What Will It Taste Like? Ajinomatrix Opens a Recipe-to-Perception Workflow - Then Lets Reality Grade the AI

A recipe - including one recovered from a video - can now become a structured process file, a human-reviewed sensory prediction, and a context brief for AI reasoning. After cooking, the recorded tasting can be compared with the frozen prediction under a published scientific protocol that preserves the AI's misses as carefully as its hits.

Ajinomatrix today announced that its public tools now connect end-to-end into a single recipe-to-perception workflow: from a recipe captured from a page or video to a machine-readable process file (JRF), to a predicted multisensory profile (MP6) reviewed by a human before it is trusted, to a contextual handoff (FEELLLM) that an AI assistant can reason over - and, after cooking, to an observed sensory record that can be compared directly with the original prediction.

The central design decision is deliberately unusual for an AI announcement: the prediction is frozen before anyone tastes anything - and then reality grades it. A wrong prediction is not quietly deleted, overwritten or retrospectively adjusted. It remains part of the evidence, together with the record of where and how it differed from the observed experience.

Try the workflow in five steps
1. Capture the recipe.
At https://www.jrf.recipes, a recipe can be turned into a structured .jrf process file containing ingredients, operations, temperatures, timings and intermediate states rather than only a conventional recipe card.
2. Preview the expected sensory experience.
At https://record.mp6.app, the JRF can generate an Expected MP6: a multisensory prediction spanning sight, smell, taste, touch, sound and intuition. The user can use the fully published deterministic baseline or the currently supported BYOK AI predictor. Each predicted value remains a proposition until the user accepts, edits, rejects or skips it.
3. Ask an AI without losing the evidence boundary.
Through https://www.feelllm.org, recipe state, predicted or observed perception, user context and the actual decision question can be packaged into a brief for tools such as ChatGPT, Claude, Gemini, Grok or other assistants - without silently turning a simulation into a measurement.
4. Cook and record what actually happened.
After preparation, MP6 Recorder can capture the real experience through spoken impressions, typed notes or direct values and produce an Observed MP6.
5. Let reality grade the prediction.
At https://compare.mp6.app, the predicted and observed files can be compared attribute by attribute, only where comparison is legitimate. Missing, skipped or rejected dimensions are not silently converted into zeroes.

The demonstration case used during development is a fermented citrus-ginger tonic recovered from a Facebook video in August - the Elixir of Invincibility. The name is retained strictly as the recipe identity. It has already travelled the full chain: extraction, JRF structuring, sensory prediction, human review, native MP6 export and direct comparison inside the public MP6 Studio toolchain.

For kitchen labs and kitchens alike
For food R&D teams and kitchen labs, the workflow provides a low-friction way to test, on their own recipes and with their own tasters, whether sensory prediction can reduce blind bench iteration - while preserving an auditable record of every prediction, review, correction and tasting.

For households and cooking enthusiasts, it addresses a simpler question: what will this probably be like if I change it? - and then lets the actual dish show whether the preview was right.

For educators, it offers a structured way to record and compare perception while preserving the distinction between self-report, model prediction and measured or panel-derived evidence.

The science underneath: published methods, explicit limits
The workflow is documented in the scientific paper:
"From Recipe Semantics to Human-Verified Sensory Prediction"
Zenodo DOI: https://doi.org/10.5281/zenodo.22869296

The paper publishes the deterministic prediction baseline, the human-review evidence model, freezing and provenance rules, comparison mathematics and a prospectively specified, Methods-frozen protocol under which prediction quality can subsequently be tested.

Ajinomatrix is explicit about what is and is not claimed.
The system is engineering-validated: an earlier development campaign comprised 103 automated checks, followed by a focused 63-check production release gate covering acceptance behavior, hardening and Studio interoperability. Recorder-generated MP6 files have also been demonstrated to move directly into the existing Compare application without a Validator repair pass.

These engineering results do not establish sensory prediction accuracy, panel equivalence or panel replacement.

The scientific proposition is narrower and stronger: the public apparatus now exists to freeze a sensory prediction before observation, preserve its provenance, record what actually happened, and quantify the difference prospectively.
As the paper states in substance: a rejected machine proposition remains evidence about the machine; it does not become evidence about the food.
Open by design, model-neutral by design

JRF and MP6 are designed as portable interchange formats. The deterministic baseline is fully disclosed, and the evidence architecture is not tied to a single AI provider.

Ajinomatrix does not need every food company to use the same AI. The larger objective is to make predictions portable, falsifiable and comparable in the same sensory language.

Food-tech platforms, ingredient companies, sensory-software developers, universities and research groups are therefore invited to read, emit or consume JRF and MP6 files.

An interchange layer becomes more useful as more independent tools can speak it.

Next: "Predict Before You Cook" - a public benchmark
Ajinomatrix is preparing a public Predict Before You Cook benchmark built on the same evidence contract:
recipe frozen → prediction frozen → preparation → observation → residuals published

The benchmark is intended to be model-neutral. Deterministic baselines, general-purpose AI systems, specialized food models and human experts can all submit sensory predictions against the same frozen recipe and the same later observations.

The objective is not to declare one system "best" in advance. It is to make different prediction methods measurable under the same prospective rules and to retain failures as carefully as successes.

Chefs, universities, food-science students, AI developers, ingredient companies and prospective industrial partners interested in early benchmark participation can contact Ajinomatrix.

Try the workflow:
https://www.jrf.recipes
https://record.mp6.app
https://compare.mp6.app
https://www.feelllm.org
Scientific paper:
https://doi.org/10.5281/zenodo.22869296
Website:
https://ajinomatrix.org

Ajinomatrix SRL
Drève de la Chapelle, 34
1430 Rebecq, Belgium
Press contact: François G. Wayenberg, Founder & CEO
via https://www.ajinomatrix.org

Ajinomatrix, based in Rebecq, Belgium, develops technologies for the digitisation, modelling and interoperability of sensory and food-related information.
Its current public stack includes:
• JRF / Ajinoverse for machine-readable recipe and process semantics;
• MP6 for portable multisensory representation;
• MP6 Recorder for predicted and observed sensory evidence;
• MP6 Viewer, Validator, Player, Compare and Report for downstream interoperability;
• FEELLLM for contextual AI handoff;
• SEG / ABED for provenance-aware sensory evidence and feedback observability.
The broader research programme is developed within Ajinomatrix and in collaboration with external scientific and technical contributors.

This release was published on openPR.