How we found
JG25’s balance.

How to read the experiments

Three kinds of evidence, three different jobs.

A waveform is the pattern of sounds sent over the radio. The demodulator is the receiver software that turns those sounds back into data. Changing one can change which design looks best.

Keep the units separate. A raw bit error is a wrong bit before packet validation. Packet delivery asks whether the whole checked packet arrived. “Signal-to-noise ratio,” or SNR, compares signal strength with noise; more dB means more signal relative to noise. Early sample-based SNR and later audio-band SNR use different definitions, so their numbers cannot be put on one scale.

The sequence · August–October 2026

The design evolved in eight stages.

Scroll the table sideways to see every column. Keyboard: focus the table, then use the arrow keys.

Evidence and test conditions · PDF p. 4

01 · The first tradeoff

Carry more bits without rushing every symbol.

Bell 202, the two-tone reference used here, sends one bit per symbol at 1,200 symbols per second. Four choices can carry two bits per symbol: 00, 01, 11, or 10. At 800 symbols per second, that already gives 1,600 raw bit/s. Continuous phase means the carrier’s cycle stays continuous when the tone changes.

1. 00Lowest tone
2. 01Next tone
3. 11Next tone
4. 10Highest tone

Schematic bit mapping; actual tone frequencies depend on the mode.

The earliest fast attempt doubled the raw rate but made substantially more mistakes. Slowing the symbols helped. This comparison holds the four tone centers fixed at 550, 1,250, 1,950, and 2,650 Hz.

White noise is only part of a voice-radio channel. Before selecting a mode, the team had to ask what filtering and distortion would do to it.

Scroll the table sideways to see every column. Keyboard: focus the table, then use the arrow keys.

The engineering explanation is plausible, but separate. More tones carry more bits per symbol. Fitting them into the same audio range forces a trade among spacing, symbol duration, and bandwidth. That helps explain the difficulty; it does not supply the missing higher-tone experiment.

Evidence and test conditions · PDF p. 8

02 · Fit the sound to the radio

Four tones still have to fit through a narrow doorway.

A voice radio does not pass every audio frequency equally. Imagine a doorway that is narrowest at its edges: pushing the outer tones too far out can make them weaker or more distorted. A clean-noise winner could therefore lose once realistic audio filtering was included.

Scroll the table sideways to see every column. Keyboard: focus the table, then use the arrow keys.

Chat simulation · −2 dB sample-domain SNR · 18,000 symbols × 3 runs per point · original 550/1,250/1,950/2,650-Hz four-tone plan, before later tone optimization. These are bit errors, not whole-packet losses.

The search varied center frequency and spacing for each speed. In one 1,200-symbol/s sweep, 350-Hz spacing produced 10.110% raw bit errors, versus 1.337% at 700 Hz and 1.548% at 750 Hz. Close packing hurt this detector, while ever-wider spacing was not automatically better. Different runs used different setups, so these precise BER values should not be mixed with the earlier chart.

Scroll the table sideways to see every column. Keyboard: focus the table, then use the arrow keys.

Four retained vectors trace directly to printed search winners. JG2000’s vector was also printed, but its retention was incompletely explained. The final changes to JG2200 and JG2400 were not recovered. Later comparisons help explain why those plans stayed; they do not reconstruct the missing original decisions.

Tone span is not occupied bandwidth. Modulation spreads energy beyond the outer tone centers. Later rendered-audio measurements placed 99% of the power in roughly 2.47–2.53 kHz for the three slowest modes, and up to 3.17 kHz for JG2400. That is audio bandwidth, not an FM transmitter’s occupied radio-frequency bandwidth.

Evidence and test conditions · PDF p. 13

03 · Help the receiver read the introduction

The message cannot arrive if its instructions are lost.

A packet starts with a known preamble and a synchronization pattern. They help the receiver find the beginning, estimate frequency offset, and line up the symbol timing. Then a small header says which speed and error-protection setting the message body uses.

Early tests could find the preamble but lose that header. The answer was to slow and protect the short header, rather than slow every bit of the body. The 350-symbol/s four-tone choice was the best tested airtime-versus-reliability compromise in this contest.

Scroll the table sideways to see every column. Keyboard: focus the table, then use the arrow keys.

The selected opening has 48 preamble symbols in NORMAL or 96 in ROBUST, plus 20 sync symbols, all at 600 symbols/s. ROBUST adds 80 milliseconds. Including the 38-symbol header at 350 symbols/s, the opening takes about 222 or 302 milliseconds before the body, excluding radio keying delays.

Under one simulated burst-interference condition at −6 dB, complete acquisition rose from 19.6% with NORMAL to 57.1% with ROBUST. That is a different test from the 96.7% isolated-header result above: finding the whole opening remained harder than decoding a header whose position was already known.

Zero false acquisitions in 320 simulated no-packet recordings was a bounded test, not a zero-false-alarm guarantee. The receiver later learned to recognize either acquisition profile even when the station’s own transmit setting differed.

Evidence and test conditions · PDF p. 19

04 · Spend airtime where it earns delivery

Redundancy helps only if the receiver can use it.

Forward error correction (FEC) adds structured redundancy so the receiver can repair some damaged bits. Interleaving rearranges the coded bits before transmission, spreading a short burst of damage across the decoder’s work. The receiver also keeps confidence information instead of immediately treating every tone decision as certain.

Scroll the table sideways to see every column. Keyboard: focus the table, then use the arrow keys.

The underlying error-correcting code is shared; F1 and F2 omit selected coded bits to reduce overhead. The selected interleavers use 8 rows for the header and 32 for coded bodies. Tests also compared 64-, 128-, 256-, and 512-byte bodies. Direct trials made 256 bytes a useful historical compromise; an early extrapolation favoring 512 bytes was too optimistic.

The later production-code comparisons qualify the early “more coding always helps” intuition. Redundancy takes airtime, and receiver behavior matters. A small run with no failures does not prove a setting can never fail.

Evidence and test conditions · PDF p. 22

05 · August–September 2026 · Repair what the radio changes

First restore the balance. Then preserve the timing.

An FT-8800-to-FT-817 recording path showed roughly −4.3 to −4.4 dB of receive tilt per octave: higher tones were progressively weaker. Software emphasis boosted them back toward balance. Replaying selected recordings recovered 8/8 JG2000-F1, 8/8 JG2000-F2, and 12/12 JG2400-F1 frames that had failed without conditioning.

But two filters can make tones equally loud while delaying their components differently. That delay can smear one symbol into the next. A controlled comparison held the +6 dB/octave magnitude correction nearly identical and changed its phase realization—the timing behavior of the filter.

The receiver also stopped treating fast symbols as four independent energy contests. At JG2400’s rate and spacing, the tones are not perfectly separable inside one short symbol window. A phase-aware detector followed the carrier’s continuous cycle across symbols, keeping information that the simpler detector discarded.

Scroll the table sideways to see every column. Keyboard: focus the table, then use the arrow keys.

Its 48 states track carrier phase; they are not 48 transmitted tones. A continuous synchronizer could also find a packet inside a recording opened by noise or a transmit transient. Missing audio, runtime queues, and keying defects were fixed separately.

The selection lesson: a failed decode may reveal a receiver defect rather than a bad transmitted waveform. These replay populations are different tests, and their totals should not be added into one reception rate.

Evidence and test conditions · PDF p. 26

06 · September 2026 · Make the comparison fair

Do the new tones still win after the receiver is fixed?

Field failures reopened the tone question. Lower tones suffered more relative delay on the fitted radio path, so moving them upward sometimes helped the earlier receiver substantially. The team then compared the candidates using the repaired receiver, including channel compensation and phase recovery.

A delivery threshold is the signal-to-noise ratio where the model reaches a target—in this case, 90% packet delivery. Lower is better because less signal is needed. These simulations use fitted radio channels and long F0 frames; they are not calibrated RF-input sensitivity measurements.

Scroll the table sideways to see every column. Keyboard: focus the table, then use the arrow keys.

JG1800 and JG2400’s selected plans won all three comparisons against both candidates. JG2200 was close, with no decisive improvement established by these rounded figures. JG2000 was the clear exception: candidate B still needed about 1.7–2.4 dB less signal.

Evidence and test conditions · PDF p. 31

07 · September–October 2026 · Measure the resulting family

A faster clean-link default, with a real sensitivity cost.

The later reference sweep used the actual modem transmit and receive code (C++), matched packets and noise seeds, and seven FM channel models. It gives a more useful comparison than ranking raw rates alone. Its waveform payload rate includes acquisition, header, checks, and coding, but excludes retries, keying setup, turnarounds, acknowledgements, and application envelopes.

Scroll the table sideways to see every column. Keyboard: focus the table, then use the arrow keys.

Production-code simulation · 256-byte payloads · thresholds are medians across seven channel models at 90% delivery. Lower threshold is better. The median of seven thresholds is not a delivery guarantee on an individual channel. The complete F0/F1/F2/F3 matrices are in the PDF.

F3 improved the threshold at every speed, while slowing delivery. The F0 sensitivity ladder was not monotonic: choosing a lower mode number did not always improve the modeled threshold with this receiver. There was no single “best” setting for every path.

Scroll the table sideways to see every column. Keyboard: focus the table, then use the arrow keys.

On these three models, JG2400/F0 needed roughly 3.4–3.9 dB more signal than AX1200 at 90% delivery. F3 closed much of that gap. The practical achievement was a wider operating menu and a faster strong-link default, not a universal sensitivity victory.

These reference channels use FM-shaped noise measured in the 300–3,000-Hz audio band, with a documented exception for the double-de-emphasis model. They omit the measured transmit third harmonic. Thresholds were interpolated from finite packet counts; one radio pair underlies most fits. These figures are not sensitivity readings in dBm.

Evidence and test conditions · PDF p. 35

08 · The resulting design

The parts were selected together.

Scroll the table sideways to see every column. Keyboard: focus the table, then use the arrow keys.

Some receiver features and defaults arrived after the original waveform search. The record explains why the system survived repeated challenges, while leaving genuine gaps: the original higher-tone results, the final JG2200/JG2400 tone changes, and over-air validation of the promising JG2000 retune.

Automatic rate ladders, compact multi-frame bursts, and short inter-frame markers were explored in the research campaign. They should not be described as current automatic behavior. The implemented format uses a full opening and header for each frame, with JSQ service reliability above the modem.

Evidence and test conditions · PDF p. 40

Read further · Review dated 23 October 2026

The evidence remains part of the story.

This account adapts the reviewed report from application snapshot 1cd39dea4441. No radio trials or sensitivity sweeps were rerun to write it. The report’s corrections table explains disagreements in the recovered histories, including tone-count wording, rate arithmetic, and swapped comparison labels.

• Review corrections and unresolved evidence · PDF p. 42
• Source notes and methods · PDF p. 44
• Published JG25 specification — a separate published snapshot covering FM and HF.
• Website evidence notes — earlier decoder results and waveform rate versus file throughput.
• JG25 on the modem page — the operator-facing description.