V1's principal error was architectural. A symmetric magnetic stage was being asked to supply distortion that these recordings divide between a broadband overloaded analog path and a strongly frequency-dependent magnetic path. Increasing magnetic drive depleted LF fundamentals while leaving mid/high frequencies relatively clean and exposing peaks. Accurate numerical solutions did not validate that premise.
V2 selects an asymmetric rail-junction input circuit, a nonlinear flux circuit with weak asymmetric reluctance, and downstream AC coupling. CORE / IRON / MIX, fixed offline compensation and host infrastructure remain. This is a measured terminal model and an auditionable new hypothesis, not a recovered SSL schematic or a claim that listening approval has occurred.
The primary evidence is I:\Test\SSLDATA: 93 WAVs, 31 complete triplets,
all stereo 48 kHz. Seven levels at 20/50/100/1000 Hz accompany the complete
136-second mix, 16-second drum loop and half-second kick.
Manifest,
hash inventory, and
automatic pairs
record source/condition, format, length, peak and RMS. Nothing is normalized
before analysis. Sources remain immutable at the supplied location.
A is No Processing, B is SSL No Transformer, C is SSL With Transformer. A→B identifies the complete captured driven analog path, including conversion. B→C identifies the effect of the transformer switch within that path. C−A is never interpreted as an isolated transformer output. Absolute dBu and internal component geometry are not inferred.
Program alignment finds positive polarity, no identified channel swaps and essentially zero integer offsets. Fractional offsets are recorded per pair; the nearly mono kick makes channel identity ambiguous. Apparent long-record drift is under approximately 0.01 ppm and is not time-stretched away. Static least-squares gain is descriptive and kept separate from absolute measurements.
Independent sine files have unrelated starting phases. Absolute fundamental
phase shifts are not uniquely identifiable. Time-invariant harmonic phase
arg(Hn)−n*arg(H1) and fundamental-locked trajectories remain useful, but
discard fundamental phase and do not alone prove hysteresis. Invalid raw
complex subtraction was excluded rather than presented as a physical result.
Sine CSVs cover both channels, H2–H40 where below the measurement bandwidth, fundamental levels, gain/compression, odd/even ratio, THD, DC, asymmetry, crest and temporal stability. Near-noise harmonics are not mechanistic evidence. Program analysis includes spectra/deltas, K-weighted energy, envelopes, transients, short-time spectra, compression, stereo and residuals. K-energy is ungated mean-square, not certified LUFS. Fractional shifts can change sample peaks: candidate crest conclusions use original grids and separate reconstructed 8x true-peak estimates. Complete reference report.
B remains almost linear through −18 dBFS. At −12 it develops about 5% THD; at −3, about 32%, with output fundamental near −11.2 dBFS at every frequency. Predominantly odd harmonics and modest asymmetry support broadband headroom limiting, not magnetic flux saturation as the sole mechanism.
Harmonic phases support roughly 1.75–1.8 Hz AC coupling after overload. The H3-derived poles at strongly driven levels are about 1.80–1.81 Hz; H5 is broadly consistent with one weaker exception. Its location inside SSL versus the capture chain is not uniquely known. Phase evidence.
For the full mix, B lowers crest factor from 12.81 to 6.88 dB and RMS from −14.86 to −17.59 dBFS. It supplies substantial broadband peak limiting before the transformer contributes its distinct behavior.
At −3 dBFS, C's fundamental/THD is −25.02 dBFS / 142.37% at 20 Hz, −17.38 / 68.75% at 50 Hz, −12.53 / 16.98% at 100 Hz and −11.23 / 31.99% at 1 kHz. THD above 100% means harmonics exceed the remaining fundamental. At 50/100 Hz, H2 exceeds H3 at the lowest captured level; at 20 Hz H3 already exceeds H2 there. Onset moves substantially with frequency, so a blanket assertion that H2 always comes first would be wrong.
At 100 Hz, engagement reduces H3 from B's −10.74 to C's −30.60 dBc at the highest input, while H5 stays strong. Harmonic amplitudes and phases interact; distortion orders do not simply accumulate.
C increases the full mix's crest factor from B's 6.88 to 9.53 dB. It is not merely another compressor. Mix B→C energy falls about 4.05 dB at 20–60 Hz while 12–20 kHz rises about 1.15 dB; drums show a similar pattern. Subjective depth is therefore not explained by bass boost or blanket treble removal. The real capture includes severe LF saturation. Reserving that for high IRON is a product constraint, not grounds to reject the measurement.
The frozen engine was rendered on every RAW sine at 15 IRON positions and all three cores (1,260 renders), and every complete program file at all 45 settings (135 renders). No manual matching was required.
At Nickel 85%, −3 dBFS / 1 kHz produces only 0.0264% THD and a −0.85 dBFS fundamental. Simultaneously, 20 Hz produces 222.90% and −31.16 dBFS. SSL C instead gives roughly 32% at 1 kHz and 142% at 20 Hz. One magnetic drive cannot put both regions in the correct regime.
The full mix has V1 crest factor 17.77 dB versus C's 9.53 dB. After a separate RMS match, V1 still has 6.25 dB less 20–80 Hz energy and 2.26 dB more 8–20 kHz energy. Compensation scales relatively unaffected highs while LF fundamentals collapse. This explains depleted weight and exposed peaks measurably. V1's nominal symmetry also misses the observed even harmonics.
All sine solves converged. Interior coherent alias probes put intermediate Nickel settings around −120 dBc, although extreme drive does alias (about −57 dBc for the 7 kHz probe, improved at 16x). Aliasing and backward-Euler damping are concerns but do not explain the principal intermediate mismatch. V1 has actual magnetic history; calling it literally memoryless would be false. Full audit, sweeps, residuals and renders.
Retained: the production C++ bridge, oversampling/FIR framework, fixed latency, reset/state discipline, independent channels, smoothing, exact dry endpoints, state IDs, UI and fixed offline compensation principle. The V1 source and binary are frozen with hashes. Shared JA code and TUBE/BUS are unchanged.
Replaced: the isolated symmetric JA-only premise, extreme voltage allocation and unsupported material-specific claims for current CORE profiles. The production nonlinear system now combines input limiting with an inverse flux law. Trapezoidal integration replaces the previous winding network's backward Euler. Default 44.1/48 kHz oversampling rises to 16x after alias comparisons.
Nine variants were built and rendered: input-only; JA, reversible flux and stop/play flux with coupling before or after magnetics; and two weakly asymmetric cubic-flux variants. These independently test missing input overload, magnetic model class, minor-loop state, coupling placement and weak asymmetry.
The selected output_biased_flux uses a steep positive-reluctance knee and a weak shifted cubic. The latter supplies early asymmetric harmonics without requiring permanent output DC. The lower scalar fitting cost did not win: the input-coupled biased model had poorer program residuals and an undesirable small-signal LF pole. Stop/play state gave no consistent held-out advantage. No supply-sag, viscosity, thermal or excess-loss state was added without discriminating evidence.
This is a rail circuit feeding a flux differential equation and capacitor state, not static waveshaping followed by EQ. MODEL.md gives the equations, coefficients, drive mapping and units. Research documents primary sources, including SSL's hardware guide, TI's rail/impedance macromodel guidance and coupled magnetic modeling.
These laboratory results are capture-referred, before product autogain. All 28 sine fundamentals have 0.100 dB RMS error, 0.231 dB maximum. For levels ≥−18 dBFS and reference orders above −60 dBc, H3 error is 1.27 dB RMS and H5 2.53 dB. Across all H2 points above −60 dBc, error is still 11.09 dB RMS. That poor match is not hidden by the compression score.
| Held-out source, common aligned support | RMS error | Mean absolute band error | Residual relative to C | Estimated true-crest error |
|---|---|---|---|---|
| Kick | +0.020 dB | 0.951 dB | −22.97 dB | −1.49 dB |
| Drum loop | +0.131 dB | 0.343 dB | −17.91 dB | −1.18 dB |
| Full mix | −0.026 dB | 0.237 dB | −19.03 dB | +0.82 dB |
The selected circuit has the lowest absolute waveform residual on all three held-out sources among measured candidates. Some alternatives win individual spectral/crest dimensions. Residuals include phase, filtering, noise and nonlinearity together; they are not isolated distortion. Program scores | Sine scores.
The candidate residual table excludes a 32-sample interpolation boundary; true-crest comparisons use the original complete signals. A separate complete kick onset audit retains sample zero using explicitly assumed zero prehistory. The selected model remains best over the whole kick (−22.77 dB residual), but ranks only fifth over the first 5 ms and 5–30 ms among the eight C-target candidates. Its first-5-ms residual is −16.11 dB. Some JA variants reproduce that earliest attack more closely. The final production analysis also supplies uncropped anchors: its standard 64-sample valid-support crop would otherwise omit the hardware kick's peak at sample 61. This limitation is not concealed by comparing a cropped model crest with an uncropped reference crest.
The final production engine was also measured on complete original durations, with no sample-peak resampling or onset crop. These capture-referred results use Nickel 80% before its fixed compensation:
| Complete source | RMS error | 20–80 Hz band error | 8–20 kHz band error | Sample-crest error |
|---|---|---|---|---|
| Kick | +0.024 dB | −0.027 dB | +1.035 dB | −1.358 dB |
| Drum loop | +0.131 dB | +0.004 dB | +0.693 dB | −1.413 dB |
| Full mix | −0.026 dB | −0.052 dB | +0.681 dB | +0.951 dB |
Uncropped measurements and whole-duration aligned residuals retain the early kick peak. Residual RMS is 7.09%, 12.37% and 11.08% of C for kick, drums and mix respectively. The corresponding product output is exactly 0.79761 dB louder through fixed compensation; crest is unchanged.
At the difficult 20 Hz points, V2's capture-referred THD is 3.37% versus 1.13% at −30 dBFS, and H2 is approximately −54.52 dBc versus −27.51 dBc at −3 dBFS. Good fundamental matching plainly does not resolve all onset and asymmetry errors.
IRON 80% anchors the capture-referred circuit before compensation. Lower IRON uses different continuous input/magnetic excitation schedules so input density arrives before severe LF collapse. Nickel anchors the measurements; Alloy and Steel provisionally increase flux headroom by 15% and 30%. They no longer claim independently validated JA material identities.
Autogain uses synthetic references only and depends solely on CORE/IRON. It cannot reconstruct lost fundamentals or erase compression. The compensated product is deliberately louder than the captured raw circuit at its matching setting. Production comparisons explicitly separate this fixed scalar.
IRON automation uses a separate, disclosed control convention: falling excitation contracts stored states; rising excitation preserves them. Earlier rules produced large transient overshoots, including energy injection when flux was rescaled upward. The selected one-sided contraction avoids that injection and prevents falling gain divisors from amplifying stored voltage. It adds damping during knob movement and can dissipate history under repeated modulation. It never responds to input envelopes or clears signal history. This is a product behavior, not a mechanism inferred from the SSL recordings.
The artifact index supplies commands and links to raw CSVs, plots, hashes, calibration, solver tests and listening renders. Final production sweeps include both compensated output and a clearly labeled uncompensated diagnostic at the reference setting. Build/test results and remaining numerical limitations are recorded alongside those measurements.
The new production build passes all four CTest gates and pluginval strictness 10 (GUI tests skipped). The 756 supplied-source sine cases have zero failed flux solves. Worst THD at IRON 0%/20% is 0.00808%/0.06680%; clean-range gain is approximately −0.048 to +0.003 dB. The 28-point uncompensated production anchor reproduces the laboratory fundamental result (0.1002 dB RMS error). The sine fixed-compensation identity holds to 2.39×10⁻¹⁵ maximum sample error. These checks establish reproducibility and behavior, not listening approval.
The control fix was independently checked against frozen pre-fix bridges: all 756 supplied-source sine cases and six complete program comparisons (product and uncompensated paths) are bit-identical, with maximum sample error zero. Thus the fixed-control measurements and calibrated gain table remain applicable. Exact equivalence proof. Another 96 cases verify the measurement build against the final delivered build, including dynamic controls and irregular block partitions, with all 384 comparisons bit-identical. Build parity. Final dynamic probes were rerun on the revised control logic. Earlier source-change warnings and failed intermediate controls remain preserved, with their relationship to the final build explained in production findings.
The expanded IRON-only tests hold CORE/MIX/bypass fixed: 144 phase-varied transitions have at most 0.498 dB peak excess over the louder settled endpoint; 72 additional transitions including DC offset reach 0.767 dB with no failed solves. These are measured bounds for those probes, not a click-audibility claim. The failed intermediate rules remain identified in the evidence.
At 7 kHz/−3 dBFS, the final 16x path measures alias ratios of approximately −121.6, −79.4 and −64.5 dBc at IRON 50/80/100%. The extreme result remains a limitation. Cross-rate comparisons use shared harmonic bandwidth: a 44.1 kHz measurement otherwise excludes H3 of 7 kHz while 48 kHz includes it. The corrected convergence analysis reports the distinction instead of interpreting that bandwidth change as DSP instability.
The actual VST3 host output agrees with the production engine within 1.49×10⁻⁸ in the float-path host test. The 16x stereo DSP regression timed approximately 0.55 CPU seconds per audio second on this workstation while other validation jobs were active. This is a disclosed performance sample, not a DAW track-count guarantee or a worst-case benchmark.
Old dist packages remain V1 history; they are not relabeled V2. Old sessions retain their controls but necessarily sound different with the redesigned DSP.
