Hardware Compressor Capture

Hardware Compressor Capture

Use this skill with the repository AGENTS.md, mc-patcher, and hardware-audio-emulation. This skill owns the capture campaign and dataset release gates; the hardware-emulation skill owns general model selection, antialiasing, perception, and plug-in validation.

Respect the requested operation

• For review or planning, inspect and write the protocol only. Do not operate the user's DAW, change hardware controls, power-cycle equipment, or start a capture.
• For capture execution, confirm the exact physical route and safe level limit before generating or playing high-level probes. Never open a tube unit or request internal measurements unless a qualified service technician owns that work.
• When the user freezes the converter control plane, treat that as a hard non-mutation boundary: do not change converter routing, patching, trims, gains, clock, sample rate, buffer size, presets, or any other converter state. Only read-only verification is allowed. A route diagnostic must never write converter state, even temporarily.
• When the user also stops converter discovery, do not enumerate, discover, probe, or refresh converter identity through CoreAudio, CueMix, AVB, USB, network discovery, or another device API. Treat the user-declared device name and channel contract as the only retained authority. If a capture is active when the stop arrives, interrupt it, quarantine any unpaired partial artifact, and never resume automatically; offline annotation, validation, packaging, and upload may continue without device access.
• Treat user-declared capture CLI channel numbers as literal arguments after confirming the tool's numbering convention. Do not pre-decrement or rebase a one-based declaration when the tool already performs that conversion. For this user's MOTU capture route, every capture invocation must pass output 33/34 and return 37/38 exactly, and the wrapper must fail closed if any other channel is requested. Retain but quarantine recordings made through a mismatched route; they are never training-eligible.
• Treat raw recordings as evidence. Never normalize, destructively denoise, overwrite, or discard them during alignment or dataset preparation. The only authorized deletion is the exact post-Release cleanup below, after remote upload verification has closed.

Freeze the claim and unit identity first

Write the intended claim as a bounded behavior domain: unit and revision, source/load conditions, sample rates, input levels, controls and modes, stereo topology, warm-up state, and longest history to be reproduced. “Complete” or “exact” capture is not an acceptance condition.

User-specific Tegeler topology restriction

For this user's Tegeler capture work, do not enter, toggle, qualify, or capture DualMono in any mode except an explicitly designated Stereo Transformer mode. For every other mode, keep the unit in linked Stereo and do not touch the DualMono controls. If Stereo Transformer mode is explicitly active and DualMono is required, initialize both channel Threshold controls to the same exact raw value and verify that equality before and after each operation; any unexpected mismatch is a restore/stop condition. This restriction is campaign-specific and does not generalize to other compressors or users. Where a referenced Tegeler campaign document requests broader DualMono qualification, this user-specific restriction takes precedence for this user's sessions.

Capture-priority precision exception

When the user explicitly prioritizes completing the capture over exact control precision, an unreachable raw target is not by itself a reason to abort the round. Use the nearest value that the supported control path can set and hold reliably, but record the requested target and the observed /info raw value separately for every affected control. Base the canonical state identity on the observed hardware value, never on the requested value; if it collides with an existing state or creates an unplanned state, retain that disposition in the ledger and do not silently deduplicate it. Mark the round and affected files as precision-degraded/approximate, keep them out of exact-state claims and training eligibility until independently reviewed, and preserve the normal restore, clipping, busy, route, and provenance gates. This exception does not permit unverified levels, unstable controls, hidden drift, DualMono use under the Tegeler restriction above, or reporting an approximate capture as exact, hardware-fit, or model-ready.

For the user's Tegeler Audio Manufaktur Schwerkraftmaschine Mode 5 deep black-box campaign only, read references/tegeler-schwerkraftmaschine-deep-blackbox.md before inventory closure, parameter-design freeze, capture, or Release sync. Its 800-state floor, two-round phase gate, and second-round 12-hour cadence are campaign-specific and must not be generalized to other compressors.

Do not generate a full matrix until the unit identity, control topology, safe headroom, and service state are known. Record serial, revision, modifications, tube or gain-cell state when documented, recent service, control detents, source/load impedance, and channel condition. For a Manley Stereo Variable Mu, read references/manley-variable-mu.md.

Qualify the chain before the device

Read references/capture-protocol.md before designing routing, calibration, stimuli, or parameter coverage.

Require these gates in order:

1. converter loopback and monitor-path headroom;
2. the user's declared converter calibration authority and safe-level mapping;
3. bypass plus integer/fractional delay, polarity, channel mapping, and drift;
4. declared warm-up and stabilization check;
5. active no/low-gain-reduction baseline and repeatability anchors;
6. static, dynamic, nonlinear, stereo/link, and recovery pilots.

Stop at the first failed gate. A green engineering script cannot compensate for clipped converters, an unidentified revision, an unstable unit, or a missing declared safe-level mapping.

For this user, use the MOTU output calibration/specification and current output trim as the level authority. Never require an external true-RMS meter or audio analyzer as a prerequisite. Record the exact MOTU model, output range/reference, channel trims, digital waveform convention, load assumption, and safety margin; label the result MOTU-nominal dBu, not terminal-measured absolute dBu. Verify the usable range with conservative loopback, bypass, ADC-headroom and clipping checks. Missing terminal-voltage measurement narrows the analogue-level claim but does not stop capture when those MOTU-based gates pass.

Preserve three different signals when available: the original digital stimulus, a simultaneous analogue-input reference tap, and the wet output. The digital stimulus is normally the production model's canonical input; the tap is evidence for delivered level, timing, drift, and converter behavior. Do not train on the tap and then silently feed pre-DAC samples to the plug-in.

Design the campaign by behavior, not duration

Start with a pilot, then characterize the axes that actually vary, then record the first training set, and add settings only after residuals or model disagreement identify a gap. Do not prescribe a universal number of hours.

Separate controls by function before sampling:

• enumerate switches, detents, ratio modes, side-chain filters, and routing;
• sample continuous attack, threshold, drive, and other nonlinear controls only after their measured law and repeatability are known;
• test whether output attenuation is a separable gain or a loading-dependent analogue behavior before adding it to a neural parameter space;
• keep input drive and threshold distinct when one changes audio-path coloration and the other changes side-chain action;
• treat stereo Link as a measured control-state coupling, not automatically as summed audio, left-master behavior, or an arbitrary M/S matrix.

Use multiple stimulus families because no single sweep or noise file identifies all layers: low-reduction complex response, steady transfer curves, stepped and interrupted envelopes, tones and sparse multitones, periodic random-phase multisines with repetitions, asymmetric stereo/link probes, recovery/settling sequences, and held-out program material. Label each probe by the hypothesis it tests and the observation that would reject it.

Preserve state and provenance

Read references/data-contract-and-qa.md before recording production data or releasing a dataset. Copy assets/capture-session.template.json for each session and validate it with scripts/validate_capture_manifest.py.

Paths beginning with assets/, references/, or scripts/ in this skill are relative to the directory containing this SKILL.md. Run those commands from that directory, or resolve the path explicitly before invoking them.

Keep canonical original recordings under campaign/raw/<session-id>/; do not mix them with derivatives, reports, training products, or sync packages. For training or remote synchronization, also read references/training-consumption-and-sync.md. Create an exact claim under campaign/training/ready/<dataset-id>/ before a training run and seal its receipt and artifacts under campaign/training/completed/<training-run-id>/ afterward. Mark consumed raw evidence with SHA-256 sidecars under campaign/training/ledger/; never mark or rewrite the audio file itself. A loader or new training run must fail closed when the same raw hash has already completed the same training objective.

Close every capture round with file-level handoff annotations. Copy assets/capture-round-annotations.template.json into the round directory as round-file-annotations-v1.json, add one entry for every persistent raw, stimulus, derivative, analysis, accepted, rejected, or superseded data file created or relied on by the round, and write a concise ROUND_NOTES.md for human readers. Never write annotations into raw audio metadata because that changes the evidence hash. Validate the sidecar with scripts/validate_round_annotations.py --check-files; when the round has a self-contained output directory, also pass it as --inventory-root. A round is not complete and its capture/live-resource ownership is not released for handoff until all retained files have a purpose, control/provenance reference, QA state, training eligibility, exclusion note where applicable, and verified hash. PENDING, UNKNOWN, or unreviewed annotations are an explicit handoff blocker.

When a round publishes an exact dataset allowlist, treat that allowlist as the training-eligibility authority and mirror its per-file booleans in the round annotations. Validation and withheld files remain ineligible; rejected, quarantined, and superseded files remain ineligible. If primary and comparison target variants are both eligible, give them distinct training uses and explicitly forbid silent mixing. Before closing or syncing the round, run scripts/validate_round_dataset_eligibility.py --check-files in addition to the ordinary round annotation validator. A validator that reads only one of the two files cannot close this gate.

After the round sidecars pass file-hash validation and are stable, sync the closed round to an immutable GitHub Release. GitHub Releases are the required remote target for large capture and training data; the repository Code page stores only lightweight indexes, annotations, ledgers, and Release download instructions. Keep the local source unchanged while packaging and uploading, and never sync a half-written round. Remote re-download is not required. The deletion gate is: the Release is published, every required asset reports uploaded, each provider-exposed size and SHA-256 digest matches the locally verified package asset, the exact package manifest matches every current local source member path, size, and SHA-256, and the lightweight Code receipt and restore instructions identify that same Release. Also confirm that no capture, validation, or training process is using the local files.

After all of those checks pass, permanently delete the locally completed capture payload files and local package parts covered by that exact verified manifest. This is a standing user instruction and does not require a second confirmation. Delete only an explicit path-and-SHA allowlist under the declared capture/package roots; never delete a directory recursively, a changed file, an uncovered file, an active file, or lightweight annotations, manifests, hashes, ledgers, and restore instructions. Write an append-only lightweight deletion receipt under the authoritative Release index with each deleted path, size, SHA-256, Release tag/assets, remote-verification time, deletion time, and result. A failed or partial upload, provider digest/size comparison, source hash comparison, Code receipt, or process-ownership check deletes nothing. Release upload and local deletion do not promote a round or dataset to RELEASED or model-ready status.

Keep session order, preceding stimulus, preroll, reset method, silence, and thermal cycle in the manifest. Split train/validation/test by whole program, setting groups, stimulus families, and capture sessions; adjacent chunks from one recording never cross splits. Estimate deterministic behavior and stochastic noise/drift separately from multiple repeats.

Hand off evidence before choosing the final model

Read references/model-handoff.md when the user asks for training, architecture choice, or plug-in acceptance. The capture does not pre-authorize a large neural architecture. Compare at least one simple compressor gray-box baseline with justified recurrent, convolutional, or state-space candidates. Add audio-rate nonlinear residuals, output feedback, multirate paths, or joint stereo state only when the pilots falsify the simpler model.

Do not use a two-repeat scalar inequality as proof of equivalence. Predeclare metrics and practical tolerances, estimate hardware repeatability distributions from multiple repetitions, compare systematic bias separately, and retain an independent listening set.

Stop conditions

Stop full capture and report the missing evidence when any of these applies:

• unit identity or a behavior-changing modification is unresolved;
• the required dBu level exceeds verified DAC, ADC, pad, cable, or DUT margin;
• loopback/bypass polarity, clipping, drift, or channel mapping is unresolved;
• repeated pilots show unexplained thermal, service, or intermittent behavior;
• two campaign revisions fail under the same modeling hypothesis;
• a derived gain-reduction target is being treated as ground truth without a validated control-voltage or envelope measurement.

Report the result by evidence layer

Report the exact unit and route, calibrated domain, raw-data location and hashes, passed/failed capture gates, parameter and stimulus coverage, repeated conditions, split policy, alignment/drift method, known exclusions, and the strongest justified claim. Keep capture integrity, offline model metrics, realtime plug-in tests, DAW tests, listening, and deployment as separate statuses.