Greybound Lab

lab/ is the offline scientific workspace for Greybound. It is intentionally separate from the real-time Rust engine: the lab can use slower analysis tools, large generated files, SPICE renders, NAM references, plots, and training artifacts. The runtime crates should only consume artifacts after they have been reviewed, tested, and frozen.

Why It Exists

Greybound is moving toward a stronger gray-box modeling workflow. That means we need evidence before changing models:

• render rigs in a reproducible way,
• compare renders against references,
• measure latency, gain, spectrum, envelope, and residual error,
• keep metadata for every run,
• later generate SPICE datasets and fitted micro-models.

The lab is not a product runtime. It is where we search, measure, and decide.

Tooling Boundary

The lab is Python-first because the R&D work needs the scientific audio ecosystem: numpy, scipy, plotting, optimization, neural tooling, and eventually SPICE automation.

Rust remains the target for accepted runtime work:

• deterministic inference,
• bounded latency,
• no Python dependency in the live path,
• golden tests for frozen artifacts,
• explicit model descriptors and controls.

In short: Python for research, Rust for the engine.

Neural Cell Strategy

The current neural-cell decision is deliberately split:

PyTorch trains.
Greybound exports.
Rust runs.
ONNX verifies.

PyTorch is the default training and R&D environment because it gives the lab the strongest scientific workflow for fitting, plotting, debugging, and comparing small circuit-cell models. The live engine should not run PyTorch.

Accepted cells should be exported as versioned Greybound artifacts:

• model.greybound.json for architecture, controls, state, normalization, provenance, validation metrics, and runtime requirements,
• weights.greybound.bin for packed numeric weights.

The Rust runtime should implement only the tiny set of operations we accept for real-time audio: dense layers, small activation functions, causal convolution or explicit state update when justified, normalization, control conditioning, and safety clamps. This keeps the audio path deterministic, bounded, and free from generic graph-runtime behavior.

ONNX may still be exported for inspection, compatibility tests, or external runtime comparison. It is not the source of truth for Greybound's live runtime.

Important uncertainty: this is a current engineering decision, not a proven benchmark result. It may change after the first full loop if SPICE quality, PyTorch export stability, Rust CPU cost, or validation metrics contradict the assumption. The detailed plan and decision gates live in lab/experiments/006-spice-to-neural-cell-plan.md.

Current Workflow

The first implemented loop is complete-chain WAV analysis.

1. Render a Greybound rig to a WAV file and metadata JSON.
2. Compare that WAV against a reference WAV.
3. Optionally provide segment markers for local diagnostics.
4. Generate synthetic stimuli when the metric needs controlled input.
5. Run SPICE fixtures for bounded circuit-cell references.
6. Import NAM renders as integration references.
7. Read the generated Markdown report to decide what to investigate next.

This is deliberately before NAM integration or neural training. Without a stable comparison loop, training would produce numbers without engineering meaning.

Setup

From the repository root:

uv --project lab sync --dev
uv --project lab run pytest

The lab commands are exposed through greybound-lab.

The neural-cell work starts with spice-dataset, then later adds PyTorch training and Greybound artifact export.

Render A Rig

Use render-rig to call the release greybound-cli, write a WAV into lab/renders/, and write provenance metadata next to it:

uv --project lab run greybound-lab render-rig \
  --rig rigs/nox30-driven.json5 \
  --input-wav "lab/references/tone3000-inputs/Brit - Guitar.wav" \
  --output-wav lab/renders/nox30-driven.wav \
  --metadata lab/renders/nox30-driven.run.json \
  --render-seconds 10 \
  --sample-rate 48000 \
  --period-size 16 \
  --output-db -18 \
  --ir lab/references/tone3000-irs/celestion.wav

The metadata file records:

• git revision,
• rig path,
• exact render command,
• input WAV,
• sample rate,
• render duration,
• input and output gain,
• IR enabled state,
• local environment summary.

Sweep A Rig Against NAM

Use sweep-rig-vs-reference to generate rig variants in memory, pipe each rig to greybound-cli --rig -, render WAVs, and compare them against an already rendered NAM reference WAV.

The current NAM comparison protocol is amp-head only: do not add an IR to the NAM render, and do not pass --ir to Greybound during the sweep.

uv --project lab run greybound-lab sweep-rig-vs-reference \
  --rig rigs/nox30-driven.json5 \
  --sweep volume=0.64,0.76,0.88 \
  --sweep drive=0.68,0.80,0.92 \
  --sweep sag=0.55,0.70 \
  --input-wav "lab/references/tone3000-inputs/Brit - Guitar.wav" \
  --reference-wav lab/reports/nam-diagnostics-ac30hwh-topboost-gain5-brit-noir.wav \
  --output-dir lab/reports/sweeps/nox30-volume-drive-sag-vs-topboost-gain5 \
  --report lab/reports/nox30-volume-drive-sag-sweep-vs-nam-topboost-gain5.md \
  --metadata lab/reports/nox30-volume-drive-sag-sweep-vs-nam-topboost-gain5.run.json \
  --render-seconds 10 \
  --sample-rate 48000 \
  --period-size 16 \
  --output-db -12

The sweep report ranks points with a composite diagnostic score. The score keeps log-spectral distance important, but also penalizes weak null residual, envelope mismatch, and large gain correction. This is intentionally not an objective tone score: if a control feels musically reactive, keep that listening evidence. The score is only a guardrail against selecting a static NAM anchor from one metric.

The first Nox30 grid sweep against TopBoost-Gain5 stays near volume = 0.760, drive = 0.800, and sag = 0.550. Treat that as a coarse anchor, not as a final calibration: the null residual and envelope metrics still indicate dynamic mismatch that needs a broader sweep or targeted greybox work.

A segmented follow-up compared the best composite point, best spectral point, and current driven-style point against the same NAM render. The working measurement anchor is now volume = 0.640, drive = 0.800, sag = 0.700. This keeps the musically reactive drive = 0.800 region while favoring null and envelope behavior over a purely spectral fit. See lab/reports/nox30-anchor-segmented-summary-vs-nam-topboost-gain5.md.

Generate Stimuli

Some metrics are not reliable on a musical DI because the input does not isolate the behavior being measured. Use generate-stimuli to create controlled WAVs and matching marker files:

uv --project lab run greybound-lab generate-stimuli \
  --output-dir lab/stimuli \
  --sample-rate 48000

The generated set currently includes:

• sine-level-sweep.wav: harmonic distortion and level-dependent behavior.
• two-tone-imd.wav: intermodulation-oriented input windows.
• aliasing-stress.wav: high-frequency sine and sweep stress for nonlinear aliasing triage.
• sag-bursts.wav: repeated low-frequency bursts for supply/compression recovery behavior.
• pluck-attacks.wav: synthetic plucks for attack timing and overshoot.

Generated stimuli are ignored by git. The generator code is the source of truth.

Run SPICE Fixtures

Use spice-run to execute supported ngspice fixtures and import their output into the lab:

uv --project lab run greybound-lab spice-run \
  --fixture common-cathode-12ax7 \
  --output-dir lab/references/spice

The first supported fixture is common-cathode-12ax7. It produces:

• copied SPICE wrdata output,
• a Markdown report with DC operating point,
• settled 1 kHz transient gain metrics.

This is the bridge from full-rig observations to cell-level validation. The first report gives the common-cathode stage a reproducible electrical anchor: plate around 250.544 V, cathode around 0.402 V, B+ around 277.322 V, and small-signal plate gain around 14.88x.

Use spice-dataset to package that fixture as a local dataset artifact:

uv --project lab run greybound-lab spice-dataset \
  --fixture common-cathode-12ax7 \
  --output-dir lab/datasets/spice

The command writes:

• common-cathode-12ax7.dataset.npz,
• common-cathode-12ax7.dataset.json.

This is the first small multi-stimulus corpus. It runs generated SPICE netlists for several 1 kHz sine levels, two-tone IMD cases, first burst/decay dynamic probes, and a deliberately hard bias-recovery stress probe. It writes raw traces, packs a .npz, and records hashes, node roles, train/validation/test splits, component values, generated netlists, and operating point. Its purpose is to make the SPICE-to-training contract executable before we expand to source/load impedance sweeps, B+ perturbation, component tolerance sweeps, stronger transient probes, and real DI windows.

Train the current experimental MLP artifact with:

uv --project lab run --with torch greybound-lab train-neural-cell \
  --cell common-cathode-12ax7-mlp \
  --dataset-manifest lab/datasets/spice/common-cathode-12ax7.dataset.json \
  --output-dir lab/models/common-cathode-12ax7-mlp-current \
  --epochs 1200 \
  --hidden-size 32 \
  --learning-rate 0.0005 \
  --stride 8

The command writes:

• model.greybound.json,
• weights.greybound.bin,
• training-report.md.

This model is still a static cell, but the current working artifact is no longer only an export smoke test. It maps normalized input_v to normalized plate_ac_v through a small MLP and proves the PyTorch-to-Greybound artifact path, Rust inference path, and explicit Nox30 first-stage insertion path. It still does not model tube memory, capacitance, source/load interaction, or B+ perturbation.

The Rust core now has an experimental neural_cell loader for this artifact shape. It can parse model.greybound.json, read weights.greybound.bin, and run deterministic scalar MLP inference with tanh activations. Nox30 can use it explicitly in shadow or replace mode for R&D. For subjective local listening, Nox30 also loads lab/models/common-cathode-12ax7-mlp-current/model.greybound.json by default when that artifact exists and inserts it in replace mode. Use --disable-neural-cell for a purely analytic render.

Export Python reference vectors and verify the local artifact through Rust with:

uv --project lab run greybound-lab export-neural-cell-vectors \
  --descriptor lab/models/common-cathode-12ax7-mlp-current/model.greybound.json \
  --output lab/models/common-cathode-12ax7-mlp-current/equivalence-vectors.json

make lab-check-neural-cell-rust

The generated vectors are local artifacts. The Rust test is optional during normal test runs and becomes active when GREYBOUND_NEURAL_CELL_DESCRIPTOR and GREYBOUND_NEURAL_CELL_VECTORS point to the local files.

The Rust side now has two layers:

• ExperimentalNeuralCell: descriptor/weight loader and scalar reference path,
• NeuralCellRuntime: preallocated streaming runtime intended for future audio integration.

The generated-vector test runs through NeuralCellRuntime, so the integration candidate is checked against Python-exported golden values without allocating per sample.

Nox30 has an explicit first-stage neural path. The CLI form is:

target/release/greybound-cli \
  --rig rigs/nox30-nam-anchor.json5 \
  --input-wav "lab/references/tone3000-inputs/Brit - Guitar.wav" \
  --output-wav target/greybound-nox30-monitor.wav \
  --render-seconds 20 \
  --sample-rate 48000 \
  --period-size 16 \
  --monitor \
  --neural-cell nox30.first_stage=lab/models/common-cathode-12ax7-mlp-current/model.greybound.json \
  --neural-cell-mode shadow

shadow runs the neural adapter beside the analytic first_stage; monitor telemetry reports shadow first abs err avg/max, and the sound still comes from the analytic stage. replace feeds the neural output into the rest of Nox30. That mode is intentionally explicit and remains an R&D diagnostic, not an accepted model-quality gate.

The convenience command is:

make lab-shadow-nox30-first-stage

This is an integration diagnostic only. Promotion to audio replacement still requires better model-quality and dynamic-state evidence.

The full first integrated network loop is:

make lab-evaluate-integrated-neural-cell

It renders three versions of the same offline rig:

• analytic.wav: current Nox30 path only.
• shadow.wav: neural nox30.first_stage runs beside the analytic first stage and emits monitor error.
• replace.wav: neural nox30.first_stage feeds the rest of the amp.

The command writes its report to lab/reports/integrated-neural-first-stage-anchor-current.md. It uses rigs/nox30-nam-anchor.json5, no IR, lab/segments/guitar-chords.markers.json, and the NAM TopBoost-Gain5 render as external reference. This is the first full-chain gate for a neural component: shadow telemetry gives local component error in volts, replace-vs-analytic audio metrics show the complete audible/runtime impact, and NAM comparison shows whether replacement moves the full chain toward or away from the external oracle.

Current integrated result:

• shadow first-stage average error: about 0.137 V,
• replace-vs-analytic null residual: about -24.7 dB,
• replace-vs-analytic log-spectral distance: about 3.71 dB,
• analytic-vs-NAM log-spectral distance: about 12.43 dB,
• replace-vs-NAM log-spectral distance: about 12.68 dB,
• weighted NAM score: 0.5308 analytic versus 0.5335 neural replace,
• program-material NAM log-spectral distance after preroll: 10.87 dB analytic versus 11.10 dB neural replace,
• program-material weighted NAM score: 0.4956 analytic versus 0.4978 neural replace.

The neural path is therefore integrated and measurable, but not promoted. The working artifact is now much closer to the analytic chain after high-amplitude coverage was added, but it does not improve the NAM-facing weighted score. Segment deltas show the largest apparent regression in the quiet opening preroll; after excluding preroll the NAM metric still moves slightly away, but by a smaller amount. Promotion still needs local and external-reference evidence to agree. The promotion rule is NAM-first: replace should beat analytic on the weighted NAM score, while replace-vs-analytic remains a stability guardrail rather than the source of truth.

An offline neural blend sweep confirms this decision:

make lab-sweep-neural-blend

The command blends analytic.wav and replace.wav for several alpha values and scores each result against NAM with the same weighted score. Current result: alpha=0.000 is best globally with score 0.5308, and alpha=0.000 is also best after excluding preroll with score 0.4956. Therefore no partial blend of the current first-stage neural cell improves the NAM objective. The next useful work is not a blend control; it is a better neural target, richer cell representation, or a full-chain objective tied to NAM.

Evaluate the artifact against the SPICE dataset in physical units:

uv --project lab run greybound-lab evaluate-neural-cell \
  --descriptor lab/models/common-cathode-12ax7-mlp-current/model.greybound.json \
  --dataset-manifest lab/datasets/spice/common-cathode-12ax7.dataset.json \
  --report lab/models/common-cathode-12ax7-mlp-current/spice-evaluation.md \
  --stride 16

The current report evaluates 13 stimuli and 38,763 decimated samples. Adding a 400 mV sine to the training split and a 300 mV sine to validation changes the interpretation materially: the previous bias-recovery failure was mostly high-amplitude domain coverage, not proven memory. Weighted RMSE is now about 49.6 mV; the bias_recovery_probe_20mv_after_400mv test is about 67.8 mV instead of the earlier 1.77 V. The history-probe gain delta remains near zero, which means this fixture exposes domain coverage more strongly than low-level post-stress gain memory.

Compare the existing Rust analytic common-cathode stage against the same SPICE dataset with:

make lab-evaluate-analytic-common-cathode NEURAL_STRIDE=32

The static MLP is now in the same range as the analytic baseline on the expanded SPICE dataset and is much closer to the analytic Nox30 chain in replacement mode. The integrated Nox30 result still does not justify promotion because it moves the NAM log-spectral metric in the wrong direction. The immediate research target is now better external alignment: decide whether neural or fitted components can reduce held-out SPICE error and full-chain NAM distance without creating a large replace-vs-analytic residual.

The analytic evaluator also reports a diagnostic residual after a small integer-latency search and optimal linear gain. The current weighted residual only moves from about 80 mV to about 70 mV, so the mismatch is not explained mostly by gain or latency. That points the next R&D loop toward model shape: nonlinear transfer, dynamic biasing, solver discretization, and exact fixture equivalence. The best-lag columns are diagnostic only; periodic sine stimuli can produce phase-equivalent lags and negative gains that should not be interpreted as physical circuit latency.

The same evaluator now reports level-normalized harmonic and IMD shape. The first result is deliberately conservative: THD is within about 0.2 dB across the sine sweep, and the hotter two-tone IMD case is within about 0.1 dB. Therefore the remaining time-domain residual is not obviously a gross static transfer error. The next cell-level investigation should isolate dynamic state: cathode bypass memory, supply movement, phase behavior, operating-point trajectory, or exact SPICE/Rust fixture equivalence.

NAM References

NAM is used as an integration oracle, not as Greybound's internal architecture. The preferred protocol is:

• use NAM A2 captures only,
• find a VOX AC30-family Amp Head capture,
• render the same dry DI through NAM,
• do not add an IR to the NAM render,
• compare against Greybound rendered with cab/IR disabled.

This keeps the comparison focused on the amp core. Speaker/cab IR matching is a separate validation axis and should not be mixed into the primary NAM amp-reference score. A full-rig NAM is still acceptable as a broad end-to-end sanity check, but the report must be marked as cab/mic-confounded.

Current candidate source:

• First candidate: https://www.tone3000.com/tones/ac30hwh-6580
• TONE3000 VOX AC30 category: https://www.tone3000.com/categories/makes/VOX%2BAC30
• Filter target: gear Amp Head, platform NAM, architecture A2, clean or edge-of-breakup Top Boost style capture.

For AC30HWH-6580, the public page exposes useful capture semantics in the model names and description: Normal Bright, Top Boost, and Hot Mode variants, gain positions 3, 5, 7, or Full, optional TopCut, Top Boost treble and bass at noon, and Top Cut at 6/10 when enabled. This is useful for experiment selection, but it is not a complete machine-readable knob schema.

Metadata for imported references should follow lab/schemas/nam-reference.schema.json. Local NAM renders and downloaded model files belong under lab/references/nam/ and are ignored by git.

For a manually downloaded pack, write a source-safe manifest with:

make lab-inspect-nam-pack

The current manifest is lab/references/nam/manifests/ac30hwh-6580.json. It records the 22 local NAM files, their architecture/sample-rate/training metadata, parsed capture semantics, and the four priority models for the first comparison pass.

NAM rendering is handled by a wrapper around an external A2 renderer:

make lab-render-nam \
  NAM_MODEL=lab/references/nam/AC30HWH/TopBoost-Gain5.nam \
  NAM_INPUT_DB=-70 \
  NAM_OUTPUT_DB=-12

The wrapper writes the output WAV and run metadata in the same shape as Greybound renders. The default adapter uses the official Python neural-amp-modeler package in a temporary Python 3.11 uv run environment, selects the highest-quality A2 submodel, and keeps NAM inference outside the runtime engine.

External DI Inputs

TONE3000 exposes public input audio in the neural-amp-modeler-wasm repository. These are useful because they are already meant to drive NAM-style amp comparisons: short mono WAV examples with guitar and bass playing styles.

Download them into the local lab with:

uv --project lab run greybound-lab download-tone3000-inputs \
  --output-dir lab/references/tone3000-inputs

The command downloads WAV files from:

• https://github.com/tone-3000/neural-amp-modeler-wasm/tree/main/ui/public/inputs

It also writes a local manifest.json with the original GitHub URL, raw download URL, SHA, size, and local filename for every sample. The generated WAVs and manifest stay ignored by git.

Important rights boundary:

• the upstream repository is public and its code is MIT-licensed,
• the input WAV files are contributed audio samples,
• Greybound treats them as local R&D references only,
• do not redistribute the downloaded WAVs from our repository unless the sample rights are explicit.

External IR References

TONE3000 also exposes public impulse-response WAV files in the same neural-amp-modeler-wasm repository. These are useful as quick cab and reverb references when a NAM capture is an amp-head model and needs an external IR.

Download them with:

uv --project lab run greybound-lab download-tone3000-irs \
  --output-dir lab/references/tone3000-irs

The command downloads WAV files from:

• https://github.com/tone-3000/neural-amp-modeler-wasm/tree/main/ui/public/irs

It writes the same local manifest.json structure used by the DI downloader. The generated WAVs and manifest stay ignored by git.

Use this as a reference set, not as Greybound's canonical shipped cabinet library. Many commercial and free IR packs are licensed for end-user use but not redistribution, and some require account or mailing-list access. Those should be imported manually into lab/references/ and described by metadata rather than scripted as project assets.

Compare WAV Files

Use compare-wav to compare a candidate WAV against a reference WAV:

uv --project lab run greybound-lab compare-wav \
  --candidate lab/renders/nox30-driven.wav \
  --reference lab/references/nox30-reference.wav \
  --metadata lab/renders/nox30-driven.run.json \
  --segments lab/segments/guitar-chords.markers.json \
  --report lab/reports/nox30-driven-vs-reference.md

The current report includes:

• sample-rate and duration information,
• estimated candidate latency,
• gain correction,
• RMS, peak, and crest factor,
• null residual after alignment and gain correction,
• log-spectral distance,
• envelope error,
• optional segment-level local gain, residual, spectrum, and envelope metrics,
• band-local residual across low, low-mid, mid, presence, and air ranges,
• attack peak timing, rise timing, and overshoot deltas,
• two-tone intermodulation diagnostics,
• high-band residual checks for aliasing triage,
• sag drop and recovery deltas on dynamic windows.

These metrics are diagnostic, not a single quality score. They should tell us whether the difference is mostly alignment, gain, spectral shape, transient behavior, or nonlinear dynamics.

Metric Families

Latency and gain alignment:

: Makes comparisons fair. A model can look wrong simply because it is shifted by a few samples or louder than the reference. The lab estimates candidate latency and applies an optimal gain before residual metrics.

RMS, peak, and crest factor:

: Checks gain staging and dynamic shape. Crest factor is especially useful for spotting over-compression, missing transient energy, or unexpected clipping.

Null residual:

: Measures what remains after latency and gain correction. It is sensitive and useful for regressions, but it should not be treated as a complete perceptual score.

Band residual:

: Splits the residual by broad musical ranges: low, low-mid, mid, presence, and air. This tells us where to investigate. A mid-band error points toward gain stages, tone networks, or speaker color; a presence/air error may point toward cab/IR, anti-aliasing, or high-frequency nonlinear behavior.

Log-spectral distance:

: Tracks broad EQ, harmonic balance, cab/IR color, and frequency-dependent differences. It is useful for finding tonal drift, but it can hide transient and dynamic errors.

Envelope error:

: Tracks amplitude shape over time. It helps identify compression, sustain, decay, tremolo/modulation depth, and slow dynamic mismatch.

Attack diagnostics:

: Compare peak timing, rise timing, and overshoot in short windows. This is important for pick feel, immediacy, and whether a model blunts or exaggerates the front of notes.

Harmonic diagnostics:

: Compare THD and H2-H5 balance on stable sine windows. This is useful for diode clippers, triodes, overdrives, fuzzes, and any nonlinear stage where even/odd harmonic structure matters.

High-band / aliasing diagnostics:

: Measure high-frequency energy and high-band residual. This is a triage metric for nonlinear aliasing; it should be used with generated high-frequency stimuli, not trusted blindly on arbitrary guitar DI.

Sag diagnostics:

: Compare level drop and recovery inside burst windows. This targets supply sag, bias shift, compression memory, and other slow dynamic behaviors.

Intermodulation:

: Compares two-tone intermodulation products such as 2F1-F2, 2F2-F1, F2-F1, and F1+F2 relative to the main tones. This matters for nonlinear stages because intermodulation often reveals harshness and chord smear better than THD.

Segment Markers

Segment files live in lab/segments/ and follow lab/schemas/segments.schema.json. They define named time windows:

{
  "schema_version": 1,
  "segments": [
    {
      "name": "opening_attack",
      "kind": "attack",
      "start_s": 0.0,
      "end_s": 0.35
    }
  ]
}

Supported segment kinds today:

• general, sustain, decay, and silence: local gain, residual, spectral, and envelope metrics.
• attack: adds peak timing, rise timing, and overshoot comparison.
• harmonic: adds THD and H2-H5 deltas; use fundamental_hz when the segment is a known sine or stable note.
• imd: adds two-tone intermodulation diagnostics; use first_hz and second_hz when the tones are known.
• aliasing: adds high-band and high-band residual checks.
• sag: adds drop and recovery comparison across the segment.

The first committed marker file is lab/segments/guitar-chords.markers.json. It is a coarse analysis anchor for the bundled dry guitar sample, not a canonical transcription of the performance.

Directory Layout

lab/experiments/

: Committed experiment plans. The first one is 001-chain-reference-analysis.md; the first completed controlled-stimulus pass is 002-nox30-stimulus-batch.md.

lab/schemas/

: Committed JSON schemas for reproducible metadata.

lab/segments/

: Committed segment marker files used for local diagnostics.

lab/stimuli/

: Generated synthetic WAV stimuli and generated marker files. Ignored by git; regenerate them with generate-stimuli.

lab/datasets/

: Local generated or imported datasets. Ignored by git by default.

lab/models/

: Local generated neural-cell artifacts, checkpoints, optional ONNX exports, and experimental descriptors. Ignored by git by default except for source-safe notes.

lab/references/

: Local reference WAVs, NAM renders, measured captures, or SPICE exports. Ignored by git by default unless redistribution rights are explicit.

lab/renders/

: Local Greybound WAV renders and run metadata. Ignored by git by default.

lab/reports/

: Local generated comparison reports. Ignored by git by default.

Current Status

Implemented:

• Python package greybound-lab,
• render-rig command,
• compare-wav command,
• run metadata schema,
• segment marker schema,
• SPICE dataset manifest schema,
• neural-cell artifact descriptor schema,
• segment-level diagnostics for attack, harmonic, high-band/aliasing, and sag windows,
• band residual diagnostics for every segment,
• intermodulation diagnostics for two-tone segments,
• generated stimuli for harmonic, high-band/aliasing, sag, attack, and future intermodulation protocols,
• first controlled-stimulus Nox30 clean/driven batch,
• first common-cathode SPICE import/report,
• NAM reference protocol and metadata schema,
• first real NAM A2 render and comparison against Greybound Nox30,
• tests for metric alignment and metadata generation,
• first experiment plan for chain reference analysis.

Not implemented yet:

• plots in reports,
• NAM render batch workflow for priority models,
• broader SPICE fixture automation beyond the first common-cathode import,
• generated SPICE datasets using the dataset manifest schema,
• first experimental PyTorch MLP training from a SPICE dataset,
• first Greybound neural-cell descriptor and packed weight export,
• experimental Rust neural-cell descriptor/weight loader and scalar MLP inference,
• optional generated-vector Python/Rust equivalence check for local neural-cell artifacts,
• SPICE evaluation report for exported neural-cell artifacts.
• analytic Rust common-cathode baseline evaluation against the same SPICE dataset.

Next Research Steps

The next useful improvements are:

• add plots to comparison reports,
• define a real reference WAV protocol,
• add a stronger aliasing score that separates harmonics from folded non-harmonic residuals,
• extend the common-cathode SPICE fixture to level sweeps and two-tone IMD,
• generate the first common-cathode SPICE dataset manifest,
• train a dynamic or better-conditioned cell only after the analytic comparison identifies where the current approximation fails,
• inspect analytic residuals by waveform/phase/harmonic content to decide whether the residual is caused by initialization, capacitor discretization, solver tolerance, or nonlinear law mismatch,
• decide how neural-cell artifacts should be mounted into an amp model without violating real-time constraints,
• standardize one Nox30-style NAM comparison,
• decide the first fitted cell only after the reports show where the largest error is.