Finish a tabular experiment.

Make the task specific before fitting

The project asks a modest question: how well can sepal length predict petal length, both in centimetres, for this educational Iris comparison? A sepal is an outer flower structure; a petal is a different flower structure. The task predicts a measurement, not a species label. Species is used only to balance partitions and inspect error slices.

This distinction matters. A classifier’s percentage correct cannot be compared directly with a regression error in centimetres. Write the input, target, available measurement time and intended use before choosing a model. For a real collection, also ask whether measuring the target directly would be more useful than estimating it.

Read the data contract

The local file contains 150 rows from the UCI Iris dataset, licensed CC BY 4.0. The workbench links the source, license and exact local CSV hash. The source measurements are unchanged. Generated row IDs are not verified specimen identities.

Identical original four-measurement vectors are grouped and deduplicated, leaving 149 measurement groups. Within each species, first-occurrence group order assigns index modulo five: remainder 0 to evaluation, remainder 1 to development, and remainders 2–4 to training. “Modulo five” means the remainder after dividing by five. This gives 89 training, 30 development and 30 evaluation groups.

Grouping uses all four original measurements to avoid exact duplicate vectors across partitions. The fitted models receive only sepal length as their input. This historical, deterministic protocol is not a random sample of future flowers, and measurement fingerprints cannot detect every repeated specimen. No synthetic missingness or unit damage is applied in this project.

Iris has appeared in earlier lessons. Reserving rows inside this workflow does not make the public dataset previously unseen. The experiment teaches a frozen comparison, not authenticated independent performance.

Give the simple reference a fair chance

A mean baseline predicts the training average petal length for every input. If training targets were 1, 3 and 5 cm, that prediction would be 3 cm. It provides a reference for whether input-dependent predictions add value.

The line fits an intercept and slope from the same training groups. The ensemble repeatedly fits trees to training residuals and adds their scaled predictions. A residual is the observed target minus the current prediction. The earlier regression lesson and boosting lesson explain those calculations step by step.

Compare using the same task, groups and units. A line may beat a more flexible ensemble on development data. Complexity also affects explanation, implementation and maintenance. Choosing the simpler model is a valid conclusion when the evidence supports it.

The workbench retains twelve configurations at most. Decide an allowance for model choices before examining many results. Settings are recorded even when repeated; repetition is not fresh evidence. The controls that configure a new candidate are temporary, while recorded configurations and reports are saved locally.

Read both error summaries

For observed targets 2 and 6 cm and predictions 3 and 4 cm, the absolute errors are 1 and 2 cm. Mean absolute error (MAE) is their average: (1 + 2) / 2 = 1.5 cm.

Mean squared error (MSE) squares each error before averaging: (1² + 2²) / 2 = 2.5 cm². Squared units matter. MSE gives a large individual error more influence than MAE does; neither summary identifies every important failure.

Nearby practice: absolute errors 0, 0 and 3 cm give MAE 1 cm and MSE 3 cm². Compare that with three errors of 1 cm: the same MAE, but MSE only 1 cm². If a single large error would matter in the intended use, inspect individual errors or a predeclared worst-error criterion as well.

Freeze the recipe before looking

Choose a recorded candidate and explain the development evidence supporting it. Freezing records its model configuration and training/development identities. Evaluation replays fitting from the original local source; imported scores do not become evidence just because a file says they are correct.

The project does not silently add development rows to training. Its frozen result evaluates the exact selected fitting recipe on the reserved partition. A different protocol could refit after selection, but would need to declare that change before evaluating it.

Freezing alone does not reveal results. Use the separate evaluation control when ready. Reopening development retains known evaluation exposure. If evaluation errors influence subsequent choices, report that use rather than calling the same rows untouched again. Local exposure records and fixed cases are inspectable and self-administered, not a secure examination system.

Explain an error without inventing its cause

Open the table of evaluation predictions. Select a row and state its input, observed target, prediction and absolute error. Then compare the species slices, each with ten groups in this protocol.

A slice difference can suggest an investigation: perhaps one input does not capture enough variation, or the training range is unrepresentative. It does not prove that explanation. You would need additional measurements, a carefully controlled comparison or a new collection to distinguish possibilities. A small slice should not support a broad reliability claim.

Turn the record into useful evidence

The guided variant asks for a reproducible experiment and explanation. The implementation variant additionally asks for runnable source, environment/run instructions, tests and comparison with browser results. The decision variant asks whether the measured behavior meets the stated purpose, which errors matter, what operating assumptions remain and whether to proceed, defer or avoid the model.

For implementation practice, download the Python comparison starter. It runs a mean baseline immediately and leaves line fitting as your task. Its visible checks cover known arithmetic, constant inputs, invalid values and unintended mutation. Add your own split-isolation and reproduction tests. The starter freezes a line explicitly; extending it to an ensemble is optional further implementation work.

All variants share the same data and workbench. Any learner can choose another depth. The site stores implementation code as inert text; it never executes imported code. Minimum text lengths and matched configurations check that a report has evidence to review, not that the reasoning or implementation is correct.

Use the separate delivery-depot and concentrated-error cases to check transfer. Their arithmetic and situations differ from the Iris example. First submissions remain distinct from practice retries. Even passing both does not establish independently reviewed software implementation or deployment readiness.

Download the report for inspection and use a My Work backup to preserve editable state. Progress stays on this device unless you export and import it. Take the report to a peer or teacher who can challenge one calculation, one data assumption and one claim. Human artifact review and learner pilots remain pending.