Recipe Flowchart

The eval tool

Every one of the 60 runs below was scored by hand against a fixed taxonomy (click a cell, tag why it's wrong, no free text) in a local tool built for exactly this. It isn't part of this static site — it writes real annotation data to disk, which a GitHub Pages site can't do — so here's what it looks like in use instead.

Worked example

One recipe shown in full — the rendered graph plus how all 6 models did on it. Picked for size, not drama: compact enough to read without a lot of scrolling, but still dense with the kind of convergence (multiple ingredients folding into one operation) that makes a Gozinto chart worth looking at. It's also the most visually distinctive input in the set — a photo of my own handwritten, bilingual kitchen notes, with hand-drawn brackets already marking which sub-steps run in parallel.

Handwritten: Choux au Craquelin

Source: My own kitchen notes (bilingual shorthand, hand-drawn dependency brackets)

brown sugar115g	cream	mix until crumbly	press into rectangle6x8 rectangle	roll out and freeze12x14, 1/8 inch, freeze 5 min	cut rounds18x 2 inch rounds	pipe and top with craquelin	bake400F, 10-12 min	rest with oven off and door open30 min
butter115g
AP flour115g
salt1/8 tsp
water235g	boil panade170F	cool	mix in eggs
butter84g
sugar8g
salt2g
AP flour128g
eggs4 large

Model	Input tok	Output tok	Cost	Latency	Ingredients	Operations	Merges
Claude Haiku 4.5	3678	923	$0.0083	5.9s	10	8	5
Claude Sonnet 5	4485	1294	$0.0329	14.0s	10	11	5
GPT-4o mini	38125	532	$0.0060	8.7s	10	5	4
GPT-4o	2395	642	$0.0124	7.0s	10	8	3
Gemini 3.5 Flash-Lite	2553	834	$0.0029	3.5s	10	7	4
Gemini 3.5 Flash	2553	925	$0.0122	16.6s	10	9	6

All 10 recipes

Chosen to stress-test specific things, not just to pad the count — genuine parallel prep, raw unedited web-scrape noise, several variations on ingredients that split across steps, and one recipe added specifically to check whether the bug fix above generalized.

• Cooked — Serious Eats
• Baked — NYT Cooking
• Handwritten — My own kitchen notes (bilingual shorthand, hand-drawn dependency brackets)
• Web Scrape — Serious Eats (J. Kenji López-Alt)
• Pancakes — Serious Eats (J. Kenji López-Alt)
• Eggnog — Alton Brown
• Mac & Cheese — Food Network (Alton Brown)
• Potstickers — Damn Delicious
• Shredded Beef Tacos — RecipeTin Eats (Nagi)
• Tikka Masala — Savory Tooth

Same recipe, two different (and equally correct) structures

Both of these models scored zero human-flagged issues on the same recipe — and produced genuinely different graphs, the biggest structural gap of any pair that both scored clean. GPT-4o folds "boil water" into the potato-boiling step and does the whole roast as one operation. Claude Sonnet 5 splits both out: its own "boil water" step, and the roast broken into its two real phases (undisturbed, then flip-and-continue). Neither is more correct than the other — they're different, equally valid choices about how finely to decompose a continuous process, which is the real reason this project didn't try to grade against one canonical answer.