Start with a measured Reflex
Open the Reflex catalog and choose a measured test drive. Bearing, Toothbrush, Magic Wand, Gearbox and Transport cover different classification tasks. The test drive opens a read-only project with a frozen Reflex, replay probes and the evidence from its original campaign. You do not need to upload your data or connect a development kit to inspect it.
Start in Results. Follow one input through the actual comparisons to its terminal class. Change a finite feature value and replay the decision. The browser evaluates the graph; it is not running an nRF54LM20 DK. The displayed board figures come from a separate PPK2 campaign. The full historical training panels are not published as downloadable rows, and the small replay probe set is not an accuracy test.
Read the Reflex identity and evidence badge before comparing values. A demonstration graph can have a different variant from the measured artifact. Where that distinction is recorded, keep it visible: a successful browser replay does not make that graph newly measured.
Create your own project
Go to Projects and create a project for your classification objective. Choose the intended target and budgets. Give the project a name that identifies the task rather than a particular experiment: its sessions will hold the experiments. The target is an engineering intent, not evidence that your new model fits or has been timed on that board.
In Data, import a numeric CSV or use the sample dataset offered by the project workflow. Review the parsed columns and assign a target label, ordered features and any group or time metadata. Rows stay in this browser unless you explicitly authorize a service transfer. Related observations from one person, machine or recording should stay together in a grouped split; a random row split can hide leakage.
Train locally, then inspect the result
In Prepare, choose the task, split policy and numeric contract. The browser baseline is a bounded tabular classifier, not unrestricted AutoML. Start a named session in Train. It fits on training rows, selects on calibration rows, freezes the chosen graph and then evaluates the test partition. Local metrics describe those supplied partitions, not a prospective field trial.
| What you see | What it establishes |
|---|---|
| Local split Accuracy and confusion | Performance on this session's local split. |
| Verified Exact replay or receipt | The stated implementation or numeric property, within its recorded scope. |
| Not measured Target cost | No board measurement has been attached to this new Reflex. |
Use Results to inspect confusion cells, class support and literal decision paths together. A wrong label and a wrong branch are different problems. Repeated tuning against the same test set can make that test optimistic; keep a genuinely untouched holdout when making a release decision.
Separate integration from measurement
Deploy offers integration examples and evidence exports. A new compiled C package needs the private compiler service, which is being connected to the v3 workbench. A locked action should say Needs service rather than generate a pretend result. Request access when you are ready to compile; inspect the reviewed payload before consenting.
For device work, read Deploy on nRF54, replay the package’s probes on the actual DK, and then measure the same firmware in Measure. Published test-drive latency and charge cannot be copied onto your own session. Export the project metadata and keep the original CSV separately: the metadata export does not contain the dataset rows stored in IndexedDB.
Last reviewed .
Source of truth: published facts in content/public.ts; CURRENT_MEASURED_RESULTS.md; CONNECTED_BUILD_WORKFLOW.md; standalone C emitter and v3 endpoint contract. Reference examples are not physical-board test receipts.