Skip to main content
Meallions

Validation · The accuracy record

Every prediction is scored against the line.

A language model produces a plausible number. Meallions produces a number that has been checked against a running line — and tells you which of the two you are looking at. This page exists to show the checking.

1–2

Confirmation runs to a passing batch

SIMULATED

< 60s

Batch recall

LINE-VERIFIED

7years

Immutable records

LINE-VERIFIED

4

Parameters scored every run

SOURCED

How a prediction becomes evidence

Four steps, in this order, every run.

  1. 01

    The window is recorded before the run

    The predicted operating window, the setpoint package and the confidence are written down and versioned before any material is committed. It cannot be edited afterwards.

  2. 02

    The run is executed from the approved setpoint package

    An expert approves the package at the production gate. The line runs those values; deviations are logged against the record, not around it.

  3. 03

    Inline and lab measurement are reconciled

    Inline sensors and NIR scans are reconciled with lab assays against the batch specification. Disagreement between the two is itself recorded.

  4. 04

    The delta is stored against that line and recalibrates the model

    Predicted minus actual is stored per parameter, per line. That error tightens the twin for that specific machine — not for a global average of every line in the world.

Published accuracy

Predicted against actual, per parameter.

We publish the error, not just the wins. A vendor that shows you only its successes has not told you its accuracy. The series below are demonstration data in the exact shape a connected line produces; once your line is connected, these charts carry your runs and are marked LINE-VERIFIED.

Demonstration modelLINE-VERIFIED

Die pressure (bar)

Demonstration model

Mean absolute error 0.9 bar (2.1%) · target band shaded · predicted solid, actual dashed

Moisture exit (%)

Demonstration model

Mean absolute error 0.31% absolute · target band shaded · predicted solid, actual dashed

Specific mechanical energy (kWh/kg)

Demonstration model

Mean absolute error 0.006 kWh/kg · target band shaded · predicted solid, actual dashed

Throughput (kg/h)

Demonstration model

Mean absolute error 9 kg/h (1.8%) · target band shaded · predicted solid, actual dashed

Calibration

The model learns a specific machine.

Prediction error against verified runs on one line (% MAE, die pressure)

Demonstration model

Error decays as the model is calibrated on runs from that specific machine. The shape of this curve is the product: a global model is the starting point, the line-specific calibration is what you buy.

Operating-window hit rate

86%

Share of first confirmation runs landing inside the predicted window, after eight or more verified runs on that line.

Demonstration model

Below eight verified runs the window is wider and the hit rate is lower. We state the run count next to the figure rather than quoting the best case.

Stated limits

What we are not good at yet.

A stated limitation is worth more than a claimed strength. In these cases the model declines and says so, rather than producing a number you cannot check.

High-fat and high-sugar confectionery extrusion

Melt behaviour is dominated by fat phase transitions we have not verified across enough runs. The twin declines rather than extrapolating.

Single-screw equipment

Our calibration history is twin-screw. On single-screw lines the twin returns SOURCED benchmarks only, and will not emit a setpoint package.

Wet texturized protein above 60% moisture

Cooling-die dynamics need line-specific runs before confidence rises. Expect a wider window and more confirmation runs.

Novel ingredients with no verified specification

Without a supplier spec or a verified reference, the input is not admitted at stage 01. That is a refusal, not a gap we paper over.

How the record is governed →Why this is not an LLM →