A third section, alongside the aircraft and the weather sensors, and for the
same reason as both: a scan stops on a signal, records it and moves on, while
APRS is a two-second transmission every few minutes from a hundred stations
sharing one frequency. A sweep catches whichever happened to key up as it
passed. `bandsaunter aprs` parks on the channel and catches all of them;
`bandsaunter packets` reads a log back.
Four layers, three of them new.
The link layer was already here, opportunistically, in the generic decoder --
a correlator, NRZI, HDLC and a checksum, run on whatever a scan happened to
record. It is now a receiver. What had to change is the state that survives
a block boundary: the tail of the audio so the correlators see no edge, the
phase of the sampling loop so a bit is not lost where one block meets the
next, the tone the line was last at, and the bits themselves. A packet is
most of a second and a block is about one, so frames straddling the boundary
are not an edge case, they are most of them.
Above that, the APRS information field, which is not one format but about
twenty, chosen by the first character and accreted over thirty years.
Positions uncompressed and compressed; Mic-E, which every Kenwood and Yaesu
mobile sends and which hides the latitude inside the destination callsign
because in 1995 those six bytes were carrying the word "APRS" and nothing
else; weather with a position and without; messages, acknowledgements,
rejections and bulletins; objects and items; status; telemetry; third-party
traffic, credited to whoever originally sent it rather than to the gateway.
Course and speed, altitude, power and antenna height, range and the precision
extension, all of which ride in the comment. Every one has a writer beside
its reader, so a packet goes in and the same packet comes out.
Above that the section: a registry of who is out there and what each last
said of each kind, distances and bearings from --at, a log keeping the whole
frame under whatever was made of it, a spreadsheet, a map, and a channel full
of stations that are not there for --simulate.
One rule is worth naming because it is the difference between a decoder and a
liar. A packet whose format does not match what its first character promised
comes back as unparsed with its text intact. Thirteen characters of a
*malformed* uncompressed position are perfectly good base-91, so trying one
format and falling back to the other does not fail on a bad packet -- it
succeeds, as a confident and completely different place, usually a thousand
miles away. The specification makes the two unambiguous, a leading digit
always meaning uncompressed, and the rule is read rather than guessed at.
Two faults found by building it, both by measurement rather than by reading
the code again. The framer handed back frames it had already reported,
because it trimmed its buffer to before them rather than after -- every packet
counted twice, for ever, which only shows up once the same signal is read
across more than one block. And the invented channel truncated a
transmission at the end of the block it began in rather than carrying the
remainder over, which was invisible for as long as the simulated clock
advanced in exact seconds and put every transmission at a block boundary; the
moment it was paced against a real clock, nothing decoded at all.
204 new tests against seven deliberately broken builds, one of which survived
until a test was written for the case it actually breaks. Full suite 2597
passed. Built as 2026-09-20_02.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016PsWPTweCT6pwxKngvVxcg
Still nothing, with rtl-433 receiving the same sensors on the same aerial from
a different receiver. That settles where the fault is not: not the aerial,
not the sensors, not the band. So the sensible thing is to stop differing
from the configuration known to work on that aerial, and this differed from it
in three ways, every one of them mine.
It tuned a quarter of a megahertz to one side of 433.92 and shifted the signal
back in software, to keep the receiver's own spike off a signal that works by
being switched off. That is a real effect and avoiding it this way is a bad
trade: at the sample rate this ran at, the shift is followed by a filter, and
a filter narrow enough to reject the spike is narrow enough to lose a
transmitter that has drifted -- or to lose the signal outright if a dongle
presents its samples the other way round, which is not a thing to depend on.
The spike is a steady addition to the envelope and a burst rises clear of it.
Tuning straight at the sensors now, which is what the established tools do.
It sampled at a megasample a second where a quarter of one is plenty: the
shortest pulse these send is two hundred microseconds, which is fifty samples
at the lowest rate a dongle will do. The extra rate bought nothing but the
room for that filter to exist in. At 250 kS/s nothing after the mixer is
narrower than the band, so the offset now does nothing at all whatever it is
set to, and says so.
And it turned on the RTL2832's digital gain control along with the tuner's.
The two pump: the gain winds up through the silence between one burst and the
next, lifting the noise towards the signal and squeezing the very difference
the burst detector works on. It matters here in a way it does not for
aircraft, where a frame is found by correlating a preamble over microseconds
rather than by comparing a burst with the quiet around it.
Three more faults found while going over the rest of it, all the same mistake
in different clothes -- treating the middle of a distribution as though it
were the quiet part of one.
The check that skips an empty block measured the peak against the median. A
recording that is mostly burst measures its own burst against its own burst,
finds no difference and is discarded as silence, which is what happened to
every short capture. The gate's scatter had the same trouble one level down
and could come out above the peak, which is the one setting that cannot be
right, so it is now capped below it.
And the rule deciding where one message ends keyed on the middle gap in a
burst. Where a one is drawn as a gap three times a zero and most of the bits
are zeroes, the middle gap is the short one, twice it still falls inside the
message, and every one-bit ended a burst -- the message coming apart into
pieces of three pulses. It keys on the widest gap now, which is a fact about
the message rather than about the data it happened to carry.
The pieces of a message are also put back together after being sliced rather
than before. Grouping has to be tight, because the group is what fixes the
threshold and a group holding two sensors of unequal strength fixes it on the
louder; but the gaps inside one message run from two hundred microseconds on
the newer sensors to four thousand on the oldest, and a grouping tight enough
for the first tears the second into a bit at a time. So each piece is
measured at its own amplitude and joined to its neighbours afterwards, and
anything that comes out longer than the longest message there is gets cut at
its largest gaps.
Measured rather than argued: across four hundred and eighty combinations of
pulse and gap timing, 464 now read where 417 did; across twenty-one gap-keyed
combinations, 18 where 12 did; and eighty seconds of receiver noise still
yields nothing at all.
--from-iq FILE reads a saved capture instead of the receiver, so a recording
made where the aerial is can be worked on anywhere, as many times as it takes.
Everything downstream of the dongle is the real thing, which is what tells a
receiver problem and a decoder problem apart. --save-iq writes the settings
beside the samples, a file of raw samples with no record of its rate being
unreadable by anything. The invented garden now waits like a dongle instead
of running as fast as the machine allows, which it should have done from the
start: --seconds meant nothing against it and a capture came out fifty times
too large.
Full suite 2331 passed; this work checked against sixteen deliberately broken
builds, two of which it survived until the tests were made to catch them, and
one change was removed for being unable to earn a test at all. Built as
2026-09-07_03.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016PsWPTweCT6pwxKngvVxcg
Reported as reading nothing at all with several sensors in range. Two faults,
either of which is enough on its own, and both of them things I assumed rather
than checked -- this was written against messages I generated myself and never
against a sensor.
The first is the threshold. Bursts were found by setting one level per second
of band, halfway between the noise floor and the loudest thing in that second.
That is the obvious way to write it and it is wrong: a sensor on the windowsill
and a sensor at the end of the garden differ by forty decibels, so a level set
halfway to the near one sits above everything the far one ever does. The far
ones do not come through weakly, they vanish -- and vanish only while the near
one is transmitting, which is as confusing a symptom as radio produces. A
block with one loud sensor in it yielded exactly one sensor however many were
out there.
Finding bursts is now two passes. The first asks only where anything happened
at all and asks it against the noise -- the bottom fifth of the second, which
is noise however busy the rest was, and which does not move when something
loud arrives. Whatever clears that is grouped into regions, and the second
pass re-thresholds each region against its own high and low. Every sensor is
sliced at its own amplitude. Six sensors spanning eighty times in strength
now all come back from one second of band.
The second fault is that the slicer knew how a bit is drawn. It read a pulse
by comparing it with the gap that followed, which is right when the gap is the
complement of the pulse so that every bit takes the same time, and wrong when
the gap is a fixed spacer: a two-hundred-and-twenty microsecond pulse against
a two-hundred microsecond spacer is the longer of the two and reads as a one,
which is the wrong bit, and then every message fails its checksum having said
nothing about why. Nothing is assumed now -- not which of the pulse and the
gap carries the bit, not whether the gap is a complement or a spacer, not
which of long and short means one. The same burst is read half a dozen ways
and the checksums say which reading it was, at most one being able to satisfy
one. Copies are counted per message rather than per reading, or two readings
of one burst would corroborate each other and the rule protecting the two
thinly-checked models would protect nothing.
Both were caught the same way: by measuring, rather than by reading the code
again. A thousand seconds of the invented garden still yields no sensor that
is not there, and reception of the ones that are is up by a quarter, because
bursts that used to be masked now decode.
And, because none of the above should have needed me: `bandsaunter weather
--diagnose` prints each second taken apart stage by stage -- the noise, the
level a burst must clear, the loudest thing in the block, then every burst
with the lengths of its pulses and gaps and whatever was made of them. Those
lengths are the useful part: a real message has two or three of them and
nothing in between, which says at a glance whether the trouble is the radio or
the arithmetic. At the end it says which of five things it was: nothing
arriving, nothing above the noise, something never keyed, bursts that framed
as nothing, or messages that framed and arrived only once. `--save-iq FILE`
keeps the raw samples for whatever that cannot settle.
Full suite 2286 passed; the new work checked against six deliberately broken
builds. Built as 2026-09-07_02.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016PsWPTweCT6pwxKngvVxcg
A consumer weather station is two things. The display on the kitchen wall is
one of them; the other is a plastic box on a fence post that says what it can
see every sixteen seconds, in the clear, to anyone who happens to be
listening. This reads the box.
A section of its own, like the aircraft one, and for the same reason: it does
not fit through the scanner. A sensor message is a burst of a carrier
switched on and off, a fifth of a second long, and the scan path is a squelch
and a recorder -- it would record the bursts as clicks in a WAV file and
decode nothing. `bandsaunter weather` listens, `bandsaunter readings` reads
a log back, `bandsaunter sensors` says what is out there. Item 6 in the main
menu is the same thing without a command line.
Five families: the Tower 592TXR, the 5-in-1, the 6045M lightning detector,
the 609TXC and the 606TX. Temperature, humidity, wind speed and direction,
rainfall, strike counts, how far off the storm is, and battery state from all
of them. Every one is implemented from its published description and checked
against frames built from the same description, which proves the framing, the
parity, the checksums and the arithmetic and is not the same as having held
one of each.
The naming is the point. A sensor broadcasts an identity, and that identity
is a number that came out of a hat in a factory; it tells one sensor from
another and is no use at all for telling which is which. So press n while
listening: the display comes down, the sensors are listed, you name one, and
it goes back up, with the receiver running throughout. That is the moment it
is possible -- the sensor is on the screen saying 3.1 degrees, and the person
watching is the one who knows that the cold one is the shed. An hour later it
is a list of hexadecimal again. Names are written the instant they are given
rather than at exit, to a neighbouring file renamed over the old one, and one
given before a sensor has ever been heard waits under its identity and moves
across when the first message says which model it is.
Four things keep the neighbours' doorbells off the display. The checks the
message carries; a second copy, for the two models that carry only one byte
of check between them; a plausibility range, because a checksum can be
satisfied by a message the hardware could not send; and where in the burst
the message sits. That last one is the one that is easy to miss: a seven-byte
message read out of the front of a real eight-byte one is made of that
message's own payload bytes, whose parity is already correct, so the parity
bits contribute nothing and one byte of sum is all that is left -- and
corroboration cannot help, the three copies being identical. What gives that
window away every time is that it ends a whole byte before the burst does.
The Atlas is nine bytes like the lightning detector and lays its payload out
differently, so every decoder insists on a message type it knows. Anything
else that frames correctly is reported with its identity and no weather,
because wrong weather under somebody's sensor name is a worse answer than
none.
ism.py now delegates to this rather than keeping a second implementation of
the tower sensor, which fixes the channel letters -- A is 3, B is 2, C is 0,
and there is no D -- and the battery bit, which is set while the battery is
good. The two thinly-checked models are not reported from a scan at all: a
scan hears one burst, and they need two.
The option menus are now handed the module that owns the options rather than
importing the aircraft one, so one set of screens drives both sections and
will drive a third.
169 new tests, checked against nineteen deliberately broken builds; two of the
tests were too weak to notice their own mutation and were rewritten. Full
suite 2252 passed. Built as 2026-09-07_01.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016PsWPTweCT6pwxKngvVxcg
An audit rather than a feature, prompted by wanting this fit to hand to
somebody else.
Five options had no command-line flag written down in the table the manual
is generated from -- location, hold, schedules, tile_url and speed_unit --
so five flags that exist were missing from the manual. Four of them did
exist under other names and are now recorded; hold had no flag at all and
has one.
Four switches could be turned off from the command line and not back on:
--no-lookup, --no-basemap, --no-airports and --no-labels had no positive
halves, so an option turned off in the saved settings could not be turned on
again for one run. All four now have both.
And adsb, which opens the window and draws a map when it stops, could not be
given any of the settings that decide what those look like: no --at, no
--radius, no --tiles, no --map-brightness, no --width, --fps, --trail,
--fade, --stale, --airports or --labels. It takes all of them now.
The manual had no list of the aircraft options at all -- the ADS-B sections
were hand-written prose -- so five of them appeared nowhere in it. It now
generates an AIRCRAFT OPTIONS section from the same table the menu and the
flags come from, and the readme carries a table of all thirty-four with
their flags and defaults. Three tests hold the three of them together: one
that every option records its flag, one that every flag the table claims
actually exists on a command, and one that the readme names them all.
The installing instructions now list every dependency rather than only the
optional ones: the four Python packages with their names in Debian, Fedora
and Arch, and librtlsdr, which is a C library and therefore the one thing
pip cannot bring and a virtual environment cannot supply. What reaches a
network is written down too -- which host, when, and which file it is cached
in -- since somebody installing this on a metered or air-gapped machine has
to be able to see that nothing is fetched behind their back. build-repo.sh
needs dpkg-dev and apt-utils, which a minimal system does not have, and now
says so.
Verified rather than asserted: a clean virtual environment, pip install from
this tree, and a real log read back through the installed command. It pulls
seven wheels rather than the four the page claimed, the other three being
what Rich brings with it.
Also: matplotlib is gone from the readme's dependency table, nothing having
imported it; ffmpeg and Qt are in it, both having been missing; ffmpeg is a
Suggests on the package; and resume.sh is ignored.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016PsWPTweCT6pwxKngvVxcg
A callsign is a flight number rather than a leg, and the free registers hold
one route per number, so an aircraft over Arizona kept being handed a hop
between two airports in Texas. Nothing on the air settles it: ADS-B carries
no origin or destination. A commercial schedule service does know, because
it holds the day's actual movements.
Four are wired up and all four are optional: FlightAware AeroAPI,
Flightradar24, OAG and Cirium. Each is asked before the free databases and
each answers for the moment the aircraft was overhead rather than for the
flight number in general, so the leg chosen is the one that was in the air.
With no keys set nothing changes at all: a source with no key is skipped
rather than asked and refused, and the free databases answer as before.
Keys come from the environment and are never written to the settings file,
because a settings file is meant to be copied between machines and pasted
into a message asking for help, and an API key is not. There is a test that
holds that line.
None of the four has been run against its live service, since each wants a
paid account. They were written from the published response shapes and are
tested against those shapes, so each reader finds what it recognises and
returns nothing otherwise: a service that has changed since costs a route
rather than a scan. Cirium's plain departureTime is local and carries no
offset, so the UTC field is preferred where it is there -- reading the local
one as UTC is up to half a day out, which is exactly far enough to pick the
wrong leg of the same number. Reading now happens inside the same guard as
asking, as an answer shaped differently from the documented one is the
failure most likely to actually happen.
And the map. The zoom is now chosen from how wide the picture is rather
than from the area alone, with half again over the width fetched and
averaged down, since a downscaled tile is sharp and an upscaled one is not.
The window fetches a little more world than it shows so panning does not
leave the ground blank, and now fetches that bigger piece at the bigger
piece's own size: rendering it into the window's own pixels and stretching
it back was a fifth of an upscale over the whole map, which is what a sharp
map looks like when it looks blurred. The comment in the fetcher said the
opposite of what the code did, which is how it stayed hidden.
At 1920 by 1080 over a hundred miles the tiles now hold about 1.6 times the
pixels the window wants. At 3840 by 2160 the tile budget is reached, the
zoom stops climbing and the map is enlarged after all; a smaller radius buys
the detail back, and somebody else's tile server is not a thing to fetch a
thousand tiles from for one picture. The README says so rather than
implying otherwise.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016PsWPTweCT6pwxKngvVxcg
The terminal board says what is overhead. This says where: a real map
with the aircraft moving on it as the frames arrive, and beside each one a
box carrying everything known about the flight -- type and registration,
who operates it, where it came from and where it is going, height with a
rate of climb, speed and heading, how far away and on what bearing, its
position, how many frames it has sent and how long since the last one.
Qt is asked for and not required. Four bindings are tried, the module
imports on a machine with none of them, and asking for the window without
one gets the instructions rather than a traceback -- before the receiver
is opened, since nothing is gained by taking the dongle for a window that
cannot be drawn.
In the menu, "listen now" is now "passive capture" with a realtime
display beside it. Closing the window leaves exactly the files pressing
control-C leaves, because listen and watch share one read loop and one
finishing step; the receiver runs on its own thread, so a slow repaint
cannot cost a frame and a slow tile fetch cannot stall the picture.
The animation's labels grew to match: flight level and speed, type and
registration, and both ends of the route, each with a small flag of the
country its airport is in. The flags are a table rather than a network --
twelve pixels by eight, where a flag is the arrangement that makes one
recognisable rather than a rendering of the real thing -- and a country
not in the table is named by its two letters, since a flag that is nearly
another country's is worse than none. Where a route arrives as bare
codes the country comes from the ICAO prefix.
Four things found on the way. The window ignored --seconds, so "listen
for ten minutes" meant something different with a window open; it closes
itself now. The register was being asked twice per aircraft, once for
labels and once for airport positions. Cached routes had no country in
them, so the first real redraw drew no flags at all -- routes are
versioned now. And past fourteen aircraft on one frame the labels go
back to the callsign, the height and the speed, because five lines beside
each of three hundred aircraft is a page of overlapping text with a map
somewhere behind it.
Long names are folded rather than allowed to stretch a box, breaking at
the arrow of a route so the two ends stay whole; and the animation's
label placement gained the same ring search the window uses, having only
ever tried four spots.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016PsWPTweCT6pwxKngvVxcg
The program had no licence file at all, and pyproject claimed MIT into
the void. It is now the GNU General Public License, version 3 or later:
LICENSE holds the text verbatim, pyproject declares it with the OSI
classifier, both .deb builds write /usr/share/doc/<pkg>/copyright in the
machine-readable format Policy requires, both manuals carry a COPYING
section, and --version prints the GNU notice on both programs.
INSTALL.md is the step-by-step: what you need, the Debian package, the
virtual environment for everywhere else, how to check it worked, every
optional dependency with what it buys and what happens without it, and
the errors people actually hit first -- PEP 668 at the top, because on
Debian a plain "pip install ." refuses and reads as a broken program.
Speech transcription gets its own four steps, because it is the only
part with a real download in it: the recogniser into the environment
bandsaunter runs from, checking it took, the model (base.en, 148 MB,
from Hugging Face into ~/.cache/huggingface, fetched deliberately rather
than in the middle of a scan), then turning it on. With the sizes of
every model, the offline routes, and what to do when --engines says no
although pip says yes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016PsWPTweCT6pwxKngvVxcg
The Dust Council