The OSD-lite decoder was the source of FT2 false positives. It tries
685 CRC-14 checks across 5 passes (1 + 16 + 120 per pass), giving a
~4% chance of accepting random noise as a valid decode.
The reference implementation (decode174_91) verifies OSD results
against the received signal; the trx-rs OSD-lite only checked CRC.
Add ft2_count_hard_errors_vs_llr() which counts how many of the 174
coded bits in an OSD candidate disagree with the received hard
decisions. A legitimate correction disagrees in very few positions;
a false CRC match on noise disagrees in ~40-50 parity positions.
Reject OSD results with more than 36 hard errors.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stanislaw Grams <stanislawgrams@gmail.com>
The ft8_wrapper.c references ft4_encode and ft8_encode from encode.c,
but encode.c was not included in build.rs, causing linker errors.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stanislaw Grams <stanislawgrams@gmail.com>
The previous SNR formula (cand->score * 0.5 - 29.0) used the adjacent
tone bin as a noise reference. On a crowded FT8 band that bin is often
occupied by another station, inflating the apparent noise floor by
10-15 dB and capping reported SNR at around -10 dB even for strong
signals.
Replace with ftx_post_decode_snr(): re-encode the decoded message to
obtain the exact per-symbol tone sequence, compare each signal bin
against the minimum of the remaining (noise-only) bins, average over
all valid symbols, and apply the WSJT-X 2500 Hz bandwidth correction
dynamically per protocol. This produces accurate SNR estimates for both
FT8 and FT4 regardless of band occupancy.
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: Stanislaw Grams <stanislawgrams@gmail.com>
The OSD-3 (triples) path over 5 LLR passes was doing ~11,600 CRC checks
per candidate. With a 14-bit CRC this gives ~0.7 expected false positives
per candidate — far too high.
Remove OSD-3 entirely. Cap max_candidates at 16 for OSD-1/OSD-2, giving
136 CRC checks per pass (680 total). Gate OSD-lite behind a check that
LDPC reached within 6 parity errors of converging, so it only fires when
the LLRs are already trustworthy. Combined false-positive rate drops to
~0.04 per near-miss candidate.
Also remove the now-unused ft2_osd_decode and ft2_codeword_distance
functions.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Increase BP/SP iteration count from 30 to 50 to match WSJT-X reference
and give belief propagation more opportunities to converge near-threshold
candidates.
Replace the parity-based OSD-1/OSD-2 fallback (which required LDPC to
have nearly converged) with ft2_osd_lite_decode applied to all five LLR
combination passes. The CRC-based decoder works directly from raw LLRs
without depending on LDPC convergence, searching the 24 least-reliable
systematic bits for up to three bit errors via OSD-3.
Also increase max_candidates in ft2_osd_lite_decode from 12 to 24 for
broader coverage of likely error positions.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Diagnostic logging showed the FT2 BP/SP decoders consistently reach
1-8 residual parity errors rather than zero — the LLRs are correct
in direction but LDPC belief propagation stalls just short of
convergence.
Add ft2_osd_decode() implementing Ordered Statistics Decoding orders
1 and 2: after the five-pass BP/SP loop fails, sort the 174 codeword
bits by |LLR| ascending and trial-flip single bits (OSD-1, always)
or all pairs of the 50 least-reliable bits (OSD-2, when the remaining
error count is <= 4). Each trial costs one O(83) parity check;
worst-case overhead is ~1300 checks per candidate, negligible next to
the 5 x 30-iteration BP/SP passes already performed.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Add err=N/N/N/N/N to the FT2 window diagnostic log line, showing the
minimum number of unsatisfied parity equations across all candidates
for each of the five LLR passes. This makes it possible to distinguish
between a signal-quality-limited failure (small error count) and a
systematic decoder bug (large error count), which is the key unknown
in diagnosing the current FT2 LDPC non-convergence.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Normalize FT2 log-likelihoods before LDPC and fall back to\nthe standard waterfall candidate decoder when the raw FT2\npath produces no decodes.\n\nCo-authored-by: OpenAI Codex <codex@openai.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Use an FT2-specific raw sample window and candidate acquisition path.
Keep the build clean by removing the stale monitor warning.
Co-authored-by: OpenAI Codex <codex@openai.com>
Signed-off-by: Stanislaw Grams <stanislawgrams@gmail.com>
Use FT2-specific analysis settings and a wider receive span.
Switch server-side FT2 decoding to a rolling async window.
Widen FT2 candidate timing search in the vendored decoder.
Co-authored-by: OpenAI Codex <codex@openai.com>
Signed-off-by: Stanislaw Grams <stanislawgrams@gmail.com>
Implement a distinct FT2 protocol path in the decoder stack and align\nits timing with the confirmed FT2 framing used by Decodium.\n\nCo-authored-by: OpenAI Codex <codex@openai.com>
Signed-off-by: Stanislaw Grams <stanislawgrams@gmail.com>
Skip decoded candidates where ftx_message_decode() returns a non-OK
status instead of forwarding a synthetic error string.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
The score from ft8_lib is an averaged uint8 difference between Costas
sync tones and their neighbours (each unit = 0.5 dB). The previous
score * 0.5 gave the signal-above-adjacent-noise in dB relative to a
single 3.125 Hz waterfall bin, yielding values of +5 to +50 dB —
all wrong.
Subtract 10*log10(2500/3.125) ≈ 29 dB to normalise to the 2500 Hz
reference bandwidth used by WSJT-X and expected by PSKReporter:
snr = score * 0.5 - 29.0
This maps score 10 (minimum decodable) → -24 dB and score 60 → +1 dB,
matching typical WSJT-X SNR report ranges.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Add [workspace.package] version = "0.1.0" to the root Cargo.toml and
switch all 21 member crates to version.workspace = true so the entire
workspace is versioned from a single place.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Move trx-ft8 and trx-wspr into src/decoders/ alongside a new trx-aprs
crate that extracts the Bell 202/AX.25 decoder from trx-server, giving
all three modems a consistent crate-per-decoder layout.
- src/decoders/trx-ft8/ (moved from src/trx-ft8/)
- src/decoders/trx-wspr/ (moved from src/trx-wspr/)
- src/decoders/trx-aprs/ (new — Bell 202 AFSK + AX.25/APRS decoder)
- trx-ft8/build.rs: fix external/ft8_lib relative path after move
- trx-server: drop decode::aprs module, use trx_aprs::AprsDecoder
- AprsPacket stays in trx-core (mirrors Ft8Message / WsprMessage)
- Workspace Cargo.toml updated with new member paths
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
sjg
