Better has to
be measurable.
Connect outcomes to an algorithm. Combine useful signals. Test the next version against data from a later period, then take the winning artifact with you.
Build data you can reproduce.
Join measurements and OMNIA signals, create derived columns, inspect quality and lineage, and freeze a version with its complete research recipe.
Open dataset workspace →What are you measuring?
A machine, an experiment, a recording, or something entirely new. Define your own signals, units, outcomes, conditions, and supporting files.
Start an experiment
Try the farm-equipment example, load your own experiment, or bring a report from your custom algorithm runner.
The example uses synthetic sensor readings. It demonstrates fuel-consumption prediction, not proven fuel savings.
Baseline and candidate signals
Enter the coefficients of your current linear model. Select signals for OMNIA to fit on training data. Custom algorithms use the runner below.
Linked weights must predate the training cutoff. Final test cases keep equal weight.
Candidate input signals
Creates four fitted candidates with different regularization strengths. Only training data influences fitting.
Your code. The same test.
Use a customer-side adapter for classification, custom models, or an API client. Return predictions and portable artifacts, then import the report here for metric verification.
Imported execution is customer-reported. NOUS recomputes metrics; it cannot certify how outside code was trained.
Integration format
Download the example setup for the experiment schema. Examples include id, lineageId, timestamp, inputs and an observed outcome. Keep related episodes within a single partition.
Run sdk/algorithm-lab-runner.js with your setup and a local adapter exporting fit(config, training) and predict(artifact, inputs). Full instructions are in docs/ALGORITHM_IMPROVEMENT_LAB.md.
Experiment history
No experiments yet.
Put improvement to the test.
The result will show which candidate was selected, how it performed on the final partition, and whether it passed each check.
- Timestamp and episode separation
- Bounded believability weights for training
- Equal weight for every final test case
- Approximate uncertainty across independent episodes
- Error limits and group regression checks
OMNIA EFV indices can be supplied as candidate features. Their value must be measured against a baseline without them. Recorded frequency in Hz remains a distinct physical measurement.
Candidate comparison · selection period
| Candidate | Score |
|---|
Final test groups
| Group | Cases | Baseline loss | Candidate loss |
|---|
From a test to a useful product
A model artifact and its evaluation travel together. Use fresh data for the next decision, validate operating conditions, and connect measured field outcomes before changing equipment behavior.
This lab evaluates offline predictions. It does not deploy machine controls or change production algorithms.
Claim IntelligenceTrace the statement. Inspect the sources. Follow the outcome.
See what connects.
One continuous path from a public statement to reusable knowledge.
Give your data a track record.
Track a claim to see real sources and outcomes connected here. No example findings are mixed into your data.
Owner-reviewed outcomes · Connections show recorded lineage, not causation. Motion is decorative; it does not represent measured EFVT or live transfers.