LESS GUESSWORK. A MORE THOUGHTFUL SERVICE.

Enough for lunch.
Room for uncertainty.

Turn service history into a preparation decision.
See the tradeoff before you turn on the stove.

One kitchen.
One menu category.
One decision you control.

NEXT SERVICE

A plan, with its assumptions.

HUMAN IN CONTROL

Service date Known special event Flag a planned event before service.

SCENARIO-BASED PREPARATION

—portions

Historical residual band—

For recorded requests, not guaranteed demand.
Nominal 80%; actual held-out coverage is below.

YOUR TRADEOFF

What does a shortfall cost?

Relative penalty per unavailable portion versus one leftover portion. These are preference units, not money.

Shortage penalty 3×

More weight on leftoversMore on availability

—average surplus

—average shortfall

UNCERTAINTY YOU CAN SEE

Past errors, tomorrow's scenarios.

Historical-error scenariosYour preparation

Export the inputs, frozen model, evidence and chosen plan.
Reproducible JSON. No orders are placed.

SHOW THE LOSSES, TOO

Does learning help here?

FIXED CHRONOLOGICAL TEST

A ridge model learns calendar and event patterns locally. Two simple alternatives keep it honest. Selection happens only inside training data, never on these held-out results.

60%Train + select

20%Calibrate errors

20%Untouched holdout

Known-label holdout performance · lower error and loss are better
Candidate	MAE ↓	RMSE ↓	Band coverage	Loss 1:3 ↓

Independent stress tests: where this approach fails

A separate evaluator froze six fictional worlds before seeing engine outputs and reproduced results with an independent NumPy implementation. Seasonal/event ridge MAE was 4.46 versus 9.98 for the best baseline. But abrupt-shift coverage was 0 of 18 for every model; unseen-event coverage was only 11–12 of 18. In the no-signal world, the training-selected recent average lost to weekday median. These are synthetic checks, not kitchen validation.

A later adversarial audit found and fixed an integer tie edge case. The corrected engine passed 5,509 independent assertions; those reruns are post-inspection regression, not fresh blind validation. Read the full independent report.

Reproduce the learning and selection

Fixed ridge penalty λ = 5; intercept, weekday indicators, calendar-day trend standardized on training rows, and pre-known event flag. No same-day outcomes, prepared counts, inferred requests, or future normalization enter the features. Weekday median and recent-seven mean use only known training labels.

An expanding-window test on the last up-to-14 known training rows chooses the candidate by MAE. Ties favor weekday median, then recent mean, then ridge. All candidates fit once on the first block; they are not refit on calibration or holdout.

Every held-out service, including unknown demand


Date	Recorded requests	Prediction	Band	Prep	Absolute error

NO INVENTED LABELS

What the kitchen actually knows.

Selling every prepared meal gives a lower bound unless all requests were counted. Unknown demand is excluded from fitting and exact scoring, but stays visible here.