A binary format you can read.

This is the workbench without its interactive layer. Every byte, table, figure and note the app carries is below, in file order — what is missing is the ability to select a field and see its bytes light up.

The capture

A binary format you can read

CSIQ is what this project's nine radios write to disk. Its design goal is that a capture should be interpretable by somebody holding only the file and the specification. This page is that claim, tested: a real capture out of the archive, decoded in front of you.

Every number here came out of the file beside it. Nothing is written into the copy. The page decodes the committed capture with the same reader the specification documents, so a figure cannot drift from the bytes you can download.

This capture is the archive's session monad09_walk-abba-48-night_20260827-161301 — monad09, walk-abba-48-night, 5 GHz channel 48 at HT20, written by csid 0.2.0. Its own numbers disagree with each other, so it is excluded from every corpus total. It is in the catalogue deliberately, so a reader can see what one looks like.

The conformance kit

Every sample here is downloadable as a valid, self-contained capture, with the decode this project's reader produces for its first records as JSON. Point your own reader at the file and diff against the answer key. Where the two disagree, the specification wins and both readers are suspect — that is the document's own rule, and it is why there is a published fixture at all.

Defects the format admits

• SEQ is a driver record counter. Not the 802.11 sequence number its name suggests. Corrected 2026-08-24; the name is kept because changing it would break every existing reader.
• Two reference readers rendered an absent width differently. One as an unknown code, one as its own default. Two sentinels for one fact, found only when both were decoded side by side. Neither is right: absence should be reported as absence.
• The header version cannot say which fields a file carries. Adding a type code is not a version bump, so two captures both stamping version 1 can differ by five codes. The capability probe is the answer instead.

What the catalogue does not hold

The catalogue is declared by axis rather than by session, so a cell the archive does not populate is reported rather than quietly dropped. 7 declared cells are empty:

• band=2.4,condition=quarantined,width=HT20
• band=2.4,condition=clean,width=HT40+
• band=2.4,condition=quarantined,width=HT40+
• band=2.4,condition=clean,width=80MHz
• band=2.4,condition=quarantined,width=80MHz
• band=5,condition=quarantined,width=HT40+
• band=5,condition=quarantined,width=80MHz

The 2.4 GHz wide-channel cells are empty because this deployment runs 2.4 GHz on channel 11 at HT20 only. That is a fact about the fleet.

File head

Offset 0, 12 bytes.

Twelve bytes, then everything else

Four bytes of magic, a version, a flags word, and the length of the block that follows. A reader that does not recognise the magic must stop here rather than guess: everything after these twelve bytes is only meaningful if they said what they were supposed to say.

The header version cannot say which fields a file carries. Adding a type code is not a version bump, so two captures both stamping version 1 can differ by five codes. The capability probe answers that instead.

Session block

Offset 12, 3,384 bytes.

Provenance travels with the data

One JSON object at the head of every capture, carrying how the measurement was taken: the radio's channel and width, the build of the daemon that wrote it, the host it ran on, and the filter that decided which frames were kept.

The block is opaque by design. A writer adds a group without a format version bump, so a reader that dropped what it did not recognise would silently lose the newest thing anybody recorded. Every group this capture carries is listed, known or not.

Record 0

Offset 3,396, 819 bytes.

A self-describing record

A tag byte, a four-byte length, then a chain of type-length-value fields. Each field costs five bytes of header, and that is the price of the property below it: a reader skips any type code it does not know, so a writer can add one without breaking every existing reader. Adding a type code is therefore not a version bump.

A field this file does not carry is a fact about the writer, not a gap in the measurement. It changes which questions can be asked of the capture at all, so each absence below is printed with what it costs.

What this writer does and does not record

Presence was established by probing 2,000 records. Absence is only ever evidence over the records looked at.

FTM

Why one timestamp is not enough

A record carries up to five timestamps, and no one of them answers every question. Each series is drawn as seconds since its own first reading, because they have different epochs and one absolute axis would say nothing.

The FTM counter is 32 bits at 320 MHz, so it wraps. Differencing the stored form across a wrap gives a large negative interval that looks like a clock stepping backwards.

The FTM counter is 32 bits at 320 MHz, so it wraps every 13.42 s — 0 times in the 400 records above.

BW_ANTSEL

242 tones is HE20 or VHT80

A record says how many tones it carries. That number alone does not name the numerology: 242 tones is HE20 at 78.125 kHz spacing, or VHT80 at 312.5 kHz. Deriving the grid needs the bandwidth too.

The bandwidth is the field the reader rule exists for — prefer the explicit code, fall back to decoding the rate-and-flags word, and report genuine absence as absence rather than as 20 MHz.

What the catalogue's records carry

SEQ

A name the format is stuck with

SEQ is a driver record counter, not the 802.11 sequence number its name suggests. Corrected 2026-08-24. The name is kept because changing it would break every existing reader.

SRC_MAC

The one thing that is not the archive's

Record framing and lengths are bit-identical to the original capture. The only bytes that changed are the six-byte MAC addresses, relabelled into the RFC 7042 documentation range — here, in the session block, and inside the 272-byte driver header that keeps its own copy of the source address.

Hostnames are not redacted: this site publishes node names already, and stripping them would remove the provenance these files exist to demonstrate.

RSSI

A sentinel is not a weak measurement

Received power is stored per chain in dBm. One value is not a measurement: −127 marks a stale block — the driver handed back the previous record's bytes unchanged.

Averaged in, it becomes the weakest signal the fleet ever recorded, tens of decibels below anything real, and it drags every mean it touches. It is counted separately everywhere it appears on this page.

CSI_MATRIX

The CSI matrix, and the two ways to read it wrong

Complex coefficients across the tone grid, one block per antenna chain — the shape the channel imposed on a frame as it crossed the room.

Amplitude here is AGC-normalised, so it is channel shape only. The correlation between the squared amplitude and received power is about 0.01 on this hardware. The absolute scale is in the RSSI field, not in this one.

Two properties of the stored matrix are easy to get wrong, and getting either wrong is silent: the result looks healthy and is not. Both are visible in the impulse response — the same data seen in delay rather than in frequency — and in nothing else.

1. Coefficients are imaginary first, then real. Reading them the other way round yields i·conj(H). Amplitude is untouched, so every amplitude plot stays byte-identical, and every phase is mirrored.
2. The payload is chain-major. Chains are contiguous blocks of tones, not interleaved per tone. Read the other way, the impulse response smears instead of concentrating.

This capture inverts the specification's result, and we cannot yet say why. The specification measures the correct reading at 21.5× more early-delay energy than late across a population of real captures. On this file the ratio runs the other way. Four of the ten catalogue samples do this, and all three of their sessions are walk-* capture profiles. The page prints what it measured rather than what it expected, because a plot that contradicts the number printed beside it teaches a reader to distrust the plot.

The same measurement, across the catalogue

Record 1

Offset 4,215, 819 bytes.

Record 2

Offset 5,034, 819 bytes.