Drive baseline probes by input role and calibrate the inputs they exposed #13

Merged

peterswimm merged 1 commit from baseline-inputs into main

2026-10-22 02:25:25 +00:00

peterswimm commented

2026-10-22 02:24:01 +00:00

Owner

Scope

Baseline probes now drive each probed input with the stimulus its role needs. The models those stimuli exposed are calibrated against G1-Emu, each in both export_faust and its render_module_output mirror.

Stimulus by port role (BASELINE_PROBE_STIMULI):
- trig/trigger/retrig → ClkGen.
- gate → Keyboard gate (port 1).
- vel → Keyboard velocity (port 2).
- Note and pitch inputs are unchanged.
DrumSynth 58 velocity: level is linear in vel, clipped to 0..1, and independent of Amount. An unpatched vel reads 1.
Mod-Env 23 retrig: a rising retrig edge drops the gate for one sample, so g1_env_adsr re-attacks from its current level.
EnvFollower 71:
- Model: a peak detector with a 0.6 ms hold and a 7 ms peak release, then a one-pole attack/release stage.
- Attack: 4.38 ms at 64, doubling every 12 steps. Release: 0.91 ms + 8.97 ms·2^(v/20.3).
- Fitted to gate steps; the fixture now feeds it a gated Mod-Env burst.
CompareAB 89: adds the A>=B logic output (per Nomad2026), which makes it control-observable.
MasterOsc 97:
- Observed through a slaved OscSlvE, which makes it audio-observable.
- Pitch inputs are wired: coarse, fine, kbt, and pitch mod 1/2 with the Type II depth law.
PolyAreaIn 127:
- Sums the poly section's CVA bus at 0.5× its output level; +6 dB doubles it. Muted outputs and outputs to the main pair are excluded.
- Exportable as approximate, not Disting-released.
AudioIn 2: the G1 reference now runs with the adapter's --input-sine 220 (L 220 Hz, R 440 Hz). The model renders the same unit sines.
GainControl 44: an undriven control is 0. With shift on, the control maps to (c+1)/2.

No tolerance was widened. No numbered plan was cut and the parity matrix was not regenerated.

The v25 calibration session owns, and this PR leaves untouched:

OscA/OscB pitch-mod and FMA
ClkGen 68
LFOB/LFOC
PercOsc
Phaser
Expander

Testing

python3 -m unittest discover -s tests: 137 tests OK. New BaselineInputTests cover each item above.
tools/export_public.py into an empty directory verifies clean.
Fixture hashes against plan v24: only modules 20, 23, 46, 52, 53, 58, 71, 84, 89, 97, 100 and 127 changed (41 probes). Tier changes are 89 → control-observable, and 97 and 127 → audio-observable. AudioIn and GainControl fixtures are byte-identical; their change is in the adapter tone and the model.
Full probe sweep on a scratch plan (every v24 probe re-run, plus 22 new):

Verdict	v24	scratch
pass	842	865
level	192	191
pitch	298	294
shape-or-phase	100	110
peak-only	7	7
both-silent	36	39
silence-mismatch	14	5

Probe(s)	v24 → scratch
002-001, 002-002, 002-003 (AudioIn)	silence-mismatch → pass
044-001, 044-002, 044-005, 044-006 (GainControl, undriven)	silence-mismatch → both-silent
044-003 (GainControl shift)	level → pass
058-034, 058-035 (DrumSynth trig/vel)	silence-mismatch → pass
071-003, 071-004 (EnvFollower)	pitch → pass
071-001, 071-002, 071-005, 071-006, 071-007	pitch → shape-or-phase †
053-002 (S&H trig)	both-silent → shape-or-phase †
020-016 (ADSR retrig)	shape-or-phase → pitch †
020-015, 023-020, 046-014, 052-024, 084-008	pass → shape-or-phase / pitch ‡
089-001..004 (CompareAB, new)	3 pass, 1 pitch †
097-001..013 (MasterOsc, new)	12 pass, 1 shape-or-phase
127-001..005 (PolyAreaIn, new)	5 pass

No other v24 verdict changed, and no real v24 pass regressed.

‡ Vacuous passes, now midpoint-equivalent (not regressions). In v24 these gate/trig probes were fed a Keyboard note. On G1 an envelope gate driven by a note fires only above note 64.5, so the envelope never fired. All five shared the identical idle rms_error of 0.0022. Now that the gate fires, each matches its own module's midpoint metric:

020-015: 0.0517 (020-001: 0.0517)
023-020: 0.0525 (023-001: 0.0526)
046-014: 0.0433 (046-001: 0.0433)
052-024: 0.0424 (052-001: 0.0423)
084-008: 0.0980 (084-001: 0.0980)

What still fails is the OscA pitch-mod observer, which v25 owns.

† These are observed through the OscA control observer or depend on ClkGen phase, both owned by v25. With ClkGen's phase patched in scratch only, all four Mod-Env retrig cases pass (rms 0.004–0.006). G1 ClkGen rises at 0.0658 + n·0.125 s, while Faust rises about 49 ms earlier.

Impact Statement

Envelope, drum and sample-and-hold gate/trig/vel probes now exercise the module instead of measuring an idle observer. This removes five vacuous passes from the baseline.
Four modules gain G1-calibrated behaviour: DrumSynth velocity, Mod-Env retrig, EnvFollower and GainControl.
AudioIn and the previously structural-only MasterOsc, CompareAB and PolyAreaIn become measurable.
PolyAreaIn becomes Faust-exportable as approximate. It stays off the Disting release list.
The G1 reference adapter gains an --input-sine stimulus, used only for patches that contain AudioIn.

Follow-ups, not fixed here:

Bipolar Constant scaling is (v−64)/64 on G1; the model uses /63.
ADSR retrig is unmodelled.
Keyboard velocity reads 1.0 at MIDI 100 on G1; the model uses 100/127.
MasterOsc pitch-mod is 3.7–16.5 cents off through the shared Type II depth law (v25).

## Scope Baseline probes now drive each probed input with the stimulus its role needs. The models those stimuli exposed are calibrated against G1-Emu, each in both `export_faust` and its `render_module_output` mirror. - **Stimulus by port role** (`BASELINE_PROBE_STIMULI`): - `trig`/`trigger`/`retrig` → ClkGen. - `gate` → Keyboard gate (port 1). - `vel` → Keyboard velocity (port 2). - Note and pitch inputs are unchanged. - **DrumSynth 58 velocity:** level is linear in vel, clipped to 0..1, and independent of Amount. An unpatched vel reads 1. - **Mod-Env 23 retrig:** a rising retrig edge drops the gate for one sample, so `g1_env_adsr` re-attacks from its current level. - **EnvFollower 71:** - Model: a peak detector with a 0.6 ms hold and a 7 ms peak release, then a one-pole attack/release stage. - Attack: 4.38 ms at 64, doubling every 12 steps. Release: 0.91 ms + 8.97 ms·2^(v/20.3). - Fitted to gate steps; the fixture now feeds it a gated Mod-Env burst. - **CompareAB 89:** adds the `A>=B` logic output (per Nomad2026), which makes it control-observable. - **MasterOsc 97:** - Observed through a slaved OscSlvE, which makes it audio-observable. - Pitch inputs are wired: coarse, fine, kbt, and pitch mod 1/2 with the Type II depth law. - **PolyAreaIn 127:** - Sums the poly section's CVA bus at 0.5× its output level; +6 dB doubles it. Muted outputs and outputs to the main pair are excluded. - Exportable as `approximate`, not Disting-released. - **AudioIn 2:** the G1 reference now runs with the adapter's `--input-sine 220` (L 220 Hz, R 440 Hz). The model renders the same unit sines. - **GainControl 44:** an undriven control is 0. With shift on, the control maps to (c+1)/2. No tolerance was widened. No numbered plan was cut and the parity matrix was not regenerated. The v25 calibration session owns, and this PR leaves untouched: - OscA/OscB pitch-mod and FMA - ClkGen 68 - LFOB/LFOC - PercOsc - Phaser - Expander ## Testing - `python3 -m unittest discover -s tests`: 137 tests OK. New `BaselineInputTests` cover each item above. - `tools/export_public.py` into an empty directory verifies clean. - Fixture hashes against plan v24: only modules 20, 23, 46, 52, 53, 58, 71, 84, 89, 97, 100 and 127 changed (41 probes). Tier changes are 89 → control-observable, and 97 and 127 → audio-observable. AudioIn and GainControl fixtures are byte-identical; their change is in the adapter tone and the model. - Full probe sweep on a scratch plan (every v24 probe re-run, plus 22 new): | Verdict | v24 | scratch | |---|---|---| | pass | 842 | 865 | | level | 192 | 191 | | pitch | 298 | 294 | | shape-or-phase | 100 | 110 | | peak-only | 7 | 7 | | both-silent | 36 | 39 | | silence-mismatch | 14 | 5 | | Probe(s) | v24 → scratch | |---|---| | 002-001, 002-002, 002-003 (AudioIn) | silence-mismatch → pass | | 044-001, 044-002, 044-005, 044-006 (GainControl, undriven) | silence-mismatch → both-silent | | 044-003 (GainControl shift) | level → pass | | 058-034, 058-035 (DrumSynth trig/vel) | silence-mismatch → pass | | 071-003, 071-004 (EnvFollower) | pitch → pass | | 071-001, 071-002, 071-005, 071-006, 071-007 | pitch → shape-or-phase † | | 053-002 (S&H trig) | both-silent → shape-or-phase † | | 020-016 (ADSR retrig) | shape-or-phase → pitch † | | 020-015, 023-020, 046-014, 052-024, 084-008 | pass → shape-or-phase / pitch ‡ | | 089-001..004 (CompareAB, new) | 3 pass, 1 pitch † | | 097-001..013 (MasterOsc, new) | 12 pass, 1 shape-or-phase | | 127-001..005 (PolyAreaIn, new) | 5 pass | No other v24 verdict changed, and no real v24 pass regressed. ‡ **Vacuous passes, now midpoint-equivalent (not regressions).** In v24 these gate/trig probes were fed a Keyboard note. On G1 an envelope gate driven by a note fires only above note 64.5, so the envelope never fired. All five shared the identical idle rms_error of 0.0022. Now that the gate fires, each matches its own module's midpoint metric: - 020-015: 0.0517 (020-001: 0.0517) - 023-020: 0.0525 (023-001: 0.0526) - 046-014: 0.0433 (046-001: 0.0433) - 052-024: 0.0424 (052-001: 0.0423) - 084-008: 0.0980 (084-001: 0.0980) What still fails is the OscA pitch-mod observer, which v25 owns. † These are observed through the OscA control observer or depend on ClkGen phase, both owned by v25. With ClkGen's phase patched in scratch only, all four Mod-Env retrig cases pass (rms 0.004–0.006). G1 ClkGen rises at 0.0658 + n·0.125 s, while Faust rises about 49 ms earlier. ## Impact Statement - Envelope, drum and sample-and-hold gate/trig/vel probes now exercise the module instead of measuring an idle observer. This removes five vacuous passes from the baseline. - Four modules gain G1-calibrated behaviour: DrumSynth velocity, Mod-Env retrig, EnvFollower and GainControl. - AudioIn and the previously structural-only MasterOsc, CompareAB and PolyAreaIn become measurable. - PolyAreaIn becomes Faust-exportable as `approximate`. It stays off the Disting release list. - The G1 reference adapter gains an `--input-sine` stimulus, used only for patches that contain AudioIn. Follow-ups, not fixed here: - Bipolar Constant scaling is (v−64)/64 on G1; the model uses /63. - ADSR retrig is unmodelled. - Keyboard velocity reads 1.0 at MIDI 100 on G1; the model uses 100/127. - MasterOsc pitch-mod is 3.7–16.5 cents off through the shared Type II depth law (v25).

peterswimm added 1 commit

2026-10-22 02:24:01 +00:00

Drive baseline probes by input role and model the inputs that exposed

PR policy check / policy-gate (pull_request) Successful in 0s

Details

832a70d101

Probed trig/retrig inputs get a ClkGen, gate inputs the Keyboard gate and
vel inputs Keyboard velocity, so envelope and drum probes actually fire.
Calibrate DrumSynth velocity, Mod-Env retrig, the EnvFollower peak/hold and
attack/release laws, MasterOsc pitch (observed through a slaved OscSlvE),
the PolyAreaIn CVA bridge, AudioIn against the adapter test tone and
GainControl, and add CompareAB's A>=B output.

peterswimm merged commit 23ba636052 into main

2026-10-22 02:25:25 +00:00

peterswimm deleted branch baseline-inputs

2026-10-22 02:25:26 +00:00

peterswimm referenced this pull request from a commit

2026-10-22 02:25:26 +00:00

Merge pull request 'Drive baseline probes by input role and calibrate the inputs they exposed' (#13) from baseline-inputs into main

peterswimm referenced this pull request

2026-10-22 04:00:25 +00:00

Regenerate the parity matrix and baseline status from plan v26 #15