Skip to content

SF7/62.5 + SF8/500, fast switching and preamble 32

User-requested hardware validation, 2026-09-14. This is a finite HIL experiment, not a production reliability guarantee. No production timing defaults or automatic-preamble overrides are changed by this test.

Fixture and method

  • Heltec V4 receiver, USB serial 44:1B:F6:69:CF:98, normal chip RX gain (0x94).
  • RAK4631 9AB3B64C641BA927: stationary transmitter at 909.5 MHz, SF7/62.5 kHz.
  • T1000-E/LR1110 34A9141999729D5D: stationary transmitter at 910.5 MHz, SF8/500 kHz.
  • CR4/5, explicit header, CRC, 32-symbol preamble, synthetic 16-byte sequence-tagged payloads. -9 dBm chip output. Only one transmitter sends at a time on host request.
  • Each transmitter initializes its assigned settings once. RF-frequency and modulation-command counters must remain unchanged for every TX call.
  • 100 fixed-channel controls per profile before scanning. Require 100/100 on both before making a scan-timing inference.
  • Three scanner settings, 100 packets per profile per setting, randomized interleaving and 10–180 ms arrival delays. A miss does not cause a retry, retune, reset, or early end of the finite sample. The scan expectation expires after 1000 ms, far longer than the test frame and randomized host delay.
Case Loop reserve (not a delay) Slow visit Fast visit
4.1-chirp floor 0 ms 8.397 ms 23.171 ms
Spare time used for 5.1 slow chirps 0 ms 10.445 ms 21.123 ms
Previous reserve comparison 4 ms 8.397 ms 19.171 ms

The 5.1 case reallocates 2.048 ms from the fast visit to the slow visit; it does not increase the nominal cycle. The 4 ms reserve is omitted from the first two cases. The theoretical switch allowance is 0.6 ms per hop; actual switching is measured, never forced to equal that allowance. Changed SF/BW requires one modulation command per hop; frequency-only cache-hit timings do not apply.

The scheduler uses the same asymmetric two-profile visit budget and the actual production guarded tuneProfile() / fast RX implementation, but runs in a small RAM-only HIL application, without the full production networking/UI load. All retunes are single-pass, warm-XOSC, buffered 8 MHz SPI, zero added settling; double frequency writes, detours and rollback experiments are off. Packet/busy ownership guards are retained. Explicit 32 on both sides bypasses the historical automatic SF8/500 88-symbol rule without removing it from production.

The selected radios already had disposable HIL firmware. Updates are application-only: matching nRF52 SoftDevice IDs/address ranges and exact USB serial/topology are checked. V4 partition metadata is read back and compared before its application-only write. No bootloader, SoftDevice, settings backup, gateway firmware or unrelated radio is changed. Prior images/captures remain.

Results

Finished all 800 unique probes, 2026-09-15 00:04:54–00:11:19 UTC (September 14 local time). Both stationary controls passed 100/100. All three scans completed their full 100-packet sample per profile. No scan achieved 200/200.

Case SF7/62.5 received SF8/500 received Total Switch mean / max Idle cycle mean / max
4.1 chirps, no loop reserve 99/100 100/100 199/200 0.581 / 0.829 ms 32.717 / 34.733 ms
5.1 chirps, no loop reserve 100/100 97/100 197/200 0.572 / 0.834 ms 32.706 / 34.978 ms
4.1 chirps, 4 ms loop reserve 97/100 99/100 196/200 0.567 / 0.836 ms 28.701 / 31.195 ms

Switch timing includes every successful hop during each stage, including post-reception hops. The RX-only five-second windows averaged 0.554 ms per hop. The largest observed switch was 0.836 ms: 0.6 ms is a nominal allowance, not a measured worst-case bound. Direction-specific timing is in the validated capture. Idle-cycle samples exclude packet-held visits but still include USB command/loop scheduling; their occasional roughly 2 ms overshoots are real.

Every scan hop issued exactly one RF-frequency and one modulation command, with one optimized RX restart. There were zero added settling delays, double passes, frequency repeats, RF retries, USB result replays, device errors, retune failures, RX mode errors, cache errors or off-channel acceptances. All eight missed packets were sequence-matched RX timeouts with the receiver still in RX mode and no device error. Both transmitters acknowledged exactly 400 transmissions and zero frequency/modulation writes during every packet.

Successful scan packets had mean SNR approximately +12.0 to +12.1 dB on the slow profile and +12.8 dB on the fast profile. Mean RSSI was approximately -20.4 to -20.7 dBm and -18.5 to -18.9 dBm, respectively. These are strong-signal bench results; the controls passed, but weak-signal performance and possible strong-signal/front-end effects are not established by this test.

What the extra chirp changes

An extra SF7/62.5 symbol is 2.048 ms, equivalent to four SF8/500 symbols. Increasing the slow dwell from 4.1 to 5.1 preserves the full-sweep budget by shortening the fast dwell, but increases the fast profile's uninterrupted blind interval. With the nominal 0.6 ms hops, the 32-symbol fast preamble's remaining time after the slow visit plus two hops falls from 6.787 to 4.739 ms (13.26 to 9.26 fast symbols), before acquisition and scheduling overruns. This is a timing-budget calculation, not an experimentally proven acquisition threshold or a per-miss trace.

The observed 5.1 sample improved the slow result from 99 to 100, while the fast result fell from 100 to 97. This is consistent with a dwell/acquisition tradeoff; one finite sample per setting does not isolate the cause or establish a statistically reliable ranking. Simply fitting a sweep inside a preamble does not establish lossless acquisition. The 4 ms reserve is an allowance, not a sleep; retaining it also did not produce a lossless result here.

Using genuine spare time above the 4.1 floor remains a reasonable policy to investigate, but allocation must protect both profiles' acquisition/return budgets and measured switching/scheduling overruns. Do not automatically spend all of the reserve on longer slow visits, or treat the historic 88-symbol rule as disproven by this nearly-lossless sample. No production defaults or warning formula were changed by this hardware-test turn.

Capture integrity, deployment and cleanup

  • Raw capture: profile_pair_sf8_32_results.json, SHA-256 3472883ade89494b64657486efd1645a04b0a3bebfd0a2e31bbf914341ad792f.
  • Independent offline checks: profile_pair_sf8_32_validated.json. Checks all 800 unique sequences, 100/profile/phase, sender lifetime counts, actual RX frequency/SF/BW/payload identity, command counts, policies and cleanup.
  • Manifest records source and image hashes. Bundle SHA-256 d6cb9ff36be109af8906eddbbe8a4657ca793d1351ec4bf50ebbd4909eed2bba. Local bundle is out/pair-sf8-32-20260914; remote capture remains under /home/mikec/hwtest/runs/pair-sf8-32-20260914.
  • Initial RAK USB reset returned EPROTO before any firmware write. The preserved pre-write record has an empty writes list. Inspection confirmed the exact RAK serial/topology in its documented CDC-only bootloader PID 0x002A. Deployment was revised to accept that identity (and reset-time disconnect only after exact bootloader re-enumeration); no permission check or board-identity check was bypassed. The revision hash and all three successful application-only writes are saved in the deployment records. No bootloader or SoftDevice was replaced.
  • All three radios were rebooted and live-verified idle, with test firmware left installed. ModemManager was restored active; gateway and host CLI services remained active. The preexisting memory soak remains inactive. No unrelated radio was flashed. Wrapper exit was zero.
  • All three HIL firmware builds passed, run sequentially. The HIL host regression suite passed 63/63 tests, including the new plan and packet-policy tests. Python compile checks and scoped git diff --check passed. No commit or push.

Follow-up

The requested 4.6-chirp / 0.3 ms reserve follow-up completed 197/200 scanning packets: slow 97/100, fast 100/100, after fresh 100/100 controls on both profiles. Its captures are separate from this run.