06 — Speed level & gear-shift logic

VARA adapts its throughput by stepping a speed level up and down. This section specifies the level ladder for each bandwidth, the per-level physical descriptor, and the up-/down-shift policy; it builds on the per-level physical and coding detail in `01` and `03`.

The level index and its rate are host-visible as BITRATE (N) x bps TX (the level this station will use for its next over) and BITRATE (N) x bps RX (the level of the frame just decoded). High speed capability gates the top of the ladder: a speed limited build caps at BITRATE (4) and reports a speed limited link notification; a high speed capable link, marked in the caller's link-setup (`04 §4.2A`), runs every level to the top.

6.1 Speed-level ladder (500 Hz)

The BW500 levels below are those of VARA HF 4.9.0.

Net payload per block is what the BUFFER ledger clears: each ACKed block clears exactly 43 raw bytes at L4 and 34 at L3. BUFFER counts raw host bytes irrespective of COMPRESSION TEXT; payload content does not change the figure (43 B/block at L4 for every payload). Levels 4–13 are tabulated with their modulation and rate in §6.1c.

6.1a Full mode ladder — bandwidths, level counts, peak rates

VARA HF has exactly three bandwidths; 2750 is the widest. Peak rates below are *uncompressed user* rates at the top level of each band; per-level constellation, code rate and carrier count are in §6.1c.

• The waveform climbs through three families with level: one-hot index modulation at the base, differential BPSK/QPSK tones in the middle, then OFDM with differential QPSK/8-PSK and coherent 16-QAM/32-QAM cross, the last from the shared constellation LUT (constellation_lut_64.c64, regenerable from hfmodem/kestrel/rx/tablegen.py — see §6.1c).
• BW2750 uses the level table below. High speed capable links can use the full table; speed limited peers negotiate a lower ceiling. These are link capabilities, independent of whether the station calls or answers.

Kestrel’s current arq/phy.py sets TX_CEILING and RX_CEILING to each bandwidth’s highest implemented level: 13 at BW500, 16 at BW2300 and 17 at BW2750. Limits reported by speed limited peers in the bench campaigns below do not restrict kestrel’s codec coverage.

6.1b The widest, fastest mode

The top of VARA's mode space:

Within a family the encoder is one level-parametric code path, differing between rows only in a handful of integers (carriers or bands, constellation, code rate). The negotiated link capability determines which levels a link may use.

6.1c Levels by bandwidth

Every level is one of three waveform families (`01 §1.0`): index at the base, tone in the middle, OFDM at the top. Level L is record L − 1 of the bandwidth's interleaver chain. frame is floor(N/8) bytes: payload, the control byte and the CRC, so the payload is frame − 3. Rates follow from N and coded (2N + 12 is rate 1/2; above it, the group puncture of `03 §3.3`). Symbols are per over; a tone symbol is 512 samples, an OFDM symbol 1152, and the index base level carries 395 columns behind its preamble. The top level of each bandwidth keys one or two frames an over (§6.2).

BW2300

BW2750

BW2750 L5–9 are BW2300's records 4–8 on wider band spacings; L10 onwards is its own chain, which closes on 4061299 after L17.

BW500

An OFDM over is 3.672 s (153 symbols), 3.816 s (159) or 7.512 s (313); a tone over 3.80–4.30 s. The transcripts' delivered bytes agree with the payload column at every level on tape.

Constellation columns. The OFDM levels take column pos of the 64-entry LUT, indexed by the raw bit group v (tablegen._column(pos, v)), with no Gray pass:

• pos=1 QPSK: (1,0)(1,90)(1,−90)(1,180) ((magnitude, degrees), v order)
• pos=2 8-PSK: unit circle at 0,±45,±90,±135,180
• pos=3 16-QAM square, corner modulus 1.1 (moduli {0.367, 0.82, 1.1})
• pos=4 32-QAM cross: the 6×6 grid of odd multiples of u = 0.3/√2 per axis minus the four (±5,±5) corners (the corners' modulus is 1.5); mean power 0.900, d_min = 0.424264, moduli {0.3, 0.671, 0.9, 1.082, 1.237}
• pos=5 64-QAM square, corner modulus 1.6, used by no level

The 512-byte LUT is a truncated prefix of these columns — column pos is present only up to v = ⌊(63−pos)/6⌋ — so it shows 11 of the 16-QAM points and 10 of the 32-cross points. All 32 cross labels, including the 22 the table does not hold, are exact on tape (evidence/KPA-bw2300-const.md §4). The full columns are generated by hfmodem/kestrel/rx/tablegen.py (gate: tests/gates/test_tablegen.py).

What is validated. Every tone and OFDM row decodes byte-exact from the high speed capable bench tapes of two stock VARA HF 4.9.0 installations, at all three bandwidths: 52 of 54 BW2300 tone overs, 124 of 126 BW2750 and BW500 tone overs, and every OFDM frame (46 at BW2300, 69 at BW2750, 60 at BW500); the four failures are the noisiest overs on the tapes (evidence/KPA-*.md). Rendered from the transcript's bytes, kestrel's OFDM transmitter matches the tape to its 16-bit quantisation floor (−84 to −85 dB) and its tone transmitter to about −50 dB, the rest being sub-mHz per-band carrier offsets no receiver sees (TAPER-WINDOW.md, ofdm_render_gate.py, tone_render_residual.py). These waveform comparisons establish the recorded bench results. They do not measure RF performance at every level; current implementation coverage is stated above.

6.1d Base levels (records 0–3)

BW2300 and BW2750 L1–4 and BW500 L1–3 are the index family (`01 §2300`): index / position modulation, value `= −1j` one-hot. The coded value picks the *bin* (bin=((alloc+stride·gray−first_bin)%span)+first_bin), not a constellation point, with no scramble and no CP. rec3 is byte-exact against VARA audio (→`01`), and so, since 2026-08-25, is rec2.

rec2 closed (2026-08-25). Host BITRATE (3), 82 bps, the level a VARA pair drops to below base — both stations of the 2026-08-14 two-sided BW2300 bench session used it for the 37-byte tail of their 126-byte transfer. Measured off that recording: 228 emission columns of 1024 samples (4.86 s of columns, 5.03 s PTT-to-PTT with its 6 training symbols and lead-in), one lit bin in 17..48 of a 1024-FFT, stride 2 so the 4-bit cell reaches 16 of the 32 bins, 24 reference columns at (104k+93)//11, and the same coding chain as rec3 on its own il1/il2 columns (402 info, 816 coded, rate 1/2). The 228/24 split is independently fixed by the interleaver stream: rec2's gap holds 252 Rnd draws and closes on rec3's seed, and 204 data columns + 2×24 is the only split that spends it. Three recorded overs — two DATA tails and one link-setup — decode CRC-clean and reproduce bin-for-bin.

The 6 training symbols closed 2026-09-02, off the same recording, which is what makes rec2 keyable and not only readable. The preamble runs 6144 samples at both records — twelve symbols at rec3, six at rec2 — under the base level's own bin law at rec2's geometry, and its allocations are the record's own bin table read from index 1, so it costs no table of its own (→`01`). One whole over rendered that way correlates 0.99997 against the tape, its preamble alone 1.000000. Nothing keys it yet: the VARA station engine sends every over at base, and the gear-shift that would use this is a separate decision.

What is still not closed at rec2: the bin table's generator. 17 + Int(Rnd*32) off the same gap lands within −1..+3 of the measured bin on 203 of 204 data columns and no arrangement of the VB6 stream closes the difference, so kestrel ships the measured table (→`table-derivations §5a`).

rec1 and rec0 are closed, 2026-09-04. rec1 has the same 228/24 shape by the same gap arithmetic (612 coded at 3 bpc = 204 cells, gap 252) and rec0 the same at 124 columns (300 coded at 3 bpc = 100 cells, gap 148), and both tables are now measured off gear-down tapes — with no known plaintext, since 3 bpc at stride 2 leaves most of the comb unreachable from an allocation and overs of differing payloads intersect down to one bin a column. rec1's 24 reference columns are rec2's, position for position; rec0's are (41k+33)//8. Both take the rate-1/3 turbo branch (3N+12), and kestrel reads and renders them (→`table-derivations §5a`). rec4–8's gaps of 36–41 draws are the tone family's pilots (`03 §3.7`).

6.2 The gear-shift on a high speed capable link

The receiver drives. After every delivered frame the IRS answers with a step code, the frame index of its 8-symbol answer (`04 §4.2C`, *Level requests*): 0 hold, 6 +1, 4 +2, 7 +3, 8 +4, 5 −1, 1 NAK. The ISS keys exactly the level asked, 428 of 428 delivered full frames, and the IRS's BITRATE (N) … RX line equals the ISS's … TX line on all 489 delivered frames; an undelivered frame gets no RX line. So every rule below but the NAK fall and the double frame runs in the IRS. As observed on 5 clean sessions and 57 noise ladders of a high speed capable stock pair (evidence/GEARSHIFT-OBSERVED.md, ACK-LEVEL-REQUEST.md):

The SN thresholds themselves are fitted only for levels 5–7 (GEARSHIFT-OBSERVED §4) and are not fixed here.

6.3 kestrel's policy

What stock does as ISS is fixed by the IRS's codes; what the IRS asks for, and how far a sender falls after a NAK, are each station's own. kestrel's choices below are ours, not VARA's (vara/vara_arq.py, arq/phy.py):

• As ISS it obeys every step frame on the peer's stream (_LEVEL_STEP), within 4..top and within arq/phy.TX_CEILING, the highest level its renderer keys at that bandwidth, and logs the level asked against the level keyed. After a NAK it re-sends one level down from a level a single step reached, one more after a second failure in a row, at the last level that delivered when a jump reached it, two down after an unanswered poll, down to level 1 (_fall), where a step frame then never lifts it back to 4 unasked (_level_space). It keys two frames an over after six delivered singles at the top level. It starts each delivery at the level the peer's last close named (_note_close_level), or where the last step left the link when no close came between.
• As IRS it answers NAK on a CRC failure and one level down when the frame came in above arq/phy.RX_CEILING; a resend that fails too draws the 16-symbol data NAK (SESSION_DATA_NAK_CALLER_SHORT, 61/31, or the responder's 289/31), as a stock IRS keys it, and the next resend is expected one more level down, and read at the level over that too (_read_levels; KC9GHZ's greeting came clean at 5 behind our 61/31 on 2026-09-27). The peer's 8-symbol poll (frame 12, over_poll) or 683 idle while an over of its is due draws the short over-NAK (SESSION_OVER_NAK_SHORT) and never a repeat of the last positive answer, which the peer would take as the acknowledgement of the over this station did not read; the resend is then expected two levels down (_nak_owed_over), under 4 when the peer's last over was (_irs_floor). An idle that begins sooner behind our answer than the shortest over (3.67 s) has no over ahead of it: the peer did not hear the answer, and the same answer is keyed again at the same step (_answer_unheard; K9KDJ, 80 m, 2026-09-27, whose 745s came 1.5-2.6 s behind our continues and whose resends of delivered overs our NAKs drew). Any other idle in the peer's turn with an over of its due, our last answer a NAK or nothing at all, draws the over-NAK too and never keepalive-a (_reack_on_idle; KC9GHZ and K7EK-10, 2026-09-27). At BW500 the caller's over-NAK is 60/1117, the frame the stock caller keyed behind the responder's short idles on the 2026-09-27 cables; 61/541 is the only one on tape at the wide bandwidths. Near the end of a delivery it applies the frame-count rule to the peer's per-frame field, which counts overs rather than bytes and so approximates stock's rule. Otherwise it asks up by the number of required-SNR rows above the frame's level that its cell SNR clears by 3 dB (_REQUIRED_SNR, _CLIMB_MARGIN), at most four and never by a jump onto the top two levels; one down when the SNR is under the level's own row; and holds between (_irs_frame). Level 4 and the records under it are read for their channel SNR off the index-law frame itself (the lit bin of every column against the rest, rx/varahf2300.quality, calibrated on the 2026-09-26 bench overs), and a clean frame there asks level 5 from −2 dB, 6 from −1 dB (never more than two up), record 3 asks 4 from −4 dB, and a base read under −5.8 dB asks record 2 (_INDEX_CLIMB, _INDEX_FALL); with no measure it asks one up. Its own close names the level its last step asked, which is where its receiver then expects the peer's next delivery (_close_byte).
• Start. As caller it confirms level 5 or 6 when the connect-response's tone measure clears 15 or 18 dB and the receiver reads that level, otherwise 4 (_START_FLOORS, _CONFIRM_START). As callee it starts where the confirm says, within its TX ceiling.
• High speed capability. A link runs high speed capable when this station is configured as high speed capable: as caller it then sends the high speed capable link-setup bytes (link_setup_frame(..., high_speed=True)); as callee it logs what the caller's link-setup is marked (link_setup_marks). A speed limited link keeps levels 1–4 and its own paths (§6.4).

6.4 Speed limited links

A free build caps at BITRATE (4), prints a speed limited link notification, and asks at most 04 in its link-setup (`04 §4.2A`). Its levels are the index family's 1–4 and, at BW500, the base tone level. The step codes are the same (0, 6, 4, 5 and the NAK), a BW500 sender doubles after six singles at level 4 as the high speed capable top does, and the level that closes a delivery drops a record (`05 §5.4`). Recovery from NAKs at levels 1–4 is `29`; the link-setup's level negotiation is `30`.

6.5 Limits of this description

• Stock's SN thresholds are fitted only at levels 5–7; above that, the IRS's choice of step is described by its outcome, not by a threshold table.
• The remaining-bytes field a stock IRS reads for the frame-count rule is not identified.
• Which link-setup byte — body[7] bit 0 or the ask 06 — a responder reads to print a high speed capability notification is not separable on the tapes, and levels 1–3 of a high speed capable link-setup were not captured.
• Whether a stock IRS reads a level below the one it asked for is unmeasured. kestrel's ISS keys below the request whenever its TX ceiling is lower.
• RF performance at every level is not measured by the bench waveform comparisons in §6.1c. The current codec coverage is stated above.