FICUS , home
Reflex Lab · Docs

Data format and capture

CSV layout, column roles, holdout files and recording from a development kit over Web Serial.

On this page

Use an explicit CSV contract

Use a header row with stable column names and one observation per row. Classification features must be numeric and finite for the browser baseline. Keep an authoritative label column separate from features. Timestamps, subject IDs, device IDs and recording IDs are usually metadata, not predictors. Assign their roles explicitly in Data; a column’s name alone does not establish its role.

Illustrative training CSV — values are not benchmark data
t_ms,recording_id,accel_x_rms,accel_y_rms,label
0,run_a,0.12,0.19,idle
20,run_a,0.15,0.22,idle
0,run_b,0.81,0.43,moving

Preview parsed rows before training. Check delimiters, decimal points, missing values, duplicate names, nonfinite values and class support. Do not fill a sensor failure with a zero unless that policy was deliberately trained and validated. A flat channel can be a real steady state, a wiring fault or a parsing error; investigate its origin instead of assuming the profiler knows which.

Freeze feature order, units and partitions

The deployed array must use exactly the feature order in the matching session manifest. accel_x_rms at one index cannot silently become a different axis or unit. Float32 rounding is part of the input contract. If the firmware computes features from raw signals, match the sampling, window length, stride, scaling and boundary handling used in preparation.

RoleUse
FeatureFinite input value at a specific manifest index.
TargetAuthoritative label used to evaluate the decision.
GroupKeeps related observations in one partition.
TimeOrders events or defines a temporal split; not automatically a feature.
IgnoreMetadata not consumed by the classifier.

Fit only on training rows; select settings on calibration rows; evaluate the frozen choice on test rows. Keep duplicate feature rows and related recording/subject groups from crossing partitions. An optional holdout CSV needs the same ordered feature schema and class semantics, but should contain independent observations that were not used for selection. Repeated experimentation still reuses your supplied data.

Record a DK over Web Serial

Download the capture firmware reference sample, build it with an nRF Connect SDK version supporting your nRF54LM20 DK revision, and inspect its README before flashing. The default without a ready sensor is explicitly synthetic. A real sensor uses Zephyr’s generic sensor API and a devicetree alias; the sample does not assume that a DK includes an IMU.

Exact UART protocol — illustrative signal values
# ficus-capture v1 channels=synthetic_signal rate_hz=50
0,-1.000000
20,-0.960000
40,-0.920000

Open the project’s Data → Capture panel in a Web Serial-capable browser. Use a secure context or localhost and select the DK’s virtual COM port with permission. The firmware stream is 115200 baud. The optional header names channels in value order; each subsequent line is t_ms,v0,v1,.... Timestamp units are milliseconds since boot. There is no label column in the device stream: apply recording labels deliberately before saving a training dataset.

Check the recording before using it

Close terminal programs holding the port. If no rows arrive, check UART selection, firmware, baud rate and permissions. If rows have gaps, inspect sensor errors and actual timestamp differences. The requested sample rate is not a guarantee that every deadline was met. Keep any synthetic capture clearly named and excluded from real-data claims.

Raw capture is not the processed-feature input of an existing Reflex. Prepare must establish the transformation before Train can use it. The capture protocol is also distinct from the Field monitor’s decision log; sending sensor values to a decision-log parser will not create predictions.

Rows are stored in the browser unless you explicitly send them for a service session. Consent should bind to the actual payload, not a blanket promise to upload later. Save the original recording outside browser storage if it matters: project metadata alone cannot recover a deleted capture.

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.

On this page