Browser-local CSI capture quality checker (GL-MT6000 / mt76)

By Wu Jun · Updated 2026-11-02 · 6 min read

Establish whether a capture is analysable before spending time on cleaning, spectral analysis or machine learning. This checker parses only this project's public container: double timestamp + uint16 length + mt76 CSI payload. All CSI processing happens in the current browser tab; the selected capture is not uploaded ^{[1] [2]}.

What it checks

Verdict rules

• Pass: at least 10 closed groups, chain and subcarrier coverage at or above 98%, non-zero I/Q, and no out-of-order groups;
• Review: the file parses, but has a truncated tail, missing chains, mixed subcarriers, out-of-order groups or a maximum gap over five seconds;
• Fail: no closed groups, a malformed payload header, no non-zero I/Q, or a file that is not this capture format.

These are fast screening thresholds, not an IEEE conformance test or a guarantee of algorithm accuracy. A formal experiment should still record firmware, driver, channel, bandwidth, exact command, free disk space and daemon memory. For the full offline statistics, run the Python quality audit.

What to do with the result

Do not delete a file simply because it says “review”, and do not send a “pass” straight into model training. Trace the result back through the capture chain:

This table only answers whether the file and capture chain are safe to use in the next analysis step. It does not validate a classifier or a sensing task. When reporting an issue, include the exported JSON/Markdown report, relevant logs and a minimal reproduction. Raw captures can still contain unredacted MAC addresses and absolute time, so apply the privacy workflow before sharing them.

Important boundary: a gap is not automatically packet loss

The capture file proves only that no closed measurement group was written during that interval. The cause could be packet loss, a receiver falling behind on disk, no CSI trigger, a frame-filter mismatch, or a stopped capture daemon. Without daemon logs, interface counters and receiver-side drop counters, a single timestamp gap cannot be assigned to one cause.

An unterminated tail group is also commonly just the point where capture stopped. It should be quarantined, not silently zero-filled. This is the same treatment used by the published Python audit, so the browser quick check and offline recomputation can be compared directly.

Data format and reproduction

The page reads the analyser-side container:

repeated:
  float64 host_timestamp
  uint16  payload_length
  payload_length bytes:
    24-byte little-endian mt76 CSI header
    int16 I/Q pairs

The container is produced by csi_capture.py and csi_capture_linux.sh; the payload header follows the mt76 CSI reporting layout. The toolkit publishes the scripts, data-format contract and a real loopback test ^{[1] [3]}. To validate this page first, select “Load the public 60-second sample”. The expected result is 2,777 records, 347 closed groups, 256 subcarriers, 8 chains per group and one tail record.

You can also reproduce the same result in Python:

• Public 60-second sample
• SHA-256 and manifest
• Tools, scripts and more datasets

Privacy

The current page reads the raw CSI file, but the file is not submitted through a form, fetch or analysis API. The page still loads the site's normal JavaScript, styles, fonts and advertising scripts; those routine requests should not contain the selected capture. Exported reports omit transmitter MAC addresses and retain only the transmitter count. A browser extension, malicious script or compromised device can defeat any local-processing claim, so handle unredacted captures only on a trusted machine.