02 — Preamble, synchronization & acquisition

Transmission is burst / PTT-keyed: every over is one keyup→keydown burst. After keyup the RF energy ramps to full amplitude over ~65–82 ms (keying/soundcard ramp) and then stays flat.

The BW500 symbol is 512 samples (10.667 ms, 93.75 baud) on two sub-bands at 1350/1650 Hz, which is what recovers frames byte-exact — see `01 §500` and `03 §3.6.1`. Several acquisition metrics below are computed on a 256-sample stride; that stride is an analysis window of half a symbol, not the symbol period.

There are two burst populations (see `01` §1.5): long DATA bursts (initiator) and short CONTROL/ACK bursts (responder, ~0.34–0.68 s, distinct centre-carrier-dominated structure, carrying the ARQ handshake / acquisition preamble — §2.5). DATA-burst onset shows a short (~6–10 symbol) transient with a distinct k=−3 head before the steady alternating-tone body; whether that head is a fixed SOF/training marker is not settled here.

2.2 Timing synchronization

2.3 Frequency synchronization

Signal centre is 1500.0 Hz (DC-null). Residual CFO relative to the estimated band centre runs ~8–20 Hz — a synchronizer quantity (carrier/soundcard offset), not a fixed waveform parameter.

2.4 Frame boundary detection

Frame = burst. ARQ cadence is very regular: initiator (A) DATA bursts run ~4.35 s (~785 sym) each (rising to ~8.53 s / ~1570 sym when more data is queued), responder (B) CONTROL/ACK bursts run ~0.34–0.68 s (~65–130 sym), and the TX↔RX turnaround gap is a near-constant ~85 ms between every keydown and the opposite station's next keyup.

2.5 Control / sync bursts (responder-sent; distinct waveform)

The responder's bursts are a separate waveform from the initiator's DATA bursts and carry the ARQ handshake / acquisition (connect, ACK, keepalive, disconnect). A 256-sample-stride FFT alternation metric (swing between the k=+1 and k=−1 bins each symbol; ≈1.0 = strict per-symbol tone alternation, ≈0 = steady) cleanly separates the two populations:

Control-burst modulation: raw DBPSK on the k=−1 (1350 Hz) sub-band. Under the BW500 front-end (down-convert 1350/1650, H=512), the k=−1 (1350 Hz) sub-band of a responder control burst carries a clean differential BPSK stream — constellation collapse |E[e^{2jθ}]| = 0.80–0.89 across 15+ bursts, against 0.15 for coherent BPSK and 0.28 for coherent QPSK. The differential is raw (cell[n]·conj(cell[n−1])), not the grid480-scrambled differential of the DATA path: a DATA over needs grid480 de-rotation first (raw diff ≈0.2) while the control burst does not (raw diff ≈0.9). For reference on the same metric, a DATA over scores 0.984 under grid480+BPSK and silence scores 0.0. The k=+1 (1650 Hz) sub-band is weaker and ambiguous (0.57–0.62); whether the burst also modulates it, or 1350 is the sole payload carrier, is not settled here.

The bursts are payload-independent — bit-identical whatever message the session carries — so they carry ARQ state (seq / ACK / NAK), not the message. The exact symbol timing/rate and any CRC are not specified here.

Raw DBPSK bits; the control burst is not turbo-coded. The raw-differential bit stream (sign(Re(cell[n]·conj(cell[n−1]))), best-onset) off the 1350 Hz sub-band is stable and repeatable — identical ARQ positions demodulate byte-for-byte identical, within and across sessions. Two frame lengths appear:

• 44-bit frame (the ~0.489 s class): e.g. 1111111111111111 0000011110000 1111 00000000000. Its bits are a fixed template; only positions 0 and 41–43 toggle between states (Hamming distance 4), the bulk (1–40) invariant. One distinct variant at higher Hamming (≈19) is a different control function (disconnect/idle class).
• 32-bit frame (the ~0.361 s data-ACK class): e.g. 01110111111101110000100001111000; likewise a template with a few varying positions.

The small bit-change per state is itself informative: a turbo-coded frame (as DATA uses) would avalanche ~½ the coded bits on any info change, so the control burst is light- or un-coded (plausibly repetition — note the 4–5-bit runs), fitting a tiny ACK that needs robustness over rate.

The control channel is a small vocabulary of fixed tokens, not parametric frames. Demodulating every responder burst shows no per-burst arithmetic field, only a handful of canned waveforms distinguished by duration and a few flag bits:

So the connected-ack and idle-keepalive are the same 40-bit template with a 4-bit tail flag (0000 vs 1111); the data-ACK is its own constant token.

The 0.49 s rows are read as DBPSK here and as two-tone MFSK in [`04 §4.2C`](04-frame-block-formats.md); the readings contradict each other and §4.2C is the one measured against the spectra. The two are the same burst: 44 x 512 samples is 11 x 2048, exactly. Every 2048-sample window of that burst holds two sharp equal carriers on the tone lattice, each with its own +/-1-bin Hann skirt and nothing between them — for pairs as far apart as bins 52 and 86, which no single 1350 Hz DBPSK carrier produces. The DBPSK templates recorded above are still usable as kestrel-to-kestrel tokens and are still what vara_control emits, but they are not a demodulation of what VARA keys at CONNECT setup, and the row's name should not be read as evidence about step 5. What the *idle-keepalive* row is, this document does not fix.

The construction is a canned token set rather than a coded frame with a sequence counter — consistent with the stop-and-wait ACK carrying no seq (§5.7). Interop needs only to emit and recognise this token set.

Both sides use fixed tokens. The initiator is the same story at longer length: its idle-keepalive is a 128-symbol (~1.385 s) constant token (bit- identical across three idle gaps in one session), with a ~65-symbol connect-confirm and distinct short bursts read at the time as disconnect requests. The responder answers those by reusing the `…1111` alive token — the same waveform serves keepalive-response and that answer, its meaning set by the initiator's state, not by a distinct burst. *What those bursts mean is now open: at BW2300 the burst of that length is the turn release and not a disconnect (`05 §5.6.1`), and this BW500 token corpus has not been re-read against that. Take the token set as measured and the names in this paragraph as provisional.* So the whole BW500 control channel is a small fixed-token vocabulary on both ends, no sequence arithmetic anywhere, matching the seq-less stop-and-wait ACK.

The NAK token. When the responder fails to decode an over — its input swamped with noise until decoding fails — it emits a distinct 32-symbol DBPSK token, 01111000111101111000100010001111: Hamming 0 across the failing bursts, Hamming 11 from the data-ACK, and appearing only while overs fail. Like the ACK it is a constant seqless token ("resend the outstanding over"). So the full BW500 responder set is: data-ACK, NAK, connected-ack, alive, ready — all fixed tokens.

BW2300 control channel — RETRACTED 2026-08-23. This section asserted until 2026-08-23 that BW2300's control channel is the same DBPSK-constant-token family as BW500, and gave four band-specific bit patterns for it. It is not, and they are not. No real VARA has been recorded keying a single-carrier DBPSK control token at BW2300, and the four patterns match nothing:

• On a 2026-08-14 bench capture of VARA HF v4.9.0 against itself at BW2300 — both directions on tape, with the host-port event log beside them, so every burst is attributable to a side and to a protocol event — the session's thirteen short control bursts match none of the four patterns at Hamming 2, and a bit search over the whole of both sides at sixteen independent column phases matches none of them either.
• Off air: zero token lines across all four 2026-08-23 monitor runs, both successful connects included. A bit search over offair/KC9GHZ_2300, offair/NS0A_2300 and the three 2026-07-24 confirmed QSOs — real VARA at both ends of every one — returns 1 to 2 hits per recording at ±60 Hz, which is exactly what the same search returns on material that holds no responder burst at all (see below), and nothing clustered, repeated or in a turnaround.
• The same detector on the same bench at BW500 finds more than forty data-ACKs at Hamming 0 (test_vara_control), so it is the table that is absent, not the instrument.

That floor is the whole of the original result. At the same settings — Hamming 2, sixteen column phases, ±60 Hz — the same search returns the same 1–2 hits per recording on material that cannot hold a BW2300 token: 2 in the 73.7 s BW2300 capture that carries only station A's DATA overs, 1 in 110 s of the BW500 control-burst session read with the BW2300 table, and 0 in 30 s of clear-channel noise. The counts this section used to cite — "9 data-ACK, 3 connected-ack, 2 alive and 2 NAK" over a whole session — are that same rate, and nothing in them separates a match from a chance one. How the patterns were arrived at cannot be rechecked: the 2026-07-23 session they were read off is in no corpus. The two-sided reference session is what settles it anyway, because there the answer is not "few" but "none".

What the burst actually is. The 0.489 s BW2300 control burst — which both stations key, not only the responder — is index-modulated, not DBPSK: eleven two-tone MFSK symbols of four 512-sample columns each in a 0.489 s keying, opening on the four fixed preamble pairs that `04 §4.2C` already specifies — {64,67} {56,74} {64,69} {68,78}, 1500/1570, 1312/1734, 1500/1617, 1594/1828 Hz — with seven state symbols behind them. Measured over all eleven instances in the reference session, on both sides, the preamble is identical in every one.

The 0.361 s class the old table called data-ACK/NAK is the same modulation one size down: 33 columns, eight two-tone symbols of four columns. It shares only the ack's first symbol (1500/1570 Hz) and none of the other three preamble pairs, and the two instances in the reference session — one from each side, at the turn exchange — share nothing past that first symbol. The durations the old table gave were right; the modulation it read them through was not.

This section used to carry a caveat beside the table — "not every 0.489 s BW2300 responder burst is a constant token", four instances demodding to three different DBPSK bit strings. That is what it was seeing: three different *state fields* read through the wrong modulation. §4.2C is the reading that reproduces, and the two sections were describing the same burst and contradicting each other about it.

The acknowledgement and the connected-ack are one waveform at BW2300. In the reference session the responder's CONNECT-time burst (b2a 8.567 s) and its first per-over answer (b2a 9.943 s) correlate 1.00. Three distinct state fields appear on the responder side of that session and two on the initiator's, so there is a small vocabulary here — but nothing held anywhere separates an ACK from a NAK among them. Settling that needs a two-sided BW2300 capture with an over deliberately corrupted, which is how the BW500 NAK was solved and what BW2300 has never had.

vara_control._TOKENS_2300 is gone, along with the band parameter that dispatched to it: BW500 is the only bandwidth with a token vocabulary, so there is nothing to select. What used to index it — vara_arq._send_token("nak") — is gone: stock ignored the BW500 NAK token keyed into a turnaround (2026-09-11), and a BW500 caller now asks behind the responder's idle instead. The reference session holds the BW500 vocabulary to the same standard, resolving none of its thirteen short bursts (test_this_vocabulary_names_nothing_in_a_real_vara_bw2300_session). The short-burst path at BW2300 reads vara_arq._ack_plateau instead, in the live receiver and in kestrel/arq/phy.detect_token alike.

Token synthesis in this implementation. BW500 only; there is no BW2300 token to synthesise. A real BW500 data-ACK is multi-carrier, with energy at ~1311/1500/1600 Hz. kestrel's token synth is single-carrier: it carries the correct differential bits (its own detector, and the demod that read them off real audio, score it at Hamming 0) and round-trips kestrel↔kestrel, but its *waveform* correlates only ~0.1 with a real token. This is not what VARA specifies; it is a stopgap until a waveform-faithful synth exists. Such a synth is achievable offline — reconstructing each constant token from its measured per-column OFDM cells (≈125 complex cells) hits correlation 0.965 against the real averaged token — but whether it is *needed* is an on-air question: if VARA's receiver demodulates the DBPSK bits (as kestrel's RX does), the bit-faithful single-carrier token already decodes and the multi-carrier waveform is surplus. It is not built, because it cannot be validated without a real VARA to accept it and may be unnecessary; the on-air vara_compat test resolves which synth is required.

What a control-burst demod needs: (1) a burst-sync / acquisition front-end keyed on the 1500 Hz centre-carrier energy — the DATA-burst stride-256 active/guard energy-concentration metric (§2.2) relies on the ±1 alternating comb and does not apply to these non-alternating bursts, so a distinct detector/template is required; (2) the same joint CFO+timing search can likely re-use the 1500 Hz centre reference; (3) once synced, the modulation riding the control burst (esp. the connect-req/resp, which must carry the callsigns + BW request + speed level) must be demodulated — this is the gate on reading the ARQ seq/ACK/NAK fields (`05 §5.7`). The level a receiver asks for rides the frame index of its answer (`04 §4.2C`, *Level requests*).

A DATA over does not open on a slice of the control waveform. Every level opens on its own lead, drawn from the session lead stream (`01 §1.1`, *Session lead stream*): the index levels' training preamble, the tone family's eight lead symbols, the OFDM family's four or five lead symbols. BW2300's leads are read cell for cell off the high speed capable tapes; BW500 L4's nine lead columns are counted (18 draws) but their cells are not read.

2.6 Open questions / provisional facts

• Blind acquisition is solved for DATA bursts without any preamble template: the stride-256 energy-concentration metric locks every burst, so no preamble detector is needed for demodulation. A dedicated preamble template is still wanted for fast, robust *acquisition* (esp. the CONTROL bursts).
• Onset marker: DATA bursts show a ~6–10-symbol transient (incl. a k=−3 tone excursion) before the steady alternating-tone body. Every DATA over opens on its level's lead, a per-over draw from the session lead stream (§2.5), so no fixed SOF symbol is expected there; whether the transient is that lead is not checked.
• CONTROL/ACK burst waveform is characterised at the envelope level (§2.5): responder-sent, centre-carrier-dominated, no k=±1 alternation (alt ≈0.2 vs DATA ≈0.99), ~0.36 s data-ACK up to ~1.7 s connect-request. Still open: its internal preamble/payload split. It needs a dedicated burst-sync front-end (the DATA energy-concentration lock does not apply).
• No CP: acquisition cannot use a CP Schmidl-Cox metric; the energy-concentration lock is the working substitute.