RecipeFix

01 — Problem

Someone selects gluten-free and cooks what comes back. If the model ships wheat flour that is a medical event, not a bad recipe. The obvious fix is to check every diet with a rulebook, and the obvious fix is worse than nothing.

Approach

A diet is checked only where a violation is a property of the ingredient itself, independent of quantity. Presence-rules — vegan, vegetarian, dairy-free, gluten-free, nut-free, halal — are checked and block. Threshold-rules — keto, low-carb, low-sodium — are deliberately not checked, because a correct check needs per-ingredient nutrition keyed to quantity. Low-FODMAP is checked only for ingredients that are high-FODMAP at any normal portion, because the serving thresholds are licensed data the product does not hold.

Tradeoff

Anyone on keto gets no deterministic guarantee, and the product does not pretend otherwise. A regex there would be security theatre: it could not tell five grams of sugar from fifty, and shipping it would imply a guarantee that cannot be made.

02 — Problem

Diet compliance answers whether it is safe to eat. An integrity pass answers whether it is internally consistent. Neither answers the question the product is actually judged on: if you cook this, does it come out right?

Approach

Two checks that are decidable without judging taste. Orphan ingredients — listed but never used in any step, a pure defect. And uncoupled substitutions: if an ingredient changed and nothing in the steps or notes acknowledges the consequence, the instructions are still describing the original dish and it will fail in the pan.

Tradeoff

These warn rather than block, because unlike a diet violation there is no proof of harm — a recipe can legitimately mention a consequence in wording the check does not match. Blocking on a heuristic would throw away good recipes. The severity of an automated check is matched to the strength of the evidence behind it.

03 — Problem

Adaptation results are cached on a hash of the request. The original implementation selected which fields to hash with JSON.stringify's second argument, which does not do what it appears to do: it filters nested objects to the same key list.

Approach

The ingredient-scale object serialised as an empty object, so "2 lb chicken" and "9 cups flour" produced identical cache keys and users were served the wrong recipe. It was replaced with a hand-written recursive canonicaliser that deep-sorts keys and preserves nested structure before hashing.

Tradeoff

More code than the one-liner it replaced, and the one-liner looked correct — which is the entire lesson. The failure is recorded in the comment above the function so nobody reintroduces it.

04 — Problem

When the diet gate fails there are two repair strategies: a targeted patch, or a full regeneration. The patch costs one extra model call if it works and two if it does not, because the failed patch still has to be followed by the regeneration. Which is right depends on how often the patch works, and nobody knew.

Approach

Both were instrumented, and the metrics endpoint reports the number that decides it. Over the last thirty days: three repairs triggered, zero repaired by patch, three fell back to full regeneration.

Tradeoff

The measurement currently says the patch stage is pure overhead on the slowest path in the product. The sample is three. So it stays, instrumented, until the number means something — which is a less satisfying answer than deleting it.