: Bake 14, 21 September — 16-hour cold retard, open crumb

ChatGPT, connected to know.sh, comparing good bakes with bad ones row by row, and offering to file the next experiment as a Question.

Prompts to try

Your assistant, connected to know.sh (Claude, ChatGPT or a local model)

Using know.sh, open today’s bake in the document Country loaf, 78% on my Bread shelf. Turn my notes into a bake table in exactly the same rows as the previous bake: flour, hydration, levain, dough temp, room temp, bulk, proof, bake, outcome. Keep my crumb and crust notes under their own headings. Put anything you cannot place under “Unsorted” instead of guessing.
Your assistant, connected to know.sh (Claude, ChatGPT or a local model)

Using know.sh, read the document Country loaf, 78% on my Bread shelf. Compare its Key bakes with every bake that is not Key. For each row of the bake table, say whether it differs between the two groups and by how much, and cite the bakes. Do not draw conclusions from a single bake.
Your assistant, connected to know.sh (Claude, ChatGPT or a local model)

Using know.sh, read the bakes in the document Country loaf, 78% on my Bread shelf from oldest to newest and list every change I made to the formula or the process, with the bake it first appeared in. Then check the overview matches the latest bakes and tell me where it does not.

Variations

Keep a document for the starter itself: a finding per week with feeding ratio, flour, temperature and how long it took to peak, so a sluggish starter has a history.
Log beer the same way in a brew log, where a batch is a finding and the table holds gravities instead of hydration.
Developing bread recipes for a book? The cookbook manuscript adds testing rounds and testers to the same idea.

Where it falls short

No photos. Crumb shots cannot go in the library; name them by bake number in your camera roll and describe the crumb in words, which is often more useful anyway.
Tables have no editing buttons. They arrive from your assistant or from Markdown pasted into the editor, and editing a cell by hand on a phone is fiddly, which is why the recipe has you type notes and let the assistant build the table.
Your assistant compares the numbers you gave it. It cannot see the loaf, and a pattern across five bakes is a hint, not proof. Smaller local models also call know.sh’s tools less reliably and many cannot read notebook photos; check what they write in Revisions.
The iOS app is in TestFlight beta, and neither the app nor the web works without a connection. With no signal in the kitchen, keep the paper notebook for the bake and type it up after.

A note on raw dough, allergens and other bakers’ formulas

Do not taste raw dough or levain. Flour is a raw ingredient and can carry harmful bacteria until it is baked. Label each formula with its allergens: wheat in almost every loaf, sesame in the seeded bread, and milk, eggs or tree nuts in any enriched or walnut loaf. Never call a loaf free of an allergen; flour dust travels.

There is no meat in a plain loaf, but if the log grows into filled breads, cook fillings to the USDA safe temperatures: 165°F (74°C) for poultry and 160°F (71°C) for ground meat. Jam for the toast belongs to a tested recipe from the USDA or the National Center for Home Food Preservation, not to an improvised one.

Baker’s percentages and ingredient lists are not protected by copyright, but the wording of a book’s or a blog’s method and headnote is. When a formula comes from someone else, write the process in your own words and note where it came from: “Adapted from the country loaf in [book], chapter 3.”

Indexed under

Sourdough Hydration Levain Cold retard Bake tables Crumb Baker’s percentages

What you will use

Method

Write each formula as a document

Have your assistant type up the notebook

Record every bake the same way

Keep the phone on the counter

Mark the bakes worth repeating

Ask your assistant what changed

Change the formula on purpose