FMC2 · Level 2 cheap pipeline · Nov 8, 2026

K0 to page build, for each kind of input

K0 sorts every submission into a kind. Pick a kind below to see the path its recipes take, step by step, and where that path differs from a plain recipe. Pick Web, Photo or PDF to see how the source changes the reading.

What runs CodeHaikuHaiku, thinkingSonnetA model, not yet chosenNothing runs

How sure Testedmeasured in the Nov 5 to 7 trials Built, not re-testedfrom the Nov 2 pipeline Not built yetdecided or written down, no code Proposed, untested

This kindChanges for this kindmarks a box where this kind's path differs from a plain recipe.

"Clean" means no difference a cook would act on (catalog kinds 28 to 35) against L2 corrected. It is DB26's stand-in for "material" until your review marks decide it. L2 corrected is a cross-check, not an answer key, and it dates from Nov 5, 12:05 AM MST, before your Nov 5 to 7 rulings. "Best chain" is DB26's best end-to-end run (Nov 5): the line picker, K2's Haiku arm, K7 Haiku, K5 Haiku with the gap check and the gap-fill, and K6 Haiku with low thinking.

1. Every submission: Level 1, then K0

L1Input

Web page

Level 1 tags every line with its role, taken from the page: title, ingredient, method.

425 of the 988 recipes

L1Input

Photo or PDF

Level 1 is plain transcribed text, with no role tags.

450 photo and 113 PDF recipes

K0 · Sort

TestedHaiku, thinking

K0 reads the whole submission once and marks four things (R130):

structuremain, several, collection or none

hassection copy, printed sub-recipe, companion, add-on variant, branch variant, pointer to a recipe not in the submission, fragment

completenesscomplete, or partial with what is missing: start, end, words, ingredient list, method, steps, pointed recipes

sourcerecipe, dish_photo, menu or not_food

Method HE: Haiku 5.5 lists what the page shows by line number, then decides, with thinking at medium effort and the code reader's findings as places to check. Pack "A plus section copy", 19,160 words. Material marks right on 719 of 723 pages against Opus high, $0.72 for all 723, 3.7 seconds a page. Its version that also lists each recipe's lines does the split into recipes (Nov 5: it matched L2 corrected on 672 of 685 one-recipe pages and 33 of 38 several-recipe pages).

K0 sends each submission down one of these paths

A submission can carry several marks at once, such as a partial recipe that also has a section copy. Its path then picks up the changes of each kind it carries.

2. The path for

Steps are drawn in an order they can run. K7 sits before K5 because K5r needs K7's choice points; the two can run side by side. The old K8c (page structure and Recipe Records) is gone: K0's split and the page build now do its job.

3. All kinds side by side

Each column is a kind; each row is a step. Tap a column's name to draw its path above. The colored edge shows the strongest model a step uses there; dashed means proposed or skipped.

Where the misses are today

Best chain end to end: 582 of 946 recipes clean (62%), TEST 326 of 406 (80%). It cost $5.86 for all 988 recipes, about $0.006 each, with a median of 18.5 seconds a recipe. Per submission, 553 of 721 pages came through clean, against 309 for the old whole-page pipeline.

Of the 364 recipes that were not clean: 150 collection rows, 61 recipes on a page of several, 58 partial, 56 plain, 24 with a section copy, 10 printed sub-recipes and 5 with other marks. 283 of them are photos.
Only 33 of the 364 had a K5 difference and nothing else, and 22 of those are partial recipes. The other 331 differ at K2 or K3, mostly an ingredient line or a row missing. Reading the lines is the weak link, not writing the steps.
By source, best chain: web pages 365 of 420 clean (87%), PDFs 85 of 111 (77%), photos 132 of 415 (32%).
DB26 ran K1 to K6 on L2 corrected's split into recipes, not on K0's. K8a, K8b, K9 and K10 come from the Nov 2 pipeline and were not re-tested per recipe.