Published Recipes
Searchable by course, cuisine, skill level and serving size
From recipe to executable cooking plan — authenticated JSON-LD files that cook alongside you
Cookpit: "shaped in large part by AI, instructed by broadbaseai.com" · Hero image: AGA in a cottage kitchen, by simonsimages [Flickr, CC BY 2.0]
A recipe is prose; a kitchen is real time. The written recipe cannot tell you that the sauce needs stirring now, that the oven should have been preheating ten minutes ago, or that if the starter is to land while the main rests, three tasks must run in parallel. Cookpit closes that gap: it turns a recipe into an executable cooking plan — a JSON-LD file with ingredients, equipment, phases, timed tasks on parallel lanes, sensory completion cues and a pre-start checklist — validated against a published rulebook and cryptographically signed before your cooking app is allowed to run it.
The result, for the cook, chef, hobbyist or baker: starters, main courses and desserts orchestrated live in your kitchen, each step arriving at the optimal moment an expert chef would commit to it — with the file's authenticity guaranteed by mathematics, not by trust in whoever generated it.
Searchable by course, cuisine, skill level and serving size
Parallel workstreams so multiple dishes cook at once
Generate → Validate → Attest → Consume, one actor each
A Cookpit file is JSON-LD — JSON with linked-data context, grounded in the schema.org/Recipe vocabulary and extended by the published Cookpit v3.2 schema. Machine-readable and human-auditable at once, every file declares the same structural blocks:
// Excerpts from the real authenticated file below { "@context": ["https://schema.org", { "cookpit": "https://cookpit.org/v3.2/" }], "@type": ["Recipe", "cookpit:CookingFile"], "name": "Creamy Garlic Salmon and Shrimp", "author": "Julie Maestre", // … a live-cook task on main-course lane M1 … { "time": "00:01:30.M1", "action": "Salmon in, skin side down. Three minutes — leave it alone.", "completion": { "type": "sensory", "cue": "underside deep golden; releases from the pan without a fight" }, "timingBasis": { "basis": "sourceRangeMinimum", "source": "Cook the salmon for 3-4 minutes on each side skin side down." } }, // … and the attestation that makes it trustworthy … "attestation": { "status": "authenticated", "issuer": "https://validator.cookpit.org/v3.2", "canonicalization": "RFC8785", "keyId": "cookpit-chefs-friend", "fileFingerprint": "4a7975f6c68bd1ca06d5b062a61ca9d4…", "signature": "1qx9lPWyhqh53rr6vRJnWrH901ArfSiJ…" } }
Creamy Garlic Salmon and Shrimp by Julie Maestre — main course, 4 servings, 25 minutes. A genuine authenticated .A.jsonld file (23 KB): open it in any text editor to audit every timing basis, or load it into CookChow and cook it tonight.
Anyone — or any AI — can generate a plausible-looking recipe file. Cookpit's discipline is what turns an untrusted AI draft into a trusted artefact. Every rule has an identifier and a sentence, published immutably and dedicated to the public domain under CC0.
"Every task time is the optimal moment an expert chef would commit to that action." No random, cosmetic or filler times — ever.
"The plan is bounded by the resources declared in this file." An undeclared pan is a hard validation failure, not a warning.
"The plan does not adapt at runtime." Execution and the clock belong to the app and the cook — never to the file.
Every quantity, duration and temperature in the source must appear with the value the source states, bound by a SHA-256 fingerprint the validator recomputes.
No task may carry a time without a stated basis drawn from a closed set of evidence types — an exact source duration, a range target, an endpoint of the cook time, a "meanwhile" cue, an implied deadline worked backwards from a downstream consumer, or a professional estimate with a one-line rationale. Each basis names the specific source line that justifies it. A model that quietly rounds a temperature or drops a range fails validation, however plausible its output reads. Where the source is ambiguous, an explicit priority order resolves it — ambiguity is never resolved silently.
Four lifecycle stages, one actor at each — so the entity that drafts a file is never the entity that vouches for it.
An AI drafts the cooking file from your recipe — and must mark its output unauthenticated. The AI contract is explicit: it "MUST NOT claim status authenticated under any circumstance" and "is not trust authority".
A browser validator — which never uploads your file — checks every rule: closure, identifiers, timing bases, fingerprints. A failing file is returned unmodified with a report; the validator never repairs it.
The validator — and only it — signs a passing file with an Ed25519 key and flips its status. The signed file downloads as .A.jsonld: authenticated, tamper-evident, portable.
Your cooking app verifies the signature against a pinned public key before it runs the plan. A tampered file — even one edited after signing — fails the fingerprint check and never reaches your kitchen.
Load an authenticated file into a live cooking app such as CookChow — "Live (Real-Time) Recipe Home Cooking" — and the plan becomes a running kitchen. The ten publication lanes are parallel workstreams: while the tart base blind-bakes on one lane, the filling reduces on another and the garnish preps on a third. Alarms are derived mechanically from each phase's declared window, so nothing depends on the cook remembering to set a timer.
Instructions arrive in a controlled register — imperative and exact, enforced by lint, with "a little while" forbidden. Every completion cue names a sensory fact, because a kitchen is judged by eyes, nose and fingertips, not by the clock alone. When the plan says the sauce is done, it tells you what done looks like.
And because each dish is a bounded, validated plan, courses compose: run the starter, main and dessert together and their lanes interleave, each step surfacing at its optimal moment so the starter lands as guests sit, the main follows as plates clear, and the dessert finishes as the coffee brews. That is the real-time promise — the coordination a head chef carries in their head, made available to the hobbyist on a Tuesday night.
The cook — dinner for the family with the timing handled, and no step forgotten at the school-run hour.
The chef — training plans and service prep expressed as validated files a brigade can run identically, shift after shift.
The hobbyist — ambitious multi-course menus that would otherwise demand a second pair of hands.
The baker — where minutes and grams are chemistry: exact times, faithful quantities, and proof the file matches its source.
The open cooking file format itself: the v3.2 specification, governing rules, lexicon, AI prompt bundle, browser validator and attestation service. Dedicated to the public domain under CC0 1.0; every versioned address is immutable forever.
The public recipe repository: 60+ published recipes across British, French, Italian, Indian, Turkish, Japanese, Thai and more — searchable by name, author or ingredient, filterable by course (starter, main, dessert), cuisine, skill level and servings, each downloadable as a Cookpit JSON-LD file.
A companion recipe corpus with the same search, filtering and assist features — demonstrating that anyone can publish a conformant Cookpit collection, because the format belongs to everyone.
Live (real-time) recipe home cooking: the consumer app that verifies a file's signature, then runs the plan — lanes, alarms, phases and sensory cues — as you cook.
Browse cookpit.recipes and pick your starter, main or dessert.
Download the authenticated .jsonld file.
Open it in CookChow and cook in real time.
Give your recipe (even a PDF) to a capable AI chatbot with the standard prompt from cookpit.org — it fetches the spec, computes the deterministic ids and the source fingerprint.
Drop the generated file into the browser validator — your recipe never leaves your machine.
Have the passing file signed, and it's ready for any Cookpit-compatible app — yours to keep, share or publish.
Cookpit shows what applied AI looks like when it is done with discipline: the AI drafts, an independent validator checks, a key-holder signs, and the app verifies before anything runs — with the whole standard published openly and dedicated to the public domain. It is a small thing about kitchens, and that is the point. The same validate-then-attest method that gets your lemon tart to the table on time is the method our Project Icarus work proposes for AI safety itself: trust that lives in the signature, not in the machine's word.
Ten lanes, mechanical alarms and sensory cues put a head chef's coordination in every home.
Fingerprints and signatures guarantee every quantity and timing matches the source recipe.
The Cookpit v3.2 JSON-LD schema/format is dedicated to the public domain under CC0 1.0 Universal. Recipe content itself is not covered by that licence. The JSON excerpt on this page is illustrative; cookpit.org/v3.2 is the authoritative schema.