Skip to content

Receiver, DSP and scanning roadmap

Status: manual reception and assisted tuning have accepted slices; Phase 1 carryovers, Phase 2 proof work and the acceleration matrix remain open. WebSDR Phase 4 scanning and Phase 8 learned classification remain deferred/planned.

Scope and dependencies

The appliance owns serial-routed acquisition; browser workers own DSP, analysis, audio and display. Manual tuning, existing clients and independent lanes take priority over detection, scanning and automatic actions. Current procedures: receivers, DSP, detection, scanning.

Phase 1 tuning/UI corrections were operator-accepted on 2026-08-25; independent display-center navigation on 2026-08-31 and VFO history on 2026-09-04. Keep passband containment, preview-only ghost tuning, independent receiver release, persistent capture-wide waterfall and default-off browser AGC. The deliberate all-demodulator and authentication-enabled browser checks remain. Logical USB isolation never establishes switched VBUS: verify physical power independently on the operator's actual topology.

Phase 2 boundary corrections were accepted on 2026-08-26, without closing the broader proof matrix. Fast occupancy owns existence and parent identity; fine FFT evidence may refine only a confirmed parent's attached edges. Detached energy and colored-noise pedestals must not widen it. Candidate analysis remains passive until explicit activation. Phase 3A was accepted on 2026-08-26 and 3B on 2026-08-27; subsequent decoder authority remains separately gated.

The 2026-08-30 acceleration correction separates ingress, realtime, display, scientific, candidate and decoder consumers. Bounded branch-local loss and transferable fallback preserve audio/manual control. Scientific FFTs through 4M do not require a 4M-element JavaScript display/detector scan. WebGPU, WebGL2 and Canvas2D fallback must remain distinct. Partial smooth-display observations do not close the live matrix below.

Performance profile targets

Early benchmark proposals paired Eco with roughly 1 MS/s, 10–15 waterfall lines per second and a single lightweight decoder. Balanced targeted around 36 frames per second at 2.4 MS/s with a 65K FFT and several seconds of replay. Analysis emphasized the highest proven loss-free rate, deeper history, concurrent decoders and higher-order features. SigMF capture, optional ONNX inference and a general on-demand native decoder broker are separate scope. The browser DSP guide lists the implemented settings. Measure display cadence, RF continuity and power on the hardware under test.

A proposed bandwidth-constrained mode would channelize on the Pi and send narrowband audio (potentially Opus) and a reduced spectrum. This would trade Pi processing and capture-wide replay for lower network usage. The interactive modes currently stream raw IQ. Profile selection leaves the host CPU governor under operator control.

Acceptance criteria and open gates

These are retained acceptance records, not qualification of the current source. Checked items keep their recorded scope; implementation, qualification and operator acceptance are distinct. Later work does not close unchecked gates.

Phase 0 — Architecture and Replay Harness

  • [x] ~~Decide project license.~~
  • [x] ~~Define the web transport protocol.~~
  • [x] ~~Define receiver, signal, hypothesis, scan-plan, scan-result, and decoder event schemas.~~
  • [x] ~~Define decoder manifests.~~
  • [x] ~~Define receiver-path profiles, including RF/tuner offsets and supported scan ranges.~~
  • [x] ~~Build CU8/SigMF/WAV replay tools.~~
  • [x] ~~Establish performance and correctness benchmarks.~~
  • [x] ~~Produce the visual design system and interactive desktop/mobile mockup.~~
  • [x] ~~Validate Phase 0 deliverables and update the completion ledger.~~

Phase 1 — Manual Web Receiver

  • [x] ~~Implement the minimal rfsentry-api gateway.~~
  • [x] ~~Read live receiver inventory from /etc/rtl_tcp/*.conf.~~
  • [x] ~~Define and validate serial-referenced route policy in /etc/rfsentry/routes.d/*.conf without duplicating receiver inventory.~~
  • [x] ~~Connect through the existing public demand ports.~~
  • [x] ~~Implement one independent controller and consumer-reference ledger per configured receiver, including simultaneous sessions on distinct receivers.~~
  • [x] ~~Automatically route manual display frequencies by compatible path, purpose, priority, availability, and configured power cost.~~
  • [x] ~~Implement visible route reasoning, expert receiver/path pinning, transition hysteresis, and target-before-release handoff.~~
  • [x] ~~Implement a primary receiver lane and compact secondary lane with independent receiver status, VFO, mute, and volume.~~
  • [x] ~~Implement WebSocket IQ and control framing.~~
  • [x] ~~Implement the initial WASM DSP chain.~~
  • [ ] Deliver AM, USB, LSB, CW, NFM, and WFM listening in an operator browser.
  • [x] ~~Deliver a polished spectrum, waterfall, VFO, filters, and manual controls, including collapsible side panels and adjustable visualization and waterfall height.~~
  • [x] ~~Replace coarse tuner-gain choices with Auto plus a slider that snaps to the supported R820T/R828D gain steps.~~
  • [x] ~~Add visible, pointer-friendly low/high passband handles above the waterfall with edge-resize cursors and drag-without-retune behavior.~~
  • [x] ~~Add Auto BW beside the filter control, backed by explicit low/high DSP cutoffs and synchronized with manual slider/edge changes.~~
  • [x] ~~Make frequency entry unit-aware and draft-based so values such as 100.7 tune 100.7 MHz immediately on Enter, with suffix, blur, Escape, and invalid-input behavior.~~
  • [x] ~~Add deterministic frequency profiles for sensible starting mode, passband, and tuning step while preserving sticky manual overrides.~~
  • [x] ~~Keep a display-only view center separate from the hardware RF center so zooming and tuning keep the VFO visible without unnecessary receiver retunes.~~
  • [x] ~~Add scoped mouse-wheel navigation: selected-step tuning over the left dB gutter, horizontal view panning with a preview-only ghost over the upper spectrum/dB field, and span zoom over the waterfall.~~
  • [x] ~~Carry horizontal navigation beyond the current capture with serialized hardware-center handoffs; keep the active VFO fixed while its complete passband remains capturable, move it only the minimum required at a capture edge, and show its frequency/direction when it is outside the zoomed view.~~
  • [x] ~~Give view span a direct validated bandwidth field and logarithmic slider so narrow scientific spans retain useful adjustment travel.~~
  • [x] ~~Add direct VFO dragging with visible 1x, 5x, 10x, and 100x precision divisors, where higher ratios tune more slowly for fine adjustment.~~
  • [x] ~~Give the clicked VFO selector keyboard focus and move it by the selected tuning step with Left/Right Arrow without scrolling the page.~~
  • [x] ~~Add bounded per-lane VFO Back/Forward history beside the primary readout, restoring the complete listening state across reloads and mapping standard mouse Back/Forward buttons to the same guarded navigation actions.~~
  • [x] ~~Add translucent, color-coded band-plan labels along the bottom of the spectrum, selecting label detail by visible span to remain readable.~~
  • [x] ~~Add validated, synchronized direct bandwidth and squelch fields beside their sliders, including a 25 kHz NFM range.~~
  • [x] ~~Add a nearby passband-edge lock, enabled by default, that preserves the passband center while resizing both edges together and permits independent movement when unlocked.~~
  • [x] ~~Arbitrate pointer and keyboard interactions so filter-edge dragging, VFO dragging, click tuning, wheel navigation, and band labels do not conflict; preserve the relevant state independently per receiver lane.~~
  • [x] ~~Enforce SDR++-style SSB filters: USB fixes the carrier-side low edge and exposes only the outer high edge; LSB fixes the carrier-side high edge and exposes only the outer low edge. Disable linked/free edge policy in SSB.~~
  • [x] ~~Preserve waterfall history across local VFO-following pan and zoom by remapping existing rows instead of clearing them.~~
  • [x] ~~Replace the view-sized waterfall bitmap with a capture-wide circular history that survives panel resizing, height changes, zoom, and retunes; zooming out must reveal previously captured outer frequencies.~~
  • [x] ~~Keep browser AGC disabled by default for each fresh receiver lane while retaining the explicit per-lane manual toggle.~~
  • [x] ~~Cover otherwise unmatched frequencies with deterministic fallback mode/bandwidth/step/span profiles, and replace stale step/span state on a cross-band or outside-view jump while preserving local manual navigation.~~
  • [x] ~~Extend fine-scrub precision to 1x, 5x, 10x, and 100x, with the highest divisor at the greatest upward drag distance.~~
  • [x] ~~Release each unused receiver after its configurable route grace and validate the existing clean demand idle shutdown.~~
  • [x] ~~Validate automatic HF/upper-range selection, two simultaneous receiver streams, boundary handoff, busy/fault isolation, and configuration-only addition of another route.~~
  • [x] ~~Update AGENTS.md and README.md with the validated gateway and operator workflow.~~
  • [ ] Validate browser control from the documented DNS origins with configuration-selected authentication enabled and disabled.
  • [x] ~~Validate the live browser IQ-to-WASM-to-FFT/audio path with visible per-lane IQ, DSP, FFT, audio, and fault telemetry.~~
  • [ ] Revalidate Phase 1 deliverables and update the completion ledger.

Phase 2 — Signal Detection

Realtime Pipeline Acceleration Candidate — 2026-08-30

  • [x] ~~Move live IQ ownership off the main thread and remove serialized per-frame stop/ack coupling between realtime, spectrum, analysis, candidate, and decoder consumers.~~
  • [x] ~~Move source-rate realtime and decoder/candidate numerical work into explicit scalar/SIMD C/WASM artifacts with bounded discontinuity-aware queues and shared PCM delivery.~~
  • [x] ~~Add exclusive WebGPU spectrum/waterfall presentation with a dedicated adaptive display FFT, chronological frame pacing, configurable refresh, WebGL2, large scalar/SIMD analysis WASM, and Canvas2D fallbacks. Keep 65K–4M scientific evidence independent while bounding display and detector outputs.~~
  • [x] ~~Extend selectable scientific FFT resolution through 2M and 4M without reducing the accepted default resolution or configured 250 kS/s–3.2 MS/s sample-rate inventory.~~
  • [x] ~~Pass production build, browser-WASM export/import inspection, native DSP, 248 browser regressions, installer checks, and the Pi throughput benchmark before deployment.~~
  • [ ] Live-accept the deployed acceleration candidate in Eco, Balanced, and Analysis modes. Exercise 2.4 and 3.2 MS/s, representative 65K–4M FFT tiers, WFM/FT8/HD paths, decoder start/stop, sample-rate changes, dual lanes, waterfall history/color/detail, spectrum/VFO stability, and uninterrupted mono/stereo audio. Record the actual WebGPU/WebGL2/Canvas fallback selected by the operator's client before closing this gate.

Phase 2 Correction Acceptance Contract

The revised detector overview and the intelligent-activation injection are one cumulative Phase 2 correction plan. Neither supersedes the other. Phase 2 is not complete until every applicable checklist item below is struck through and the operator has accepted the combined live behavior.

The required delivery threads are:

  1. Control: a visible per-lane Signal detection toggle, default on, with genuine detector-off CPU/work isolation and no effect on manual reception.
  2. Data integrity: acquisition metadata plus clipping, DC, I/Q mirror, dropped/repeated sample, gain-change, FFT-artifact, harmonic-comb, and configured-spur evidence.
  3. Scene formation: bounded multi-resolution time-frequency history, robust local background and variation estimates, CFAR-style seed/grow thresholds, and two-dimensional occupancy components.
  4. Parent tracking: band-aware envelope formation, narrow-to-wide maturation, pre-publication coalescence, stable center/bandwidth history, service-aware dwell, fades, recurrence, and identity compatibility.
  5. Conservative evidence: Phase 2 classification stops at artifact, noise-like, anomalous energy, or coherent/structured emission. Separate evidence and counterevidence replace uncalibrated percentage confidence.
  6. Stable presentation: one live corridor per parent emission, one historical rectangle per activity period, a non-reordering Signals rail, pinned inspection, and an optional diagnostic overlay for internal evidence.
  7. Activation groundwork: single-click only inspects; waterfall double-click and the pinned-measurement action share one normalized, cancellable, pin-aware, undoable intelligent-activation contract. Phase 2 applies only VFO/envelope changes supported by its evidence and exposes the remaining demodulator, gain, and decoder decisions for later phases.
  8. Performance and power: Eco, Balanced, Analysis, mobile, and detector-off policies protect audio and manual tuning, bound browser CPU/memory, perform expensive IQ analysis only on candidates, and add no Pi detector daemon or idle polling loop.
  9. Proof: synthetic and long real receiver captures validate probability of detection, false alarms per MHz-hour, parent over/under-segmentation, artifacts, bandwidth, identity, toggle isolation, UI stability, and live WFM, airband, and weak-signal behavior before commit.

Phase 3 must consume the activation contract to configure demodulator, filter, gain, AGC, squelch, de-emphasis, fine tracking, and other supported settings. Decoder phases must use the same contract for automatic protocol execution; FT8 is the first timing-sensitive proof case and later feeds decoded callsigns and Maidenhead locators to the shared map. These later capabilities are not claimed complete in Phase 2, but their identifiers, evidence, progress, handoff, cancel, pin, and undo interfaces must be laid down now.

  • [ ] Add a visible per-lane master Signal detection toggle that defaults on, persists an explicit local choice, performs no detector analysis while off, and leaves waterfall, audio, recording, and manual tuning operational.
  • [ ] Separate single-click inspection from activation and route waterfall double-click plus the pinned-measurement tune button through one normalized intelligent-activation command.
  • [ ] Define stable activation request, plan, progress, applied-change, warning/fault, decoder-handoff, setting-pin, cancel, and undo contracts. In Phase 2, apply only parent-center/envelope changes supported by detector evidence and expose unsupported decisions as pending later analysis.
  • [ ] Preserve acquisition metadata and implement inexpensive IQ-quality checks for clipping, DC, I/Q mirrors, repeated or missing blocks, abrupt whole-band gain changes, and relevant FFT/window artifacts.
  • [ ] Maintain bounded short, medium, and long time-frequency representations independently of waterfall rendering cadence.
  • [ ] Replace fast/slow-only floors with robust time- and frequency-local quantile/median and variation estimates, candidate exclusion, and CFAR-style seed and growth thresholds.
  • [ ] Build a bounded rolling two-dimensional occupancy field and join cells across frequency and time before creating a track.
  • [ ] Form one band-aware parent emission from compatible internal islands, using dual-threshold envelope growth, gap closing, and maximum-envelope priors without treating the band plan as proof.
  • [ ] Permit tentative narrow-to-wide maturation, coalesce overlapping parents before publication, and lock a nature fingerprint only after the envelope stabilizes.
  • [ ] Track parent emissions with robust center/bandwidth history, configurable evidence dwell, service-aware burst policy, fade hysteresis, and guarded compatible recurrence.
  • [ ] Apply center-bin, artifact, harmonic-comb, and receiver/path spur evidence before normal publication; retain diagnostic visibility without presenting an artifact as an independent user signal.
  • [ ] Measure center, 90%/99% occupied bandwidth, peak and integrated SNR, duration, duty cycle, drift, persistence, symmetry, and acquisition context in dBFS-backed snapshots.
  • [ ] Render exactly one stable live corridor per published parent and one scrolling historical rectangle per activity period; remove duplicate annotation passes and keep the Signals rail stable under updates.
  • [ ] Replace opaque percentage confidence with Phase 2 evidence classes limited to artifact, noise-like, anomalous energy, and coherent/structured emission, including separate artifact evidence, counterevidence, alternatives, and recommended next analysis.
  • [ ] Add an optional, default-off diagnostic overlay for background estimates, seed/grow thresholds, raw islands, parent grouping, artifact flags, and rejected candidates.
  • [ ] Retain Eco, Balanced, and Analysis performance policies while protecting manual tuning and audio; deeper raw-IQ features remain candidate-triggered rather than continuous whole-passband work.
  • [ ] Expand recorded tests with long real receiver-noise/artifact sessions and validate detection probability, false alarms per MHz-hour, parent over/under-segmentation, bandwidth error, track identity, toggle-off isolation, and live WFM/airband/weak-signal behavior.
  • [ ] Validate Phase 2 deliverables and update the completion ledger.

Deferred Phase 2 Closeout Backlog

On 2026-08-26 the operator approved moving into Phase 3 without representing the remaining Phase 2 proof work as complete. Preserve these items in the backlog: audit explicit candidate exclusion and the temporal-occupancy model; add activation and detector-off isolation regressions; exercise all IQ-quality flags end to end; collect long real receiver/noise/artifact captures and report false alarms per MHz-hour, bandwidth error, over/under-segmentation, and track identity; complete live airband and weak-signal checks; then audit and strike only the checklist entries supported by evidence. These items may be completed opportunistically when Phase 3 tests exercise the same paths, but they do not silently become complete merely because later work proceeds.

Phase 3 — Assisted Tuning

Phase 3A — Accepted Groundwork

  • [x] ~~Add candidate-specific raw-IQ autocorrelation, spectral kurtosis, flatness, coherence, timing, and bounded cyclostationary analysis.~~
  • [x] ~~Require multiple independent indicators before assigning a likely modulation family and preserve ranked unknown and artifact alternatives.~~
  • [x] ~~Implement broad AM, WFM/NFM, SSB, CW, structured-digital, and artifact heuristics with bounded initial acquisition and classification hysteresis.~~
  • [x] ~~Retain complete candidate evidence across same-track geometry jitter, fresh IQ segments, fades, and compatible recurrence without concatenating samples across a gap.~~
  • [x] ~~Display measured evidence, counterevidence, warnings, acquisition state, and ranked hypotheses without presenting scores as calibrated probabilities or treating a band-plan prior as signal evidence.~~
  • [x] ~~Separate nearby narrow emissions with scale-aware grouping and let an intermittent neighbor end without widening the persistent track.~~
  • [x] ~~Keep single-click selection passive and route waterfall double-click and the pinned-inspector action through one explicit activation entry.~~
  • [x] ~~Fit explicit activation to the published occupied edges in the current demodulator, including USB/LSB carrier-side references, with only a matching detector-resolution-aware guard; do not substitute BW90/BW99 estimates.~~
  • [x] ~~Create a lane-scoped activation request and pre-change snapshot, honor the existing frequency/bandwidth pins, show applied changes, and provide a one-step snapshot restore.~~
  • [x] ~~Quarantine hardware-settling IQ, measure quality in fixed sample-domain windows, and expose stable clipping/DC warnings suitable for later staged gain decisions.~~
  • [x] ~~Prime and rebuffer browser audio deterministically, expose continuity telemetry, and validate multirate DSP output across transport chunk sizes.~~

Phase 3B — Accepted Intelligent-Policy Work

  • [x] ~~Connect a stable candidate-analysis result to the existing activation entry and select the demodulator plus VFO/carrier reference; insufficient or conflicting evidence must leave the mode unchanged rather than force a guess.~~
  • [x] ~~Extend the accepted current-mode envelope fitter into a selected-mode plan for asymmetric/symmetric filter edges, occupied-bandwidth margin, tuning step, de-emphasis, squelch, AGC, audio, and other mode-specific settings, explaining every proposed change.~~
  • [x] ~~Implement staged tuner-gain optimization using clipping/headroom, local noise, SNR, settling, route ownership, and operator pins rather than a fixed or maximum-gain rule.~~
  • [x] ~~Score plausible alternative demodulators against measured evidence and post-change observations while retaining the alternatives and explicit counterevidence.~~
  • [x] ~~Make band-plan priors configurable and bounded; keep their contribution visibly separate from IQ-derived evidence.~~
  • [x] ~~Add explicit Auto/Manual state and make the complete policy honor mode, frequency, bandwidth, gain, AGC, squelch, and decoder pins.~~
  • [x] ~~Apply automatic changes as a cancellable, lane-scoped sequence with settling, bounded remeasurement, stale-track rejection, and sustained contradictory-evidence requirements before reclassification.~~
  • [x] ~~Add bounded fine-frequency tracking with mode-appropriate limits and hysteresis.~~
  • [x] ~~Produce a decoder recommendation/handoff descriptor without starting a protocol decoder; decoder execution remains Phase 6.~~
  • [x] ~~Harden immediate manual takeover, in-flight cancellation, partial-fault recovery, and one-step undo so asynchronous control cannot leave a mixed or stale plan applied.~~
  • [x] ~~Validate that single-click never retunes; double-click and the inspector action remain identical; Unknown is safe; every change is explained; pins, cancel, manual takeover, and undo are deterministic; and audio, waterfall, detector, and independent receiver lanes do not regress.~~
  • [x] ~~Validate all Phase 3 deliverables and update the completion ledger.~~

Phase 4 — Automatic Spectrum Scanning

  • [ ] Implement visible-window, band-sweep, and channel-list scan types.
  • [ ] Implement multi-band scan plans with include/exclude ranges, overlap, dwell, settle time, scan speed, thresholds, and receiver-path offsets.
  • [ ] Implement the explicit scan state machine and visible progress/status.
  • [ ] Partition scan windows across all eligible receiver routes, operate distinct receivers concurrently, and aggregate progress and results.
  • [ ] Rebalance work around busy/faulted receivers without stealing an operator-pinned manual receiver or duplicating coverage by default.
  • [ ] Implement stop-on-detection, temporary hold, continue, and manual-control transfer policies.
  • [ ] Let the user choose one or several detected signals from the current passband.
  • [ ] Implement per-signal actions for listen, inspect, decode, audio record, bounded IQ record, snapshot, bookmark, ignore, and continue.
  • [ ] After Phase 6 decoder jobs are available, connect compatible manifested decoders to scan actions through the shared decoder manager, admission control, result events, cancellation, and resource limits. Do not create a separate scan-only decoder path.
  • [ ] Implement the scan-results history and exportable metadata.
  • [ ] Enforce receiver control leases and existing-client protection.
  • [ ] Release each receiver independently when its assigned scan work and other consumer references are finished.
  • [ ] Validate cooperative multi-receiver scans plus simultaneous manual use on an independently available receiver.
  • [ ] Validate disconnect, cancellation, power-state, and idle-shutdown paths.
  • [ ] Update AGENTS.md and README.md with validated scan architecture and operator controls.
  • [ ] Validate Phase 4 deliverables and update the completion ledger.

Phase 8 — Learned Classifier

  • [ ] Capture and label real signals from both receivers.
  • [ ] Train a coarse modulation-family model.
  • [ ] Add an unknown/out-of-distribution class.
  • [ ] Export to ONNX.
  • [ ] Run candidate-only browser inference.
  • [ ] Compare it with the deterministic baseline before allowing it to control tuning or scan actions.
  • [ ] Validate Phase 8 deliverables and update the completion ledger.

Verification

Validation requirements own metrics, long-capture requirements and the live matrix. The retained CPU-efficiency and Viterbi measurements cover the separate native receive workload; they are not browser detector acceptance.

Rolling IQ History

The browser ingress worker owns a segmented raw-IQ timeline before DSP fan-out. Live and historical IQ use the same demodulation, spectrum, waterfall, detection, candidate and decoder paths, not buffered audio or rendered pixels. ADR 0008 records the decision; DVR architecture explains ownership, storage workers and bounded replay prefetch.

Keep the 60-second Hot RAM Minimum and 1 GB Hot RAM Target defaults. Requested duration takes precedence; there is no project-imposed RAM ceiling. Tiered storage adds six-hour total retention and a configurable 100 GB OPFS budget. All IQ storage belongs to the browser workstation, never the Pi; only small preferences enter localStorage. Explicit acquisition runs preserve sample/time, RF/tuner geometry, settings, continuity and UTC provenance; presentation source epochs reject stale output without rewriting captured evidence.

Pause/Play, seek, Go Live, the position slider, Space and tap/hold W/S review history while hardware remains live and read-only. Historical tuning must fit the complete passband inside the captured RF span. Default protect-history preserves the investigated interval, sacrificing newer live history if necessary; optional protect-live returns presentation to live. Go Live clears/restarts history only after newer history was deliberately sacrificed. See storage controls.

Waterfall history is a separate capture-time/RF-indexed derived cache. Preserve it across presentation resets, replace replayed rows, prune with raw IQ and clear on full session release. Restore without blanking/remounting; rebuild detection overlays from replay instead of projecting future evidence backward. Native Float32 for every emitted display bin remains the default. Reduction and quantization require explicit per-lane choices, never change raw IQ, and do not discard existing rows. Renderer geometry, RAM/OPFS sidecars and quality options are documented in waterfall visual history.

Account for IQ and visual payload separately and expose combined nominal-target overruns. Client memory readouts explain accounted allocations, unmeasured costs and advisory heap warnings; none is a measurement of total browser RAM, free system RAM or physical VRAM.

The original deployed IQ-timeshift candidate was operator-accepted on 2026-09-05. At that checkpoint, TypeScript, the production decoder/frontend build, all 276 browser tests, native DSP, gateway routing/IQ tests, appliance configuration, demand/power, health and repository secret checks passed. These are dated results, not counts or validation from this documentation edit.

The 2026-09-22 waterfall enhancement and 2026-09-24 tiered RAM/OPFS amendment are later work. The ledger still records the tiered candidate as awaiting live operator acceptance after its production install; focused raw-IQ seek and native-row tiering regressions are part of the frontend gate, not its closure. The completed checklist below records the original 2026-09-05 acceptance:

  • [x] ~~Live-accept pause without stopping or retuning the physical receiver.~~
  • [x] ~~Rewind across stable and changed acquisition segments, retune anywhere inside each captured RF span, and verify historical audio and waterfall.~~
  • [x] ~~Run detection, assisted analysis, and at least one decoder from retained IQ through the existing downstream path.~~
  • [x] ~~Verify 1x historical playback, single-step and held seek, and explicit Go Live semantics with and without sacrificed live history.~~
  • [x] ~~Exercise Hot RAM Target/minimum warnings, decimal/binary memory units, preference restoration, hardware-control isolation, and graceful behavior when whole-window memory utilization is unavailable.~~

Durable recording/export should use SigMF metadata where practical, but it is a separate future capability and must not turn browser timeshift into Pi disk wear.