7. Read the result and keep a useful experiment record

Your goal: identify which recipe won, inspect held-out predictions, and save enough evidence to explain the result later.

A result should answer four questions: What data? Which recipe? Which evaluation? Which model can I reuse? Start with one completed run before combining several experiments.

Inspect the first native experiment

This continues 1. Run your first complete experiment. Reload the CPU export in Python/R/Octave, or the browser export in WASM. Read its effective plan and learned final nodes.

JSON

JSON/YAML are recipe and data formats, not result-query languages. Inspect the outcome with one of the language APIs in this box; keep the original exported package intact.

YAML

JSON/YAML are recipe and data formats, not result-query languages. Inspect the outcome with one of the language APIs in this box; keep the original exported package intact.

Python

from nirs4all_core import NativePipeline
model = NativePipeline.load("ridge.native.json")
plan = model.outcome["effective_plan"]
print("Candidates:", len(plan["variants"]))
print("Folds:", len(plan["fold_set"]["folds"]))
print("Learned final nodes:", len(model.outcome["execution_bundle"]["refit_artifacts"]))

R

library(nirs4all)
model <- nirs4all_pipeline_load("ridge.native.json")
plan <- model$outcome$effective_plan
print(length(plan$variants))
print(length(plan$fold_set$folds))
print(length(model$outcome$execution_bundle$refit_artifacts))

Octave / MATLAB

model = nirs4all.NativePipeline.load('ridge.native.json');
plan = model.outcome.effective_plan;
disp(numel(plan.variants));
disp(numel(plan.fold_set.folds));
disp(model.outcome.execution_bundle.refit_artifacts);

WASM / JavaScript

import {loadBrowserPipeline} from 'nirs4all';
const model = await loadBrowserPipeline(localStorage.getItem('ridge-model'));
const plan = model.outcome.effective_plan;
console.log('Candidates:', plan.variants.length);
console.log('Folds:', plan.fold_set.folds.length);
console.log('Learned final nodes:', model.outcome.execution_bundle.refit_artifacts.length);

Expected result: two candidate variants and two learned refit nodes, the scaler and Ridge. Inspect the effective fold count from the actual outcome rather than guessing it from the number of candidates. A candidate, a fold and an observation are different counts.

Interpret scores before ranking models

Result field in Python SDK	What it tells you	Check
`execution_engine`	Which runtime executed the recipe	Matches the intended example
`cv_best`	Candidate selected using CV evidence	Recipe parameters and ranking metric
`cv_best_score`	Its selection score	Metric direction and CV aggregation
`final`	Final refit prediction entry, where present	Distinct from the temporary fold fits
`final_score`	Refit test score, when an external test cohort exists	The test set was not used to select candidates
`top(...)`	Ranked stored entries	Ranking partition and metric
`num_predictions`	Number of stored prediction entries	Not the number of independent specimens

For the Python SDK exercise in 1. Run your first complete experiment, inspect the result inside its try block before result.close():

JSON