What Will It Taste Like Before You Cook It? How Ajinomatrix Accidentally Discovered Predictive Cuisine

A recipe experiment exposed a missing bridge between structured recipes,

sensory simulation and what people actually perceive after cooking.

Image: the flowchart of predictive cuisine by Ajinomatrix online JRF+MP6 Studios - you can enlarge it!

We were not trying to invent a new field.

We were trying to finish a drink.

The starting point was a recipe recovered from a Facebook video: a fermented

mixture of citrus, honey, ginger, herbs and spices carrying the memorable name

“Elixir of Invincibility.”

The name is part of the recipe’s identity. We make no claim that it makes

anyone invincible — or that it has immunity, detox, probiotic or other health

effects.

The first step was familiar.

We used the Ajinoverse recipe workflow to extract the source into a structured

recipe object.

It worked — but not perfectly.

Some ingredient information was incomplete.

That was useful.

Instead of pretending the extraction was authoritative, we moved the recipe

into AJNV4.3, completed the missing structure and froze a versioned protocol:

Elixir of Invincibility v1.1.0.

Now the recipe existed as more than text.

It contained ingredients, quantities, process stages and an explicit

transformation sequence — including a five-day fermentation at approximately

21 °C followed by straining and bottling.

Then something more interesting happened.

AJNV4.3 generated an AI-assisted sensory projection of the resulting product.

For the first time in this workflow, we could see not only what we intended to

make, but what the system expected the result to be perceived like.

The prediction appeared as MP6-style spider charts across six dimensions of

perception.

And the immediate question was:

Can we carry that sensory prediction with the recipe?

THE ABSURD STEP

At that moment, we could see the prediction — but we could not yet export the

final predicted profile as a native MP6 file.

So we did something deliberately awkward.

We captured screenshots of the spider charts.

Then we reconstructed the approximate values into an .mp6 file by hand.

And we marked that file very clearly:

simulated;

unvalidated;

reconstructed from visual output;

not measured;

not observed by a panel.

That awkward workaround turned out to be the most useful part of the

experiment.

Because it showed us exactly what was missing.

If a recipe system can predict a sensory state, that prediction should not end

as a picture.

It should become a portable, versioned sensory object that can travel with the

recipe.

FROM RECIPE TO SENSORY STATE

That insight gives each Ajinomatrix layer a very clear role.

JRF describes the culinary object:

What are we making, and how?

AJNV4.3 reasons over the structured process:

What does this recipe-process imply?

MP6 represents perception:

What is the result expected — or later observed — to be like?

FEELLLM packages the sensory state, the user’s response, the context and the

decision question so that a human or downstream AI can reason about what to do

next.

In one sentence:

JRF defines the culinary object.

AJNV4.3 computes its transformations.

MP6 makes predicted and observed sensory states portable.

FEELLLM contextualises those states for decisions.

THE SECOND HALF OF THE IDEA

A prediction alone is not enough.

The more interesting possibility comes after cooking.

Before execution, a recipe version can carry:

MP6-SIM

— a simulated sensory profile.

After execution, the same recipe version can receive:

MP6-OBS

— an observed sensory profile.

Only attributes that genuinely use the same semantic definition, scale,

recipe state and compatible context should be compared.

Missing values are not zeros.

The comparison therefore happens only on a declared comparable set.

For an aligned sensory attribute j:

observed value

minus

predicted value

gives a residual.

But the residual is not interesting merely because it is a number.

It becomes useful because the prediction and observation remain bound to:

the exact recipe version;

the exact sensory vocabulary;

their provenance;

their evidence type;

their model or observation context.

That turns disagreement into data.

PREDICT → COOK → OBSERVE → COMPARE → LEARN

This is the workflow we now call Predictive Cuisine.

Not “AI generates recipes.”

Not “AI perfectly predicts taste.”

Predictive Cuisine is sensory digital prototyping of structured recipes.

The loop is:

Predict

→ Cook

→ Observe

→ Compare

→ Learn

A recipe can be duplicated.

One ingredient or process variable can be changed.

A new sensory profile can be predicted before cooking.

Then the recipe can be prepared.

Then the actual result can be recorded.

And the predicted change can be compared with the observed one.

WHY THIS MATTERS

Food development today still depends heavily on iterative physical trials.

That is not inherently bad — cooking is physical, sensory and contextual, and

no simulation replaces tasting.

But digital prototyping can potentially make experimentation more explicit.

Instead of:

“I think this version might taste brighter.”

we can eventually preserve:

Recipe V1

Predicted MP6 V1

Observed MP6 V1

then:

Recipe V2

Predicted MP6 V2

Observed MP6 V2

and ask:

Did the predicted direction of change actually happen?

THE ELIXIR ALSO TAUGHT US SOMETHING ABOUT WASTE

The experiment exposed another process-design question.

When the fermented mixture is strained, the process does not really produce

one output.

It produces two:

a liquid beverage;

and aromatic solids.

Instead of automatically treating those solids as waste, they can become the

input of another recipe — for example a citrus preserve.

Conceptually:

STRAIN

├── beverage

└── aromatic solids → continuation recipe

Process inputs, outputs and branching are not new ideas in process modelling.

What this case did was motivate a concrete JRF requirement:

a culinary process should be able to create multiple typed outputs and carry

those outputs into subsequent recipe branches.

WHAT WE HAVE DEMONSTRATED — AND WHAT WE HAVE NOT

We have demonstrated, manually and once:

recipe extraction;

repair and versioning in AJNV4.3;

process-graph representation;

AI-assisted sensory projection;

reconstruction of a provenance-labelled predicted MP6;

FEELLLM contextual packaging;

a downstream AI interaction that asked process-relevant questions and returned

a recommendation.

We have NOT demonstrated:

validated sensory prediction accuracy;

generalisation across recipes;

autonomous recipe optimisation;

FEELLLM superiority over an ordinary prompt;

digital-twin operation;

panel replacement;

health effects.

Those are not hidden limitations.

They define the next experiments.

THE NEXT ENGINEERING BUTTON

The most obvious missing feature now has a simple name:

Generate MP6 Preview.

Instead of:

sensory spider chart

→ screenshot

→ manual reconstruction

→ .mp6

the workflow should become:

versioned recipe

→ Generate MP6 Preview

→ native MP6-SIM

The exported file must preserve:

recipe identity;

recipe version;

model/method;

simulation provenance;

confidence or coverage;

missing values;

and the fact that prediction is not measurement.

THE NEXT SCIENTIFIC QUESTION

The conceptual-definition phase is now mature enough to stop.

The next question is empirical.

Before the next recipe is cooked, we will freeze:

the exact recipe version;

the prediction model/version;

the predicted MP6-SIM;

the comparison rules.

Then we will cook.

Then we will record MP6-OBS.

Then we will compare only what is legitimately comparable.

And if the prediction is wrong?

That is not failure of the research programme.

That is the data the programme is designed to preserve.

PUBLICATION

The technical definition and methodology of Predictive Cuisine are now frozen

in:

“Predictive Cuisine: A Closed-Loop Architecture Linking Structured Recipes,

Sensory Simulation and Observed Multisensory Outcomes”

Ajinomatrix Pre-Publication Intent Paper v0.4.1

Zenodo DOI:

https://doi.org/10.5281/zenodo.22088718

The associated public white paper expands the first Elixir case and explains

the workflow in accessible form:

Check out our White Paper

CLOSING

Predictive Cuisine is not an attempt to remove intuition from cooking.

It is an attempt to make culinary intuition testable, versionable and

learnable.

The recipe expresses intention.

The simulation expresses expectation.

The tasting expresses reality.

And the difference between expectation and reality becomes the next thing we

can learn from.

Predict → Cook → Observe → Compare → Learn.