Your Worth a look page lists recipes whose stored nutrition is really the whole dish rather than one serving. Most of those can be fixed by simply dividing by the serving count, and the site already offers that. 26 could not, because the serving count itself was missing or wrong, so dividing would only launder one bad number into a legal-looking one.

Those 26 are what this pass looked at. Every recipe's title and full ingredient list went to an AI model, which was asked one question: how many servings does this actually make?

Applied on 2026-08-02. The 15 below are now corrected in your collection, and their nutrition was re-divided to match. Worth a look went from 70 recipes to 56. Both the old serving counts and the old nutrition are backed up, so any of it can be put back.

The 26 split three ways, and only 4 of them still need you.

How these were checked, not just guessed

An AI serving count on its own is a confident-sounding guess. Each answer was then tested against the same bars your site already uses to decide whether a plate is believable: the whole dish divided by the proposed count has to land between 50 and 1,200 calories a serving and under 100g of protein, and the model had to report at least 0.6 confidence in its own answer. Anything that failed any one of those is in the lower sections, not the approved list. That is why the count came out at 15 rather than 26.

Applied (15)

Each of these lands at a sensible plate once the serving count is corrected.

Not a servings problem (7), and now repaired

These were in the wrong queue: their whole dish totals were already tiny, so no serving count could repair them. A whole cherry pie recorded at 148 calories is a failed ingredient match, not a portion question.

All seven have since been run through the nutrition recovery tools and rebuilt from the USDA database. Five came back clean on the first pass. The last two, plus a chicken casserole, needed a second look: the engine matched every single ingredient line at full confidence, but its calorie figure contradicted its own macros. Instant Pot Frozen Chicken claimed 188 calories while its fat alone accounted for 182. Where every line matched, the macros are the trustworthy part, so the calorie figure was recomputed from them.

Banana Pudding at 1,028 calories a serving is high enough to be worth your eye, though condensed milk and whipped cream make it plausible.

Genuinely ambiguous, your call (4)

The model was asked not to invent a number when the ingredients are too vague to judge, and on these it said so. Three are single-serving mug recipes where the recorded count is probably right and the nutrition is simply enormous for one mug, which is a different question.

What this cost

Roughly half a cent, run on your OpenRouter key rather than the Claude plan. Three of the four batches had to be run twice: the model spent its entire output budget thinking and returned nothing, which cost about a tenth of a cent to learn.

The full nutrition pass, 2026-08-02

After the servings fix above, the rest of the queue got worked through properly. Worth a look went from 70 recipes to 23, and the repair queue from 60 to 10.

The queue was lying to you, twice

It ignored your own servings edits. It read the original imported count and never looked at your correction, so recipes you had already fixed kept being listed. That is now fixed in the code, and it is why 9 of the first 15 never needed an AI at all.

Eighteen recipes were filed as "too much protein for a dessert". They were nothing of the sort. They were whole-dish totals that happened to sit under the calorie alarm, so only the protein line tripped. Fourteen divided cleanly.

Ten are left, and none of them are lazy leftovers

• Three mug recipes (a one minute cobbler, a mug cake, a mug bread pudding) read 1,800 to 3,000 calories for a single mug. Recomputing gave the same answer, so the ingredient quantities themselves look like a full batch while the recipe claims one serving. That needs a human reading the original.
• Three chicken dishes where the recomputed macros do not add up to the calories. Writing those in would replace one wrong number with another.
• Two are not problems at all. Bread and butter pickles at 29 calories and cloud bread at 25 calories are simply light food.
• French Onion Soup claims 26 calories and 30g of protein in the same breath, which cannot both be true.
• Traditional Cornbread Dressing recomputes to exactly what is already stored, so there is nothing new to learn from the engine.

Every one was left alone deliberately. A number I cannot stand behind is worse on your site than a gap you can see.

Two tools were quietly broken and are now fixed

The nutrition boost scripts had the temp folder of the session that first ran them written into the code. That folder dies with its session, so every run since read a stale checkpoint and printed "0 to do, ALL DONE" while doing nothing. They now take the folder as a setting.

How to undo any of this

Every previous value was saved before it was replaced, in the same backup the site's own repair button uses. Tell me which recipes and I will put them back exactly as they were.