Three models you can redo

The whole cycle with the four verification stages in PASS. This is the case that admits no argument: if the package returns 273.15 for 0 °C it works, and if it returned anything else you would see it without knowing any machine learning. ON YOUR MACHINE the fourth stage will say INCOMPARABLE and the command will exit with code 3, and that is the CORRECT answer: the tolerance was measured in the environment of the package and yours is a different one, so a difference there would prove nothing. The first three do come back PASS.

predicted_kelvin = 1*celsius + 273.15

The recipe is the data, not decoration: the rows are generated from it, it ships inside the package as `data_recipe.txt`, and `verify` uses it to REGENERATE the dataset and compare its digest.

Copy them as they are into an empty directory, next to the recipe. They are all you need: the commands above generate the rest.

PROJECT CelsiusToKelvin

VECTOR Reading[1]
  celsius: Scalar[-50, 150]
END

PARAM W1 Vector[1]
END

PARAM b1 Scalar
END

FUNCTION PredictedKelvinModel
  predicted_kelvin: Scalar = linear(W1 * Reading + b1)
END

GRAPH
  Reading -> PredictedKelvinModel
END

MODEL kelvin.mxai

DATASET KelvinData
  SOURCE csv("datos/celsiustokelvin-synthetic-train.csv")
  INPUT Reading FROM COLUMNS [celsius]
  TARGET predicted_kelvin: Scalar[200, 450]
  SPLIT train=0.8 validation=0.2 seed=42
  BATCH size=8
END

LOSS KelvinLoss
  TYPE mse
  PREDICTION predicted_kelvin
  TARGET predicted_kelvin
END

OPTIMIZER KelvinOptimizer
  TYPE sgd
  LEARNING_RATE 0.5
  UPDATE W1, b1
END

RUN
  EPOCHS 200
END

• paquete.zip — THE WHOLE PACKAGE: model.onnx + predict.py + inference_spec.json + reproduce.json. It predicts with `pip install onnxruntime` and nothing else
• kelvin.mxai — the model
• kelvin.mxtrain — the training contract
• receta.txt — the data recipe
• entrada.json — the example input (0 °C); the other two are entrada_100.json and entrada_-40.json
• reproduce.json — the package manifest, to compare digests
• salida.txt — the literal output of the five steps

matrixai generate-dataset kelvin.mxai --training kelvin.mxtrain \
  --rows 300 --seed 20260825 --mode coherent --recipe receta.txt -o datos
matrixai train kelvin.mxai --training kelvin.mxtrain --output runs/v1 \
  --recipe receta.txt --dataset-manifest datos/celsiustokelvin-synthetic-manifest.json
matrixai export-bundle kelvin.mxai --params runs/v1/params.best.json \
  --outdir paquete --training kelvin.mxtrain --data-recipe receta.txt --from-run runs/v1
cd paquete && python3 predict.py --input ../entrada.json
cd .. && matrixai verify paquete.zip --retrain

R2: 1.000000
MAE: 6.50521e-17
Equivalence PASS: max_abs_diff=3.57e-08

  0 °C   -> 273.15000146627426
  100 °C -> 373.1500029563904
  -40 °C -> 233.15000236034393

manifest  PASS
R1        PASS
training  PASS
R3        PASS

Everything the previous one shows, plus the document a journal asks for to publish a clinical prediction model — written from what the run recorded, and saying which boxes it CANNOT fill.

alto: edad > 75 OR reingreso_previo > 0.5
DEFAULT: bajo

The data is SYNTHETIC and the record says so in its first line. A model trained this way is not clinically validated, and publishing the case without that sentence would be the problem, not the solution.

Copy them as they are into an empty directory, next to the recipe. They are all you need: the commands above generate the rest.

PROJECT ClassifierProject

VECTOR Input[4]
  edad: Scalar[18, 100]
  dias_ingresado: Scalar[0, 60]
  num_diagnosticos: Scalar[0, 15]
  reingreso_previo: Scalar[0, 1]
END

NETWORK Classifier
  INPUT Input
  LAYER Dense units=32 activation=relu
  LAYER Dense units=16 activation=relu
  LAYER Dense units=2 activation=softmax
  OUTPUT predicted_class: ProbabilityMap[alto, bajo]
END

GRAPH
  Input -> Classifier
END

AUDIT
  EXPLAIN Input -> Classifier
END

MODEL modelo.mxai

DATASET Reingreso
  SOURCE csv("datos/classifierproject-synthetic-train.csv")
  INPUT Input FROM COLUMNS [edad, dias_ingresado, num_diagnosticos, reingreso_previo]
  TARGET predicted_class: Label[alto, bajo]
  SPLIT train=0.8 validation=0.2 seed=42
  BATCH size=16
END

LOSS Perdida
  TYPE cross_entropy
  PREDICTION Classifier
  TARGET predicted_class
END

OPTIMIZER Optimizador
  TYPE sgd
  LEARNING_RATE 0.05
  UPDATE Classifier.*
END

RUN
  EPOCHS 60
END

• paquete.zip — THE WHOLE PACKAGE: model.onnx + predict.py + inference_spec.json + reproduce.json. It predicts with `pip install onnxruntime` and nothing else
• modelo.mxai — the model
• modelo.mxtrain — the training contract
• receta.txt — the data recipe
• entrada.json — the example input (age 92); the other one is entrada_joven.json
• ficha_tripod.en.md — the FULL TRIPOD+AI record in English, produced by the command (ficha_tripod.md is the Spanish one)
• reproduce.json — the package manifest
• salida.txt — the literal output of the six steps

matrixai generate-dataset modelo.mxai --training modelo.mxtrain \
  --rows 400 --seed 20260825 --mode coherent --recipe receta.txt -o datos
matrixai train modelo.mxai --training modelo.mxtrain --output runs/v1 \
  --recipe receta.txt --dataset-manifest datos/classifierproject-synthetic-manifest.json
matrixai export-bundle modelo.mxai --params runs/v1/params.best.json \
  --outdir paquete --training modelo.mxtrain --data-recipe receta.txt --from-run runs/v1
matrixai verify paquete.zip --retrain
matrixai report paquete --tripod --locale en -o ficha_tripod.en.md

Accuracy: 0.968750

  age 92, prior readmission 0.9 -> {"alto": 1.0,    "bajo": 9.9e-11}
  age 25, no prior readmission  -> {"alto": 0.0002, "bajo": 0.9998}

# TRIPOD+AI record
> WARNING: the training data is SYNTHETIC…
- Predictors: edad (18–100), dias_ingresado (0–60),
  num_diagnosticos (0–15), reingreso_previo (0–1)
## What this record CANNOT fill in
- Missing data · Calibration · Subgroup fairness
- Funding · Conflicts · Ethical approval

manifest  PASS
R1        PASS
training  PASS
R3        PASS

2,189 days of real weather observations. There is no recipe to regenerate that data, so `verify` leaves two stages INCOMPARABLE and says so in the package’s own words. This is what sets this product apart: what was not checked is stated.

There is no recipe here and none was invented. Inventing one so the four stages would go green would have been exactly the lie these pages claim to avoid.

Copy them as they are into an empty directory, next to the recipe. They are all you need: the commands above generate the rest.

PROJECT LluviaManana

VECTOR Parte[6]
  temp_max: Scalar[0, 45]
  humidity: Scalar[30, 100]
  pressure: Scalar[980, 1040]
  pressure_delta: Scalar[-20, 20]
  cloud_cover: Scalar[0, 100]
  precip_hours: Scalar[0, 24]
END

NETWORK Pronostico
  INPUT Parte
  LAYER Dense units=16 activation=relu
  LAYER Dense units=2 activation=softmax
  OUTPUT llueve_manana: ProbabilityMap[no, si]
END

GRAPH
  Parte -> Pronostico
END

MODEL lluvia.mxai

DATASET Observaciones
  SOURCE csv("datos/lluvia_nombrada.csv")
  INPUT Parte FROM COLUMNS [temp_max, humidity, pressure, pressure_delta, cloud_cover, precip_hours]
  TARGET llueve_manana: Label[no, si]
  SPLIT train=0.8 validation=0.2 seed=42
  BATCH size=32
END

LOSS Perdida
  TYPE cross_entropy
  PREDICTION Pronostico
  TARGET llueve_manana
END

OPTIMIZER Optimizador
  TYPE sgd
  LEARNING_RATE 0.1
  UPDATE Pronostico.*
END

RUN
  EPOCHS 40
END

• paquete.zip — THE WHOLE PACKAGE: model.onnx + predict.py + inference_spec.json + reproduce.json. It predicts with `pip install onnxruntime` and nothing else
• lluvia.mxai — the model
• lluvia.mxtrain — the training contract
• lluvia_nombrada.csv — the 2,189 REAL days, with the classes named
• dia.json — the humid example day; the dry one is dia_seco.json
• reproduce.json — the package manifest
• salida.txt — the literal output of the five steps

Daily observations from Open-Meteo (open-meteo.com), ERA5 reanalysis. Its API is free for non-commercial use and its data derives from ERA5 (Copernicus/ECMWF), published under CC BY 4.0. The CSV you download here is a selection of columns plus a derived label —«will it rain tomorrow»— computed by us; it is not the original data.

matrixai train lluvia.mxai --training lluvia.mxtrain --output runs/v1
matrixai export-bundle lluvia.mxai --params runs/v1/params.best.json \
  --outdir paquete --training lluvia.mxtrain --from-run runs/v1
cd paquete && python3 predict.py --input ../dia.json
matrixai verify paquete.zip --retrain

Accuracy: 0.762557   (2,189 real days)

  humid day, pressure falling -> {"si": 0.943, "no": 0.057}
  dry day, high pressure      -> {"si": 0.260, "no": 0.740}

manifest  PASS
R1        INCOMPARABLE — «this model has no data recipe, so its dataset
          cannot be regenerated and compared; why there is none is not
          recorded in the package; the dataset generation seed is unknown»
training  INCOMPARABLE — the dataset could not be regenerated, so there is
          nothing to train on
R3        NOT_RUN

The reason is in English because it is what the PACKAGE wrote about itself, and here it is quoted rather than translated: those bytes are covered by its own digest, and translating them would put words in its mouth.