DeepFist neural CW decoder: DeepFist + Auto (hybrid) RX CW engines, with confirmed-call learning (ONNX Runtime) by n9bc · Pull Request #7 · N8SDR1/Lyra-SDR-cpp

n9bc · 2026-07-15T10:02:58Z

Summary

Adds DeepFist, a neural (spectrogram → CNN + CTC) CW/Morse receive decoder, as
additional selectable RX CW engines alongside the existing fldigi-port ("Classic")
decoder. Three engines: Classic (default, unchanged), DeepFist (neural), and
Auto (a hybrid arbiter that runs both and shows whichever is more confident). Runs via
ONNX Runtime (CPU). Classic remains the default; nothing changes unless the operator
selects DeepFist or Auto.

Scope note: this PR grew from "second engine" to the full DeepFist feature set (Auto
hybrid + confirmed-call learning + a dev-only training-capture path) so it can be
reviewed and land as one coherent unit.

Engines

DeepFist — src/dsp/deepfist/: resampler (48k→3200), front-end conditioner,
from-scratch STFT spectrogram, greedy CTC, ONNX model wrapper, streaming adapter, and a
CTC-lattice callsign rescorer (Super Check Partial).
Auto (hybrid) — src/dsp/CwArbiter.*: runs Classic and DeepFist in parallel and
displays the more-confident source moment-to-moment, falling back to Classic on fades.
Copy is source-marked so you can see which engine produced it.

Confirmed-call learning (on by default, no UI toggle)

Decoded callsigns that are also confirmed live by an RBN/cluster spot near the tuned
frequency are remembered in a local scp_local.txt (AppData). This feeds the rescorer's
candidate set and the "known-real" (amber) highlight, so the decoder improves at the calls
you actually work.

Text only — never audio. Stores callsign strings + timestamps; capped (10k) and
aged out after a year.
Opt out with LYRA_CW_LEARN=0; delete scp_local.txt to reset. Privacy surface is
public callsigns, local-only.

Training capture — developer tool, OFF by default

A trust-tiered CW audio capture path for collecting real-world DeepFist training data. It
has no UI and is disabled unless armed with LYRA_CW_HARVEST=1. When armed, it writes
int16 WAV segments + JSON sidecars to Documents/Lyra/cw_harvest, size-capped (2 GiB,
oldest-evicted), local-only. Not part of the normal user experience.

New dependency

ONNX Runtime (Microsoft, MIT), CPU build — shipped as a DLL beside the exe (load-time
import). Headers committed under third_party/onnxruntime/; binaries git-ignored. This is
why the diff is large — ~22k lines are ORT headers.

Model & data artifacts (NOT in git)

models/deepfist.onnx (+ sidecar) and MASTER.SCP are git-ignored and distributed via the
installer/release. CMake copies them next to the exe on build.

Licensing

DeepFist is MIT (© Brent Crier, N9BC); ONNX Runtime is MIT — GPL-compatible. As integrated
it's part of Lyra's GPLv3+ aggregate; per-file SPDX headers on src/dsp/deepfist/* record
the MIT provenance. See NOTICE.md / CREDITS.md.

Verification

Builds clean (VS2022) and runs; coexists with ZeroBeat and the classic decoder. Rebased/
merged up to date with main (v0.20.0). test_rescore checks the CTC forward pass against
PyTorch F.ctc_loss; test_neural_cw / test_neural_cw_stream match the Python/ONNX
reference; Classic + DeepFist smoke-tested decoding live.

Reviewer notes

Opt-in engines, opt-out learning, dev-gated capture — Classic default; Learn is
text-only + disableable; Harvest is invisible unless env-armed.
Real Lyra code lives in src/dsp/deepfist/, src/dsp/CwArbiter.*, src/wdsp_engine.*,
src/spotstore.*, and src/qml/CwDecoderPanel.qml; the rest of the diff is vendored ORT
headers.
Known limitation: on marginal/fading signals DeepFist's presence gate is less robust than
Classic (documented in docs/DEEPFIST.md); Auto and Classic cover that case.

🤖 Generated with Claude Code

Design for adding DeepFist (self-trained CNN+CTC neural CW decoder, MIT) as a second selectable RX CW engine alongside the fldigi port, via ONNX Runtime (C++, CPU). Covers model contract, new src/dsp/deepfist module, engine selection in WdspEngine, dockable pop-out DeepFistPanel, licensing clearance, and a code-only (model out of git) delivery. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Native C++ port of the DeepFist neural CW decoder (spectrogram -> CNN+CTC ONNX -> greedy CTC), mirroring the verified Rust reference (diddle cw_neural.rs). Qt-free DSP under src/dsp/deepfist/: - DeepFistResampler : 48k->3200 Hz decimator (121-tap, 1440 Hz), exact match to the reference filter - DeepFistSpectrogram : STFT 256/hop48, 65 bins 400-1200 Hz, log1p, per-window unbiased standardize; from-scratch FFT - DeepFistCtc : greedy CTC decode - DeepFistModel : ONNX Runtime session + sidecar token table - NeuralCwDecoder : streaming adapter (rolling 6 s window, re-decode every ~1.5 s), same shape as the classic CwDecoder Vendors prebuilt CPU ONNX Runtime 1.20.1 (MIT) under third_party/onnxruntime (headers + license committed; DLL/model git-ignored, code-only PR). CMake links ORT, ships onnxruntime.dll + models/ next to the exe. Verified: scratch/test_neural_cw decodes known clips (3200 Hz direct path + 48 kHz decimator path) with a 0-mismatch exact match to the Python-ONNX reference on CQ TEST DE K3E / AGN / 5NN 001 K / CQ OK8G / R FB TU 73. Licensing: DeepFist MIT (c) Brent Crier + ONNX Runtime MIT, both permissive inside Lyra's GPLv3; no AGPL (DeepCW/HamNoise) code used. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Adds DeepFist as a second, selectable CW decode engine beside the classic fldigi port: - WdspEngine: NeuralCwDecoder member + cwEngine_ switch (0=Classic,1=Neural), Q_PROPERTY cwDecodeEngine + cwNeuralAvailable, cwNeuralText signal. The CW-mode-gated audio tap routes cwMonoBuf_ to the selected engine only. The model is lazy-loaded on first switch to Neural (from <exeDir>/models or $DEEPFIST_MODEL_DIR); engine flips to Neural only after the ONNX session is fully built, so the audio thread never touches a half-constructed session. - CwDecoderPanel.qml: Classic/DeepFist selector + model-status line. The neural engine's full 6 s-window decode shows in replace mode (cwNeuralText); the fldigi-only knobs (speed/BW/tracking/squelch/signal) are hidden for Neural, which is self-tuning. - Prefs: persist cwDecodeEngine; restored on panel load. - NOTICE.md / CREDITS.md: DeepFist (MIT) + ONNX Runtime (MIT) attribution. Not yet compiled in the full app: the v0.17.0 tree currently fails to configure because the installed Qt 6.11.1 kit is missing the Qt::SerialPort module (required by unrelated serial CW-key code). The DeepFist decode core is independently verified (see prior commit). Full-app build + live UI check are pending the SerialPort module. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Positive confirmation in the app log (%APPDATA%/N8SDR/Lyra-cpp/logs) that the ONNX model loaded, alongside the existing warning on failure. Verified via a launch smoke test: app starts with onnxruntime.dll, panel restores the Neural engine, and the model loads from the deployed models/ dir. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…decode) The neural engine emitted the full 6 s window every 1.5 s in replace mode, so the panel only showed ~6 s of text and updated in slow chunks. Replaced that with the reference live-decode approach (deepfist/scripts/tci_decode.py): - Frame-timed commit: each decoded char carries a frame time (greedyCtcFrames); it is emitted exactly once when its audio settles (~1.3 s old), streaming characters into an accumulating transcript instead of replacing a window. - Re-decode every ~0.4 s (was 1.5 s). - Inference moved to a worker thread so the faster cadence never stalls the audio callback; it idles when no new audio arrives (Neural not active). - NeuralCwDecoder.onText now delivers incremental committed text; the panel appends it (both engines now stream into the same transcript). DeepFistModel split into infer() (raw log-probs) + decode3200(); the one-shot harness still matches the reference exactly (7/7). New test_neural_cw_stream harness plays a 48 kHz WAV through the real streaming path: on DeepFist's real 25 s off-air capture it streams "…ARY 5NN … K6ZH TEST K7CO K7CO 5NN 6", matching the reference decode of that clip. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

DeepFist's shipped model is now exp15, trained on CONDITIONED single-signal audio, so inference must condition too (train == inference). Ports diddle's verified live Rust conditioner (cw_neural.rs::Conditioner) to C++: AGC (unit RMS) -> tone AFC (4096-pt FFT, 400-1200 Hz) -> complex downconvert + two cascaded 1-pole LPFs (~90 Hz) -> recenter to 600 Hz -> peak-norm. Runs on the 3200 Hz window inside DeepFistModel::infer, before the spectrogram. New src/dsp/deepfist/DeepFistConditioner.{h,cpp} (incremental-phasor form matching the Rust numerics). Verified: C++ conditioner+exp15 matches the Rust reference (diddle cargo example) EXACTLY on 5 clean clips; streaming the real audio_001 off-air capture now copies "WP3Z TEST WP3Z ..." (exp9 was patchy). Marginal fast/weak clips (untitled) sit at the model's decision boundary where Rust/Python/C++ all differ slightly — inherent model instability, not a port bug. Model artifact (exp15) refreshed in models/ (git-ignored). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Ports diddle's dsp/rescore.rs + scp.rs to C++: for each callsign-shaped word in a decoded window, swap in nearby MASTER.SCP candidates and score each against the model's CTC log-prob lattice (ctcNll = full-sequence CTC loss); the call the audio actually supports wins. Fixes garbles text edit-distance can't (WP3B vs WP3Z). Confident = margin >= 3 nats over the best different candidate. - src/dsp/deepfist/DeepFistRescore.{h,cpp}: ctcNll (log-space CTC forward), is_call_shape, run/span finding, rescoreCalls. - src/dsp/deepfist/DeepFistScp.{h,cpp}: MASTER.SCP loader + edit-distance candidate generation. - NeuralCwDecoder: loads MASTER.SCP beside the model; runs the rescorer per window on the worker thread; emits confident verdicts via onCalls. - WdspEngine: cwNeuralCall signal. CwDecoderPanel: a "Calls" readout of confirmed/corrected calls (deduped, ranked by confirmation count, green = confirmed / amber = corrected), click -> His Call. - MASTER.SCP bundled next to the exe (git-ignored, like the model). Verified: test_rescore matches PyTorch F.ctc_loss (1e-4) + flips A1B->A1C; streaming the real audio_001 capture confirms WP3Z at 5.4-9.1 nats (matches the DeepFist reference's 5.4-9.3). Runtime log: "SCP rescorer: 50038 calls". Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

# Conflicts: # CMakeLists.txt # src/wdsp_engine.h

LYRA_CW_WAV now writes UNCLAMPED 32-bit float (was clamped 16-bit) so the true decoder-input level/envelope are visible — a clamped capture hid that the live RX audio runs at ~x21 full scale (+19 dB makeup gain), which looked like clipping but is fine (the conditioner AGC normalizes it). test_neural_cw and test_neural_cw_stream now read fmt=3/32-bit float WAVs too. Used to diagnose the "horrible live decode": the audio is clean (decodes to KG4CB when a window is well-aligned); the streaming garble is exp15 being alignment-sensitive, and diddle's reference decoder garbles the same capture identically — i.e. a model limitation, not a Lyra bug. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

… finding Single resume anchor for the feat/deepfist-cw work: architecture/file map, build+run+verify procedure, environment gotchas, and the resolved "horrible live decode" investigation — the audio is clean and the Lyra port is faithful (diddle's reference garbles the same capture identically); the garble is exp15 alignment-sensitivity, a DeepFist model limitation to fix in that project. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…no chars) Ports DeepFist's live-copy-display fix (HANDOFF §18.23): the neural decoder was hallucinating streams of junk on empty frequencies because Lyra's CW AGC makes a steady ~700 Hz artifact tone whose LEVEL is indistinguishable from a real signal. KEYING separates them. Per-window keying ratio = p90/p10 of the locked tone's baseband envelope (DeepFistConditioner::keyingRatio); below 12 -> skip the decode, emit nothing, just advance the commit boundary. Chose DeepFist's keying_ratio (tools/squelch.py, VALIDATED on real Lyra x21 captures: dead air ~4, keyed CW 12-600+) over diddle's keying_depth, which gives a flat ~0.68 on Lyra audio and doesn't separate. Verified on the KG4CB capture: dead-air windows (score 3.5-5.5) now emit NOTHING; signal windows (12-671) decode. Real-signal copy quality on hand-sent CW is unchanged (still the DeepFist training frontier; exp16 not yet exported to ONNX). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…del) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…nfig) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

The CTC blank penalty was env-var-only (LYRA_CW_BLANKPEN, invisible). Make it a visible, live, persisted control so it can be tuned/A-B'd on the same signal without a relaunch: - NeuralCwDecoder: blank penalty is now a live atomic (setBlankPenalty / blankPenalty), read by the worker each decode; initial value still seeds from LYRA_CW_BLANKPEN. - WdspEngine: cwBlankPenalty property + setCwBlankPenalty (clamped 0..5). - Prefs: persist cwBlankPenalty; restored on panel load. - CwDecoderPanel: a "Blank penalty" slider (0..5, step 0.5) with a live readout, visible when the DeepFist engine is active. Higher = recover more dropped chars on weak audio; more spurious chars on strong signals. Default 0. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

- Blank-penalty slider now snaps to 0.5 steps (snapMode SnapAlways). - Space on long pause: the keying gate emits ONE space when a keyed-CW pause starts (idleGap), so copy across gaps doesn't run together (mirrors tci_decode's idle-gap space). - Removed the green "Calls" chips (rescorer verdicts) that overlapped the decode window; also unwired neuralCw_.onCalls so the per-window callsign-DB rescore no longer runs (CPU saved). Double-click a call in the transcript -> His Call still works (built into the decoded-text area). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

- WPM: the neural engine now estimates RX speed from the keying envelope (DeepFistConditioner::estimateWpm — threshold the tone envelope, dot = short on-elements, WPM = 1200/dot_ms) and emits it on the existing cwRxWpmChanged surface, so the header "wpm" populates for DeepFist too (was Classic-only). - Renamed the blank-penalty slider to "Sensitivity" with a Low/Med/High readout and a plain-language hint ("Lower = clean copy … Higher = pull weak code"). - RBN highlight: decoded call-shaped words that are currently spotted (SpotStore::isSpotted, exposed to QML as Spots) are coloured green in the transcript (RichText) — a real, active station stands out as it's copied. Re-highlights on Spots.changed(). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…aming) The double-click handler used selectWord()/selectedText, which the RichText transcript switch and the frequent DeepFist text updates (onTextChanged reset cursorPosition) could disturb -> empty/wrong call. Extract the call-shaped word straight from decodedText at the click position instead (wordAt); the right- click menu uses the same. No dependency on selection state or cursorPosition. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

- Removed the "• neural (rolling 6 s window)" reference next to the engine selector (kept only the "model not found" warning). - Sensitivity indicator now reads Low (<=0.4) or High, no "Med". Note: AR (+) / BT (=) prosigns are misdecoded by exp16 (BT -> X, AR dropped) — a DeepFist training gap, not a Lyra issue; needs more prosign data in training. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

- DeepFist neural copy now shows BT/AR prosigns as <BT>/<AR> (angle notation, matching <SK>/<KN> and the classic decoder) instead of the raw = / + aliases. Display-only: a misread prosign still shows what the model decoded. - Sensitivity (blank-penalty) slider is now bipolar (-2..+5, default 0): negative suppresses stray/doubled characters on strong signals, positive pulls weak code out of the noise. Signed numeric readout + numbered ticks (new LyraSlider showTicks/showTickNumbers). Clamps widened in Prefs and WdspEngine to match. - CwWavCapture (LYRA_CW_WAV): re-patch the RIFF/data size fields every ~1 s of audio so a capture stays a valid, playable WAV even when the process is killed before the static-local destructor runs (Qt app close doesn't guarantee it). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Add `SPDX-License-Identifier: MIT` + Brent Crier copyright to each src/dsp/deepfist/ file. DeepFist is MIT (© Brent Crier) integrated into Lyra's GPLv3+ aggregate — permissive-into-copyleft, already documented in NOTICE.md/CREDITS.md. Per-file headers make the provenance unambiguous now that these files carry an explicit license marker even if copied out of tree. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

- build3.ps1: per-machine build helper (like _b*.bat / _run_lyra_vac1.bat) - .claude/scheduled_tasks.lock: transient runtime lock - models/*.bak: local rollback copies of swapped-in models (the exp16 .bak guarding an exp27 swap) — 13 MB binaries that must stay out of git like the live .onnx artifacts already do Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

DEEPFIST_HANDOFF.md was session-handoff notes — a dated commit list, a resolved investigation log, a TODO list, and references to private local paths / a sibling repo. Replace with docs/DEEPFIST.md: a focused, public-appropriate integration reference (what it is + license/attribution, architecture & file map, model contract, front-end recipe, build/run, and verification) with the internal cruft and private paths removed. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…pot color - Decode pane: auto-scroll now triggers off the ScrollView viewport's contentHeight instead of the TextArea's. For RichText the TextArea height settled a layout pass late, so the newest (just-wrapped) line only appeared on the following update. Scrolling when the Flickable's contentHeight changes includes the new line immediately. - Sensitivity (blank penalty) slider narrowed to a symmetric -1..+1 (ticks at -1, -0.5, 0, +0.5, +1); Prefs + WdspEngine clamps narrowed to match so a stale persisted value can't sit past the slider ends. - RBN-confirmed callsign highlight brightened (#5fe0a0 -> #5fffa8) so a spotted station stands out more in the transcript. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Add the author's callsign (N9BC) to the DeepFist copyright line everywhere it appears — the 16 src/dsp/deepfist/ SPDX headers, NOTICE.md, CREDITS.md, docs/DEEPFIST.md, and the design spec — matching the "Brent Crier (N9BC)" form already used elsewhere in the repo. (DeepFist's own MIT LICENSE is updated in the separate DeepFist project.) Also fix a stale figure in docs/DEEPFIST.md: the Sensitivity slider range is -1..+1 (narrowed this session), not the old -2..+5. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

n9bc · 2026-07-15T10:04:46Z

Heads-up on the model artifacts (not in this PR by design):

The neural engine needs three runtime files that are intentionally git-ignored and not part of this diff:

models/deepfist.onnx + models/deepfist.onnx.json (the trained model + sidecar)
models/MASTER.SCP (callsign database for the rescorer)
third_party/onnxruntime/lib/onnxruntime.dll (+ import lib)

These are large binaries meant to ship via the installer / release asset, not committed to git (same pattern as other runtime binaries). CMake's POST_BUILD copies whatever is in models/ next to the exe.

What this means for review: the code builds and links fine without them (only the API headers under third_party/onnxruntime/ are committed), so this PR is fully reviewable as-is. They're only needed to run the DeepFist engine at runtime. See third_party/onnxruntime/README.md for repopulating lib/, and docs/DEEPFIST.md for the model contract. Happy to sort out the delivery mechanism (installer bundle vs. release asset vs. Git LFS) separately — just didn't want the absence of the binaries to look like a gap.

Record that the DeepFist engine is portable (non-Windows ONNX path already present; DSP is plain C++/STL) and ONNX Runtime ships for Linux/macOS — so it adds no new Windows lock-in. The blocker to a Linux/macOS build is the rest of Lyra (WDSP DLL, WinSock2; os_compat.h POSIX shim not yet wired in), not DeepFist. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…rvest) Adds Auto as a third RX CW engine: runs Classic + DeepFist unmodified and a thin engine-agnostic CwArbiter hands display ownership to whichever is trustworthy, using DeepFist's keyingRatio as the fade detector. Never goes silent; fallback copy is subtly source-marked. Second layer (CwCaptureHarvester) passively harvests a training corpus from arbiter events with trust-tiered labels, plus an RBN->SCP loop feeding verified calls back to the rescorer. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

TDD, bite-sized tasks: CwArbiter ownership state machine (pure unit + test_cw_arbiter target), onKeying feed from NeuralCwDecoder, WdspEngine cwEngine_=2 fan-out to both decoders, and the QML Auto selector with subtle source-marked (dimmed) fallback text. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…s engine switches Two-tier call colouring in the decode pane: any token that is an exact MASTER.SCP member renders amber (known-real call — DeepFistScp::contains, O(log n), surfaced as Q_INVOKABLE WdspEngine::cwCallKnown); an RBN/cluster spot still upgrades to bright green + bold (live, verified). A strict shape-only fallback (1-3 letter suffix) covers sessions where the SCP never loads; SCP arrival re-renders the transcript so earlier copy gains its amber. Future validation sources (Phase-3 RBN-confirmed local list, SDRLogger+ worked-before sync) OR into cwCallKnown without touching the QML. Engine switches no longer clear the transcript — the operator A/Bs engines on the same signal and tunes each; a " • " seam marks the switch point and the Clear chip still empties on demand. Operator-verified on air (NA2DX amber; AA0TT DE run-together no longer false-ambers). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

ScpLocal store (capped, aged, atomic rewrite; TDD), DeepFistScp::addCalls merge at model load, WdspEngine ownership + cwNoteConfirmedCall + Prefs.cwLearnCalls (opt-in per spec §6.6), QML word tap + Learn chip. Also folds in the Phase-1 escape bugfix: prefs engine clamp 0..1 → 0..2. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

… load

…f; fix engine-pref clamp to allow Auto

… list Also: restore a persisted Auto engine at startup (Prefs.cwDecodeEngine === 2 now survives restart, not just DeepFist), and wire ScpLocal.cpp/h into the main lyra target — it was only ever linked into the standalone test_scp_local unit test, leaving WdspEngine::cwNoteConfirmedCall unresolved at link time.

… placement clearDecoded() also clears the Learn tap's pending word (no cross-Clear word concatenation); the setCwDecodeEngine doc block moves back to its function (the two Phase-3 helpers had landed between them); ensureCwScpLocal's comment now states the privacy boundary (the file, not the mkpath'd dir — load() never creates it, note() is QML-gated by Prefs.cwLearnCalls). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

CwCaptureHarvester (Qt-free core): 60s 3200Hz ring fed from the CW tap via a parallel decimator, two v1 triggers (arbiter DeepFist->Classic fade -> hard_negative; RBN-confirmed call -> gold_rbn), per-tier debounce, post-roll, 2GiB retention. Segments = int16@3200 peak->20000 WAV + JSON sidecar (tier, keying trace, both texts, wpm, peak) — matching deepfist's curation-tool conventions; windowing/labeling stays offline in the deepfist repo. New CwArbiter::onOwnerChange callback; cwCaptureEnabled pref (opt-in per spec §6.6); Harvest chip. SILVER/agreement/SCP-margin tiers deferred (schema:1). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

RBN/cluster spots often land 10-60 s after a station starts up, so a cleanly-decoded call could finish its live tapWord() check before its spot existed - the transcript turned green retroactively but the learner never re-looked (W2DON on air, 2026-07-16). Re-scan the recent transcript whenever the spot bank changes; ScpLocal::note() dedupes so repeat scans are free. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

A 3-char decode garble ("M5N" on air, 2026-07-16) went green because it exactly matched a call spotted on a different frequency - isSpotted() answers for the whole spot bank. The decode panel copies the station under the cursor, so green highlight + Learn now use isSpottedHere(): same lookup, but the spot must sit within 1.5 kHz of the tuned CW carrier (same carrier convention as activate()). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ention Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

enforceCap() sorted segment paths lexicographically to pick deletion order. Filenames are "<tier>_<epoch>_<seq>.(wav|json)", and the tier string is the primary sort key, so every gold_rbn_* segment (highest trust) sorted before every hard_negative_* segment ('g' < 'h') and was evicted first regardless of actual age. Parse the age key from the trailing <epoch>_<seq> fields instead (the tier prefix itself contains underscores — "hard_negative", "gold_rbn" — so it's parsed from the end of the stem) and sort ascending by (epoch, seq), tier-independent. Adds a mixed-tier retention test that writes an older hard_negative segment and a newer gold_rbn segment, forces the cap to fit exactly one pair, and asserts the older pair is evicted regardless of tier name. Verified the new test fails against the pre-fix code (evicts the newer gold_rbn pair instead) and passes after the fix; full test_cw_harvest suite passes. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…, cwCaptureEnabled pref Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Critical defect found in review of 3cbdc12: ~WdspEngine() joined the harvest pump thread and called closeRx1(), but never stopped neuralCw_'s internal worker thread. C++ destroys members in reverse declaration order, and in wdsp_engine.h the harvest members (cwHarvestRing_/cwHarvestDecim_/cwHarvester_) are declared AFTER neuralCw_ and cwArbiter_ — so they were destructing BEFORE ~NeuralCwDecoder() ever ran (which is where its worker thread actually stops). That worker fires onText/onWpm/onKeying synchronously off its own thread (wired in setCwDecodeEngine), and those callbacks dereference cwHarvester_ (when cwCaptureOn_ is true) and unconditionally call cwArbiter_.updateKeying — a decode callback in flight during teardown could land on an already-destroyed harvester/arbiter. Fix: explicitly stop+join neuralCw_'s worker (new public NeuralCwDecoder::stop(), a thin wrapper over the existing private stopWorker(), which is already idempotent via running_.exchange(false)) at the top of ~WdspEngine(), before the harvest-pump join and before any harvest/arbiter member is touched. This covers cwArbiter_ for free, since the same worker thread is its only caller. Header member order is left untouched (lower-risk fix than reordering declarations). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The Learn re-scan re-confirms every spotted call still in the transcript on every spot-bank change, so a single call minted a near-identical gold_rbn segment every debounce window (observed on air: 5x W1AW/3 in 3 minutes). Per-call dedupe map with a configurable goldRepeatSec window (default 600 s); tier debounce unchanged. Test 10 covers same-call suppression, different-call pass-through, and window expiry. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

rescanLearn only looked at the last 1500 chars while displayHtml greens everything it renders (6000-char cap) - on air, KT4K sat visibly green but permanently unlearnable once it drifted past the window. Scan the full transcript; the per-call gold dedupe and ScpLocal note() throttle make repeat scans free. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Three on-air failures shared one root cause: the learn decision was re-derived at a different moment than the render that showed the operator a green call, and each variant lost calls to timing: - tap at word completion missed spots arriving seconds later (W2DON) - a 1500-char rescan window missed calls still visibly green (KT4K) - rescan-on-spot-change missed spots evicted from the 200-cap bank before the next bank event (KA1ULN) displayHtml''s green pass is the only place green-ness is guaranteed observed, so the learn note now fires there, memoized per call per 10 min (matching the harvester''s gold dedupe window). The invariant the operator actually expects - "if I saw it green, it''s learned" - now holds by construction. tapWord/pendingWord/rescanLearn machinery removed; spotRev re-render on spot-bank changes covers the late-spot case since the render IS the learner. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- scratch/test_neural_raw.cpp - deterministic tiled-window decode of a WAV slice through the real DeepFist pipeline (per-window keyingRatio probe + blank-penalty sweep); root-caused the presence-gate fade limitation and stays the replay tool for LYRA_CW_WAV captures. - scratch/gap_probe.cpp - keying-envelope on/off gap histogram. - scratch/rbn_fetch.py - stdlib-only RBN daily-archive fetcher/filter producing label CSVs for IQ-capture training sessions. - scratch/rbn_20260708_cwt19z.csv - proven sample output (CWT 19z, 19759 spots / 520 calls on 20m). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ev-only (env-gated) Learn and Harvest were opt-in CW-panel chips. A chip nobody understands raises support questions ("is Lyra recording me?"), and Harvest is really a training-data tool, not a user feature. Remove both chips: - Learn (RBN-confirmed calls -> local scp_local.txt, text only, never audio) is now on by default for everyone, gated by WdspEngine.cwLearnEnabled; disable with LYRA_CW_LEARN=0. Reset by deleting scp_local.txt. - Harvest (trust-tiered CW audio capture for DeepFist training) is developer- only, gated by WdspEngine.cwHarvestEnabled; off unless LYRA_CW_HARVEST=1. Both flags are CONSTANT, read once at construction. The Prefs.cwLearnCalls and Prefs.cwCaptureEnabled properties are removed entirely (no dead settings). gold_rbn capture still rides the Learn tap but only writes when Harvest is armed, so no audio is ever written without the developer opting in. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

# Conflicts: # src/qml/CwDecoderPanel.qml

… branch Brings the complete DeepFist feature set up to date with origin/main so PR N8SDR1#7 can carry it all: base neural decoder + Auto arbiter (hybrid) + RBN->SCP Learn + training Harvest (dev-gated) + render-time learn + the Learn/Harvest chip retirement. Build verified; merges cleanly into main. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

n9bc · 2026-07-19T00:58:19Z

Hi Rick — PR #7 is rebased up to date with main (v0.20.0) and shows clean/mergeable now.

Heads-up on scope: I folded the rest of the DeepFist work into this PR rather than splitting it, so it can land as one coherent unit. Beyond the original second-engine decoder, it now also includes:

Auto — a hybrid arbiter that runs Classic + DeepFist and shows the more-confident one
Confirmed-call learning — on by default, text-only (callsigns, never audio), local + capped, opt out with LYRA_CW_LEARN=0
A training-capture path — developer-only, no UI, off unless LYRA_CW_HARVEST=1

I flagged the privacy-relevant bits (Learn / Harvest) up front in the description since those are the ones worth your eyes. Classic stays the default; DeepFist/Auto are opt-in.

Two things I'd value your call on whenever you get to it:

Model-artifact delivery — deepfist.onnx + MASTER.SCP are git-ignored; installer bundle vs. release asset vs. Git LFS?
Whether one combined PR works for you, or you'd rather I split the learning/capture parts back out into a follow-up.

No rush — review at your pace.

…rf settings) into feat/deepfist-cw

…ference) into feat/deepfist-cw

brentc316 and others added 26 commits July 12, 2026 01:12

Merge remote-tracking branch 'origin/main' into feat/deepfist-cw

bd2e3fd

# Conflicts: # CMakeLists.txt # src/wdsp_engine.h

docs: handoff — keying squelch landed + export-exp16 next step

6557221

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

docs: handoff — exp16 exported + dropped into Lyra (now the active mo…

eddc337

…del) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

chore: gitignore .claude/settings.local.json (personal per-machine co…

ce4196c

…nfig) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Merge remote-tracking branch 'origin/main' into feat/deepfist-cw

4b324d2

brentc316 and others added 3 commits July 15, 2026 07:01

brentc316 and others added 26 commits July 15, 2026 13:52

feat(cw): ScpLocal store for RBN-confirmed local callsigns (Phase 3)

b9714f4

feat(cw): merge scp_local calls into the SCP candidate space at model…

6ccd017

… load

feat(cw): local confirmed-call store in WdspEngine + cwLearnCalls pre…

1f0a739

…f; fix engine-pref clamp to allow Auto

Merge remote-tracking branch 'origin/main' into feat/deepfist-cw

4c9f77a

feat(cw): CwArbiter onOwnerChange callback for the Phase-2 harvester

fdc2dc8

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

feat(cw): Phase-2 harvest support — audio ring, wav16 writer, sidecar io

1f98b35

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

feat(cw): Phase-2 CwCaptureHarvester — tiered capture, post-roll, ret…

0fc10f3

…ention Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

feat(cw): wire the Phase-2 harvester — tap feed, arbiter/RBN triggers…

3cbdc12

…, cwCaptureEnabled pref Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

feat(cw): Harvest toggle — opt-in capture of tiered CW training segments

629afae

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

docs: gold_rbn harvest requires Learn enabled (final-review M1)

cb5f22a

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Merge remote-tracking branch 'origin/main' into pr7-sync

66d0a07

# Conflicts: # src/qml/CwDecoderPanel.qml

n9bc changed the title ~~DeepFist neural CW decoder: optional second RX CW engine (ONNX Runtime)~~ DeepFist neural CW decoder: DeepFist + Auto (hybrid) RX CW engines, with confirmed-call learning (ONNX Runtime) Jul 18, 2026

brentc316 added 2 commits July 20, 2026 21:10

Merge origin/main (1f1a88a: multi-rig stages 1-4, P1+P2 discovery, pe…

527a0cc

…rf settings) into feat/deepfist-cw

Merge origin/main (bbddc17: RX Auto-LNA/auto-ATT rebuilt to Thetis re…

de1a180

…ference) into feat/deepfist-cw

Navigation Menu

DeepFist neural CW decoder: DeepFist + Auto (hybrid) RX CW engines, with confirmed-call learning (ONNX Runtime) - #7

DeepFist neural CW decoder: DeepFist + Auto (hybrid) RX CW engines, with confirmed-call learning (ONNX Runtime)#7
n9bc wants to merge 66 commits into
N8SDR1:mainfrom
n9bc:feat/deepfist-cw

n9bc commented Jul 15, 2026 •

edited

Loading

Uh oh!

n9bc commented Jul 15, 2026

Uh oh!

n9bc commented Jul 19, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

n9bc commented Jul 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Engines

Confirmed-call learning (on by default, no UI toggle)

Training capture — developer tool, OFF by default

New dependency

Model & data artifacts (NOT in git)

Licensing

Verification

Reviewer notes

Uh oh!

n9bc commented Jul 15, 2026

Uh oh!

n9bc commented Jul 19, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

n9bc commented Jul 15, 2026 •

edited

Loading