The evidence · structure-only · zero training

Every check, in one place.

An independent University of Manitoba lab, a human tasting panel, 3,230 compounds at scale, and the whole food chemical space — all predicted from molecular structure alone. Read it your way.

Showing the plain-language version — tap Detailed for the numbers, methods and sources.

← Back to the platform

1 · The instrument check

An independent lab measured it. We predicted it.

University of Manitoba e-tongue

The Richardson Centre measured bitterness on an electronic tongue; Codex predicted the same scores from structure.

R² 0.95

Predicted vs measured across 20 samples (8 bitter standards + 4 food polyphenols × 3 dilutions, triplicate): R² = 0.9509, Pearson r = 0.9751, MAE = 1.1 on the lab's 0–26 scale — above the instrument's own reference calibration (R² 0.94).

Source: Richardson Centre for Food Technology & Research, University of Manitoba — formal report, M. Janzen, March 2026.

The full e-tongue deep-dive →Download the lab's report (PDF)

2 · The human-panel benchmark

Checked against people, not just an instrument.

Published human panel + receptor assay

Compounds with bitterness established by a trained human panel and a bitter-receptor (TAS2R) assay — graded blind by Codex from structure.

10 / 10

5/5off-notes caught

5/5harmless cleared

5/5bitter vs astringent

Ground truth: Hald/Dawid/Hofmann 2018; Walser/Dawid 2024 (TU Munich — human sensory panel + HGT-1 TAS2R cellular assay).

3 · At scale

Thousands of compounds, on a laptop.

3,230 externally-labelled compounds

A large, human-curated taste database — every compound graded off-note vs harmless from structure, in about a minute.

88% balanced accuracy

89.7%off-notes caught

86.6%harmless cleared

0.76MCC

18 msper compound

Source: ChemTastesDB (external, human-curated). N = 3,230; balanced accuracy 88.1%; deterministic, zero training data.

4 · Coverage at scale

We screened the whole FooDB.

57,313 food chemicals (FooDB)

Every compound in the FooDB food-chemical database, read in about twenty minutes — each assigned a taste class and an off-note / safe call from structure alone. A throughput no tasting panel or e-tongue can approach.

57,313 screened

~21 msper compound

8,997flagged bitter

1,698flagged sweet

0training rows

This is the coverage demonstration — one fixed structural rule assigns a call to compounds it has never seen, at scale, with zero training data. The accuracy is established separately, where ground truth exists: ~0.80 on a public taste database (ChemTastesDB) and R² 0.95 against a University of Manitoba lab. FooDB shows the reach; those benches show it is right. It does not "recognise every food" — it predicts a taste call for any structure you hand it, and is correct at the rates we publish.

Watch the whole-space screen — demo →

5 · A worked example

On real canola chemistry.

It named the culprits — and cleared the rest

Given only the structures from a canola/protein panel, Codex flagged the bitter and astringent compounds and correctly called the harmless ones inert.

Compound	Codex call
Sinapine	BITTER
Sinapic acid	BITTER
Kaempferol glycosides	BITTER
Kaempferol (aglycone)	ASTRINGENT
Sucrose · citric acid · glutamate · vanillin · glycine	INERT (no off-note)

Sinapine and sinapic acid are canola's known bitter levers; the engine calls the culprit class and the bitter-vs-astringent modality from structure, and clears the harmless sugars and acids.

Simple view: it found the two compounds that make canola bitter, and gave the sugar a clean pass.

6 · What it does — and doesn't — do

Honest about the edges.

A shortlist before the bench — not a replacement for the panel.

Codex is a hand-verifiable pre-bench layer: it shortlists and de-risks before the sensory panel runs. It calls the culprit class and the bitter/astringent modality and ranks across classes.

By design it does not finely rank potency within a single congeneric series — that's the panel's job. And the honest weak spot: on peptides, specificity drops (it over-calls bitter), so peptide panels need the wet assay alongside. We say this out loud; it's why the rest holds.