SF10/125 full samples: 10 Hz versus 100 Hz first-pass detour

2026-09-14. Both 400-packet samples completed: 100 attempts on each of four channels, with no RF retries or early stop on misses. Payload loss was 64/400 (16.0%) at 10 Hz and 81/400 (20.25%) at 100 Hz. Strict same-channel reception was almost identical: 230/400 versus 231/400. The larger offset showed no improvement in this comparison.

These are two sequential finite bench samples, not proof that the offset caused the difference in payload loss. The earlier stop-first tests could not provide this error-rate estimate and are not combined with these counts.

Settings and method

RX: XIAO ESP32-S3 / Wio SX1262, COM31, 28:84:85:B4:09:80.
TX: XIAO nRF52840 / Wio SX1262, COM15, B35E71C1C3726CE7.
Uniform SF10, BW125 kHz, nominal CR4/5; 909.5, 910.5, 911.5, 912.5 MHz.
32-symbol preamble; 5.1-symbol visits (41,780 us); zero added settling.
Normal RX gain 0x94, -9 dBm chip TX request, 16-byte CHS1 payloads.
Warm XOSC, bulk SPI, fast RX; TCXO setting 1600 us. No added IRQ tracing.
Two full retunes per hop. First: selected positive RF offset, CR4/6. Second: exact original RF command, CR4/5. Duplicate frequency writes remain inside each pass: four RF commands, two modulation writes and two optimized RX starts per completed hop.
Requested 10 Hz = +10 integer RF-word steps, nominal 9.5367431640625 Hz. Requested 100 Hz = +105 steps, nominal 100.13580322265625 Hz. Neither depends on a float-MHz addition. This is programmed, not independently spectrum-measured, frequency.
Same receiver image for both runs; only the requested offset changed. Order: 10 Hz, reboot, 100 Hz. TX firmware was unchanged.
Seed 606125: shuffled channel order each round and varied arrival pauses. Each run has 400 unique probe sequences and exactly 100 TXs per channel.
Existing 10-second expectation timeout retained. A miss does not reset the receiver, tune it toward TX, reset the visit clock or clear counters. Hardware/guard/command failures still abort rather than become RF losses.

The complete comparison ran from 21:09:11 to 21:42:39 UTC (14:09–14:42 PDT), about 33 minutes. The 10 Hz sample took about 15.4 minutes; 100 Hz about 18.0. The extra elapsed time is largely the additional ten-second timeouts.

Overall results

"Payload loss" means no intact expected payload was delivered within the expectation window, regardless of reported RX channel. "Strict error" also counts intact payloads received while the scanner reports another channel.

Requested offset	Attempts	Correct-channel delivery	Intact wrong-channel delivery	Timeouts / payload loss	Strict errors
10 Hz	400	230 (57.5%)	106 (26.5%)	64 (16.0%)	170 (42.5%)
100 Hz	400	231 (57.75%)	88 (22.0%)	81 (20.25%)	169 (42.25%)

All scored delivery failures were timeouts. The receiver also counted six packet-read errors at 10 Hz and eleven at 100 Hz. The error codes were not captured, so these cannot be identified specifically as CRC failures or mapped one-to-one to timeout probes. "Zero invalid payloads" in the summary does not mean there were no packet-read/CRC errors before timeout.

Per-channel results

Each cell uses exactly 100 attempts. Loss counts therefore also equal percent.

Center (MHz)	10 Hz payload loss	100 Hz payload loss	10 Hz strict errors	100 Hz strict errors
909.5	10%	11%	26%	25%
910.5	23%	23%	51%	45%
911.5	19%	26%	48%	55%
912.5	12%	21%	45%	44%

The extra 17 payload losses at 100 Hz occurred on channels 0 (+1), 2 (+7) and 3 (+9). Correct-channel deliveries differed by only one packet overall. Neither offset eliminated wrong-channel acquisition/delivery or timeouts. This comparison does not isolate the underlying RF failure mechanism.

Switching and signal measurements

Metric	10 Hz	100 Hz
Completed hops	16,992	20,676
Mean complete two-pass switch	1263.765 us	1262.551 us
Maximum complete switch	1533 us	1532 us
Mean first pass	660.712 us	659.458 us
Mean second pass	592.344 us	592.406 us
RF commands	67,968	82,704
Modulation commands / RX starts	33,984 / 33,984	41,352 / 41,352
Maximum unheld four-channel cycle	174,139 us	173,722 us
Maximum held channel visit	590,757 us	640,492 us
Packet-read errors	6	11

Unheld cycles fit within the 262,144 us programmed preamble; packet-related held visits can exceed it. Aggregate timing without IRQ traces does not prove which hold caused any particular timeout. The approximately 1.2 us difference in mean switching time is negligible here and does not explain the losses.

All completed hops passed the internal first-offset/final-correction and CR command checks: 16,992/16,992 and 20,676/20,676, detour.bad=0. Second-pass blocks, retune failures, mode errors and cache errors were zero. Counts match four RF writes, two modulation writes and two RX starts per completed hop.

Telemetry limitation: the diagnostic detour.first_rf/first_mod fields are updated on a guard-deferred first attempt as well as an applied attempt, while final_rf/final_mod retain the last completed second pass. Thus the final snapshot's displayed pair can be from different attempts and appears equal in these captures. Those fields must not be interpreted as the last verified pair. Verification itself runs only after two applied passes and increments detour.n only then; its per-hop comparisons, zero failures and matching counts establish the commanded sequence. Original capture bytes are retained.

Packet group	10 Hz mean RSSI / SNR	100 Hz mean RSSI / SNR
Correct-channel delivery	-34.05 dBm / +8.56 dB	-34.15 dBm / +8.62 dB
Intact wrong-channel delivery	-105.18 dBm / -13.75 dB	-105.22 dBm / -13.62 dB

The markedly weaker wrong-channel readings are observations, not proof of a specific filter/image/overload mechanism.

Implementation and verification

The HIL scanner gained scancontinue 1, default off, so its ordinary completion and timeout accounting can leave scanning active. The host gained bounded --continue-on-miss --max-channels 4 with separate delivery/strict summaries. scandetouroffset 10|100 and --first-pass-offset-hz 10|100 select the integer offset while stopped. Production radio defaults and receive guards were not changed. The controller is tools/hil/profile_switch_offset_compare.py.

23 native profile tests and 51 host tests passed. Coverage includes continuing through all 400 attempts without restart/retry, balanced channels, separated error definitions, default first-miss behavior, aborting fixture failures, both integer offsets and exact restoration. Build/upload succeeded with hash verification: RAM 94,948 bytes, reported flash 370,213 bytes.

There was one cached USB RX-result replay in each sample, with no RF retransmit. All TX commands returned success. Host/device strict receive and miss counters matched; both captures are complete under fixed_sample_complete. The master manifest is complete under both_400_packet_samples_complete, with no fixture or cleanup error.

Both boards rebooted idle afterward. RX confirmed scanner inactive, zero channels, one-pass default, detour/continue disabled and normal gain 0x94. TX confirmed no autonomous TX. HIL remains installed; RAK COM29 was not opened. The temporary PlatformIO overlay was removed. No background collector remains. This work made no commits or pushes. Existing unrelated changes were preserved.

Provenance

Manifest: tools/hil/profile_switch_sf10_10_vs_100hz_full_results.json SHA-256 40ef9d70991b585abe6b99b3dc4878829412d55e0e8b5a984576b00db0326981
10 Hz capture: tools/hil/profile_switch_sf10_10_vs_100hz_full_results_10hz.json SHA-256 889a7afeb980ec8459fccb8da5a7c496feebd433da8715f84181f02c1da613aa
100 Hz capture: tools/hil/profile_switch_sf10_10_vs_100hz_full_results_100hz.json SHA-256 0b54a057c07f9da0b03f585aeffee960ffcfa2ecb5383914554697ca250aa071
RX BIN SHA-256: bddf779208cbfd24478d39d3075f3f7537b7c9275453492748a349f3ec15b630
Unchanged TX ZIP SHA-256 from the preceding validated upload: 915d507d0feec1712fad2bd4be62c2fa8d769d11400711de0da1cf98a50a4b36

Source: live keymindCascade working tree at HEAD 2b5a6f209835f0f67ed73fee78d36447a4566ce6, plus existing dirty changes and these HIL additions. The RX image hash identifies the exact tested build. Earlier captures were not overwritten.