H-Mode
v2.

Design goals

1. Maximise legal character throughput inside one voice-width SSB channel (≤ ~2.8 kHz per bonded channel).
2. Expose ionospheric substitutions, erasures, and insertions — minimal payload FEC, by deliberate design.
3. Emit per-character soft confidence, LLRs, and per-resource channel state.
4. Remain fully open and regenerable from a public PCG64 seed — nothing obscures meaning.

Two-layer architecture

HFEdge v2 stream

Base36 payloads are framed in 64-character units with global ==Fxxxx== markers. The F0000 identification frame carries H2 + callsign(10) + profile + mode + CRC32(8) + grid6 and is excluded from scoring. Profiles: PING 720 / SHORT 3,200 / STANDARD 64,000 / DEEP 1,000,000 payload characters. A Hunt → Locked → Tracking → Lost state machine tolerates marker loss — two consecutive markers to lock, three missing to Lost, with continuous resync scanning. Scoring: CER, ERAS, 36×36 confusion matrix, diff-only RLE uplink (start:length:expected:got;…).

H-Mode v2 physical layer

Five selectable waveforms share one framing, acquisition, in-band signaling, and soft-output architecture: classic and generalized OFDM, OTFS, AFDM, and Hybrid. Runtime JSON profiles (classic / robust / balanced / high_speed) retune the modem without recompilation, and a frozen on-air registry negotiates frozen profiles by ID and fingerprint. TS 2.1 adds in-band configuration and loading-map descriptors, so a receiver learns the transmitter’s exact configuration over the air.

Waveform families

The data-aided technique

H-Mode’s distinguishing move. The sequence:

1. Acquisition: preamble A/B, an 18-bit CRC-protected header, and — new in TS 2.1 — in-band config and loading-map descriptors, so the receiver learns the exact waveform, grid, and bit-loading over the air.
2. Warmup: dedicated pilots for warmup_units (default 4) data units.
3. Steady state: pilot density falls to zero. The receiver regenerates the expected stream from the public seed and uses the payload itself as the reference, driving per-resource channel estimates, residual CFO/SCO/timing tracking, and iterative MMSE equalization.

Adaptation machinery

• Adaptive bit loading per resource (Chow / waterfilling class): bits ∈ {0,1,2,4} from per-resource SNR. Fixed-QPSK OTFS capacity is 53 characters per 72 ms frame — ≈736 cps before overhead; adaptive loading on strong paths is where the ≈300–500 cps good-SNR design target comes from, roughly 8–12× Olivia-class rates.
• Hierarchical packing: a QPSK core with enhancement bits packed after all core bits; the receiver always decodes them, so poor SNR appears as low LLRs, never dropped bits.
• Adaptive sync spacing: default every 16 data units, adaptive within configured bounds ([8, 24] on balanced); three consecutive marker failures return the receiver to Hunt.
• PAPR pipeline (clip+filter, tone reservation on the two top in-band bins, selective mapping) with Tukey windowing — crest control is an audio-PA concern, handled before the window.

Soft metrics contract

Each decoded character can produce a symbol, confidence value, six log-likelihood ratios, and resource index. Optional resource channel state includes linear SNR, channel magnitude, and phase. Confidence feeds the MatchMeter and upload metadata without schema changes — the PHY’s observables flow straight into the network dataset.

SoftChar { c, confidence, llr[6], ofdm_index }
ResourceCsi { snr_lin, magnitude, phase }

Sourcing and prior art

The design stands on published work: OTFS (Hadani et al.; HF OTFS evaluations 2020–2024), AFDM (Bemani), f-OFDM/WOLA windowing, STANAG 4539 / MIL-STD-188-110 soft-decision practice, data-aided estimation, adaptive loading literature, hierarchical QAM, PAPR literature, and Watterson/ITU-R HF channel models (delay spreads ~1–2 ms; Doppler ±2–4 Hz). The specification documents why each element was adopted — and, equally important, what was not adopted: payload convolutional/turbo FEC and deep interleaving, because hiding character errors would defeat the measurement mission.

Honest status

The classic profile is bit-exact against H-Mode-1.0. V2 families require V2 peers. Throughput targets of roughly 300–500 characters per second in favorable SNR are supported by simulation and loopback testing — including Watterson-channel tests — not by published on-air measurements. On-air figures will be published as network data accumulates.

TS 2.1 discloses its own open defects: high_speed ships selective-mapping PAPR and adaptive waterfilling with known end-to-end issues, and interactive balanced is exposed to a mid-stream map-update boundary bug — the frozen hybrid_v1 registry profile is the interoperable variant. Detection signatures and anomaly classifications remain hypotheses to validate against measured field data: co-located ionosondes, GNSS TEC maps, flare lists, Es climatology, and TID catalogues. The standard, not this brief, is normative.

H-Mode applies no encryption or undisclosed scrambling; the pilots, preamble tag, descriptor CRCs, profile registry, and character mappings are documented and reproducible from public state. The classic carrier plan spans 2,625 Hz between its lowest and highest active carrier centers; actual occupied bandwidth depends on filtering, drive level, hardware, and band settings — the operator owns the signal. The standard is an engineering specification, not a legal determination.