A new section, alongside the aircraft, the weather sensors and APRS.
FT8 is the odd one out among the things this program listens to, and the
reason is worth stating because it shapes everything below. Every station on
the band transmits in the same quarter-minute slots, on the same dial
frequency, fifty hertz wide each, stacked across three kilohertz of audio.
One receiver parked on one frequency therefore hears the whole band's worth of
stations at once -- and hears most of them well below the noise, because half
of what is sent is error-correcting code. That is the entire trick: a rate of
about one half buys a mode that decodes twenty-odd decibels under what an
operator can hear. A receiver that took the loudest tone of each symbol and
hoped would decode almost nothing, which is why the tone detector reports how
confident it is bit by bit rather than what it thinks it heard.
Written from first principles except for two tables. The checksum, the
belief propagation over the sparse graph, the Costas sync search, the
waterfall, the soft-bit metric, and the seventy-seven bits that hold two
callsigns and a grid square are all here. The generator and the parity-check
matrix are not: they cannot be derived, being the code itself rather than
consequences of anything, so they are taken from ft8_lib under its MIT licence
with the attribution it asks for, and said so in the readme, the manual and
the file. No decoding logic came with them. That the two agree -- and they
are not derivable from one another, the generator's parity half running to
fifty-odd bits a row against the sparse matrix's six or seven -- is a test
rather than an assumption.
Tested against the air, not against itself. Eleven off-air recordings with
published decodes: ninety-seven of a hundred and fifty messages, no false
decodes, timing within a hundredth of a second, frequency within a hertz,
signal reports within half a decibel on average. The third not decoded are
the weakest in each slot; a mature decoder subtracts what it has decoded and
looks again in the remainder, and does ordered-statistics decoding where
belief propagation fails, and neither is built here. What is here decodes
nothing that other receivers did not also hear, which is the property that
matters in a log. Ten whole codewords lifted off the air are in the tests as
a permanent fixture, so the recordings can go missing and the regression
cannot.
Three things that looked like bugs and were not, and three that were. The
half-second timing discrepancy was the convention: a transmission is 12.64
seconds in a slot of fifteen and everybody starts half a second in, so
lateness is reported against that. Synthetic signals at known offsets proved
the clock self-consistent before anything was changed. The signal reports
were twenty-one decibels optimistic because those recordings have a receiver
passband above three kilohertz, putting a whole-band median twelve to sixteen
decibels below the real noise floor -- so noise is now measured beside the
signal, and in the tone that was actually sent rather than the loudest of
eight, the largest of eight noisy numbers being well above their mean even
with no signal at all. And the test transmitter was thirteen decibels
pessimistic, scaling its noise into a fifty-hertz reference instead of the
sampled bandwidth, which made the decoder look deaf when it was the test
signal that had been quietly attenuated.
The real bug the simulator caught was a one-block timestamp error: samples
were dated a block earlier than they were taken, which slid every slot slice a
second late and cut the first half-second -- three symbols, part of the
opening Costas array -- off every transmission on the band. One decode a slot
became six.
Reachable both ways, as everything here is. Twenty-one options, every one of
them a command-line flag and a line in the menu, both built from one table so
they cannot disagree -- and a test that says so, since an option in no group
would be settable from the command line and invisible in the menu. The band
list says which channels a plain dongle can reach and which need an
upconverter, because almost all the activity is on shortwave and finding that
out by listening to silence for ten minutes is the wrong way to learn it. The
default is two metres, which a plain dongle can hear.
--grid turns decodes into geography: every CQ says where it is, so each gets a
distance and a bearing and the furthest is named. --adif writes the log in
the form every amateur logging program imports, marked as heard rather than
worked, because nothing here transmits and an ADIF that let a logging program
treat these as contacts would put claims into somebody's log that they cannot
make.
One bug shipped and found by being used rather than by being tested: the
line that opens the receiver called a function this program has never had.
Every test reached it through the simulator, which takes the other branch, so
the one line that matters to somebody with an aerial was the one line never
run. There is now a check that every name these modules import actually
exists -- it names the missing one rather than failing somewhere downstream --
and two that say a receiver which cannot be opened is reported rather than
raised, and that nothing claims to be listening before there is one. It had
been announcing the frequency first, so a dongle that would not open read as
listening that had gone wrong.
Ninety-six new tests. Full suite 2781 passed. Built as 2026-09-21_04.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016PsWPTweCT6pwxKngvVxcg
93120b80a6