The Knowledge Outlived the Engine — How a Discontinued Project Became Two, and Then One Brain

Most software dies quietly, and its ideas die with it. This is a story about a project that was shut down — and about the two things that grew out of it precisely because someone refused to let the good part go.

The project was a recipe optimizer. On paper it generated recipes; in practice it selected them — it shipped with fourteen hundred pre-written recipes and a solver that picked from them. What it genuinely had, buried in its data files, was a small, careful library of food science: several dozen culinary truths (why browning needs 140 °C, why glutamate and inosinate together taste far stronger than either alone) and several dozen flavor transformations (what heat does to an onion, what fermentation does to milk). It had the knowledge. It never built the engine that knowledge was for. So it was discontinued.

Two engines from one substrate

Here is the decision that this whole story turns on: instead of archiving the repository and moving on, we harvested the knowledge substrate — 54 culinary truths and 48 flavor transformations — and forked it into two new projects, each built to do the thing the original never did.

One became a nutrition engine. It keeps those truths as what they are — cited claims — and refuses to let a language model invent facts at serving time; models may author and judge knowledge offline, but the thing that answers you only ever repeats what a human verified. On that foundation it grew an entire nutrition and health library: iron bioavailability, nutrient timing, dietary patterns, the environmental cost of an ingredient — each entry graded from established to bleeding-edge, each carrying a field that most knowledge bases never bother with: what would change our mind.

The other became a flavor simulator. It took the same transformations and, instead of quoting them, ran them — building a deterministic engine that evolves a dish’s flavor forward through each step of cooking and scores whether the result is coherent. It has so far built about a fifth of those transforms into working rules; the rest it lists openly as unbuilt. On that foundation it grew a sensory-science apparatus the nutrition engine never needed: the chemistry of aroma, the physics of heat, and even chemesthesis — the “third sense” that carries chili heat and menthol cool, which is neither taste nor smell.

Same genetic material. Two completely different organisms.

The discipline they share

What makes them siblings isn’t only the shared data. It’s a shared refusal.

Both engines are built on one rule — cite everything, guess nothing — and both enforce it structurally, not as a slogan. When the flavor simulator can’t derive part of a dish, it doesn’t fill the gap; it writes “I don’t know” into its own output as a typed value, and that honesty lowers the score. When the nutrition engine hits a claim the evidence doesn’t yet support, it says so in the entry itself. Neither one is allowed to launder uncertainty into confidence.

You can see this most clearly in the two projects’ most public admissions of ignorance. The nutrition engine’s flagship research piece is about a gap — the fact that we still don’t know, with real confidence, how to feed half the population, because women’s health has been chronically under-studied (a caution meant to protect women that ended up freezing the research that would serve them). The flavor engine’s own honest confession is quieter but identical in spirit: every coherence score it produces is “consistent with the cited chemistry,” never “tastes good” — because it hasn’t been validated against a human palate yet, and it says so. One gap is in nutrition, one is in flavor. It is the same honesty, twice.

That is why, when the flavor engine scores a real dish, it will happily return a 0.958 instead of a 1.0 — and why we consider that a feature. An honest 0.958 is worth more than a confident 1.0, because the 0.958 is telling you where it’s unsure. A model that can write “I don’t know” in its own output is more trustworthy than one that can’t.

The second brain

Which brings us to why this story exists as a map and not just an essay.

If you sweep both engines for every branch of science they stand on — the browning chemistry, the reaction kinetics, the trigeminal neuroscience, the nutrition and metabolic literature, the environmental lifecycle science — and you plot them as a graph, something appears that neither project shows on its own: two constellations, grown from one root, rejoining at the places they still share. Taste physiology belongs to both. The “cite everything, guess nothing” discipline runs through both like a spine. And the frontier — the fermentation science neither has mapped, the odor-mixture math no one has solved, the enzyme families still unmodeled — fades honestly at the edges, drawn as faintly as our confidence in it.

We built that map as a living thing, and we made it a standing rule that any new verified citation, in either project, gets added to it. It is, in the most literal sense, a second brain for two engines that were born from the same discontinued one — and it is designed, above all, to show you where it ends.

This is a first-party account, grounded in the two projects’ own provenance records. Every scientific claim it references is one already verified inside one of the engines. Where the knowledge is thin, we’ve drawn it thin — because on this, as on everything else, the honesty is the point.