peter-swimm/nardwahl
Watch
1
0
Fork
You've already forked nardwahl
 0
Code Issues 4 Pull requests Projects Releases Packages Wiki Activity Actions

Calibrate shared G1 model laws and track per-module parity #9

Merged
peterswimm merged 16 commits from baseline-calibration into main 2026-10-21 20:36:06 +00:00
Conversation 0 Commits 16 Files changed 9 +3342 -313
peterswimm commented 2026-10-21 20:28:41 +00:00
Owner
Copy link

Scope

Calibrate the shared model laws the v23 sweep exposed, give the remaining silent fixtures real stimuli, and start tracking per-module parity.

  • Parity tracking. nardwahl.py parity-matrix reduces a probe sweep summary to public-safe verdict counts for all 109 runtime modules, with a triage stage and each module's main gap, written to docs/G1_PARITY_MATRIX.{json,md} (a test keeps the Markdown in sync with the JSON). docs/PARITY_ROADMAP.md lays out the phases from observing every module to Disting NT audition. Linked from the README and the pipeline doc.
  • Fixtures.
    • Output and fade modules default to a captured destination / audible fade; OnOff on; LevMult off its bipolar zero. Other modules' fixtures are byte-identical.
    • Gated envelopes get a held Keyboard gate. An unpatched gate is low on G1, so ADSR, Mod-Env and Multi-Env probes previously compared only their OscA observer.
  • Levels (one measured G1 level taper, -23.3 dB at 64, 0 dB at 127): Mixer (8), LevMult, OscSineBank (plus fixed per-oscillator start phases), FilterBank (stagger-tuned band pairs), Phaser input level; Compressor static law (knee, ratio, limiter, act/mon polarity); FilterE cutoff, tracking and resonance.
  • Envelopes: measured time law, linear attack, exponential decay/release, ADSR attack shapes, AHD, AD-Env and Multi-Env segments and curves; an unpatched gate is low; Keyboard note into a gate is high above 64.
  • Pitch: ClkGen tempo (was ~4.7× fast); DrumSynth tune grid, decay, level and noise path; Sample&Hold holds on the rising edge.
  • Shape: Digitizer quantizer and sample-rate laws (was a passthrough); Expander threshold direction, ratio slope and open start; Overdrive input order; NoteVelScal gain law; OscA triangle uses os.lf_triangle (Faust's triangle ramps in over ~75 samples).
  • One stray change to the Expander detector inside the Compressor commit is reverted on its branch; the Expander calibration owns that line.

Testing

  • 125 tests pass (python3 -m unittest discover -s tests); new tests cover the parity matrix, fixture defaults and gates, and each changed law (tests/test_level_laws.py, tests/test_patch_layouts.py).
  • The merged tree is byte-identical to the tree that produced plan v24 and its sweep.
  • Full probe sweep against the private G1-Emu adapter, 1,489 probes per plan; no probe that passed in v23 fails in v24:
verdict v21 v23 v24 (this PR)
pass 207 541 842
peak-only 307 23 7
level 300 361 192
pitch 371 426 298
shape-or-phase 50 54 100
silence-mismatch 118 29 14
both-silent 134 55 36

Shape-or-phase rises because gated envelopes are now actually compared; their level and mean pitch match, and the residual is phase drift from OscA's pitch-mod depth (G1 measures 4.37–4.48 semitones per unit across levels; the model uses a constant 4.379).

Parity stages (v23 → v24): sweep-clean 10 → 17, near 18 → 22, partial 25 → 23, divergent 53 → 44, not-swept 3 → 3.

Impact Statement

  • Existing features: baseline fixtures change for output, fade, OnOff, LevMult and gated-envelope modules, so earlier plans no longer match regenerated fixtures. Faust output changes for every module listed in Scope.
  • Public APIs / contracts: new parity-matrix command; no existing command changes.
  • Schemas: new public format nardwahl-parity-matrix-v1.
  • Data / artifacts: docs/G1_PARITY_MATRIX.{json,md} added; docs/G1_BASELINE_STATUS.json regenerated from v24 (both-signal midpoints 92 → 104 of 109).
  • Integrations: target renders still need the separately installed G1-Emu adapter profile.
  • Mappings: no expected impact.
  • Backward compatibility: no module status changes; sweep verdicts and stages remain triage only.

Docs

docs/PARITY_ROADMAP.md (new), docs/G1_PARITY_MATRIX.md (generated), docs/G1_BASELINE_PIPELINE.md and README links.

Known gaps for the next round: OscA pitch-mod depth law, LFOB/LFOC startup phase, ClkGen pre-roll (AHD, AD-Env), Phaser notch structure, PercOsc pitch, OscA FMA input, Expander gate.

## Scope Calibrate the shared model laws the v23 sweep exposed, give the remaining silent fixtures real stimuli, and start tracking per-module parity. - **Parity tracking.** `nardwahl.py parity-matrix` reduces a probe sweep summary to public-safe verdict counts for all 109 runtime modules, with a triage stage and each module's main gap, written to `docs/G1_PARITY_MATRIX.{json,md}` (a test keeps the Markdown in sync with the JSON). `docs/PARITY_ROADMAP.md` lays out the phases from observing every module to Disting NT audition. Linked from the README and the pipeline doc. - **Fixtures.** - Output and fade modules default to a captured destination / audible fade; OnOff on; LevMult off its bipolar zero. Other modules' fixtures are byte-identical. - Gated envelopes get a held Keyboard gate. An unpatched gate is low on G1, so ADSR, Mod-Env and Multi-Env probes previously compared only their OscA observer. - **Levels** (one measured G1 level taper, -23.3 dB at 64, 0 dB at 127): Mixer (8), LevMult, OscSineBank (plus fixed per-oscillator start phases), FilterBank (stagger-tuned band pairs), Phaser input level; Compressor static law (knee, ratio, limiter, act/mon polarity); FilterE cutoff, tracking and resonance. - **Envelopes:** measured time law, linear attack, exponential decay/release, ADSR attack shapes, AHD, AD-Env and Multi-Env segments and curves; an unpatched gate is low; Keyboard note into a gate is high above 64. - **Pitch:** ClkGen tempo (was ~4.7× fast); DrumSynth tune grid, decay, level and noise path; Sample&Hold holds on the rising edge. - **Shape:** Digitizer quantizer and sample-rate laws (was a passthrough); Expander threshold direction, ratio slope and open start; Overdrive input order; NoteVelScal gain law; OscA triangle uses `os.lf_triangle` (Faust's triangle ramps in over ~75 samples). - One stray change to the Expander detector inside the Compressor commit is reverted on its branch; the Expander calibration owns that line. ## Testing - 125 tests pass (`python3 -m unittest discover -s tests`); new tests cover the parity matrix, fixture defaults and gates, and each changed law (`tests/test_level_laws.py`, `tests/test_patch_layouts.py`). - The merged tree is byte-identical to the tree that produced plan v24 and its sweep. - Full probe sweep against the private G1-Emu adapter, 1,489 probes per plan; no probe that passed in v23 fails in v24: | verdict | v21 | v23 | v24 (this PR) | |---|---|---|---| | pass | 207 | 541 | **842** | | peak-only | 307 | 23 | 7 | | level | 300 | 361 | 192 | | pitch | 371 | 426 | 298 | | shape-or-phase | 50 | 54 | 100 | | silence-mismatch | 118 | 29 | 14 | | both-silent | 134 | 55 | 36 | Shape-or-phase rises because gated envelopes are now actually compared; their level and mean pitch match, and the residual is phase drift from OscA's pitch-mod depth (G1 measures 4.37–4.48 semitones per unit across levels; the model uses a constant 4.379). Parity stages (v23 → v24): sweep-clean 10 → 17, near 18 → 22, partial 25 → 23, divergent 53 → 44, not-swept 3 → 3. ## Impact Statement - Existing features: baseline fixtures change for output, fade, OnOff, LevMult and gated-envelope modules, so earlier plans no longer match regenerated fixtures. Faust output changes for every module listed in Scope. - Public APIs / contracts: new `parity-matrix` command; no existing command changes. - Schemas: new public format `nardwahl-parity-matrix-v1`. - Data / artifacts: `docs/G1_PARITY_MATRIX.{json,md}` added; `docs/G1_BASELINE_STATUS.json` regenerated from v24 (both-signal midpoints 92 → 104 of 109). - Integrations: target renders still need the separately installed G1-Emu adapter profile. - Mappings: no expected impact. - Backward compatibility: no module status changes; sweep verdicts and stages remain triage only. ## Docs `docs/PARITY_ROADMAP.md` (new), `docs/G1_PARITY_MATRIX.md` (generated), `docs/G1_BASELINE_PIPELINE.md` and README links. Known gaps for the next round: OscA pitch-mod depth law, LFOB/LFOC startup phase, ClkGen pre-roll (AHD, AD-Env), Phaser notch structure, PercOsc pitch, OscA FMA input, Expander gate.
Unpatched envelope gates are low, the envelope starts 4 ms after its edge,
time knobs gain 0.2 ms per step, decay and release are exponential (40 dB
per time), ADSR attack shapes, AHD/AD-Env and Multi-Env segment laws, and
Keyboard-note gates threshold above note 64.
parity-matrix reduces a probe sweep summary to public-safe verdict counts for
all 109 runtime modules, with a triage stage (sweep-clean, near, partial,
divergent, blocked, unobserved, not-swept) and each module's main gap. The
Markdown renders only from the checked-in JSON and a test keeps them in sync.
PARITY_ROADMAP.md lays out the phases from observing every module to Disting NT
audition. The matrix is generated from probe-sweep v23.
The Compressor commit also changed the Expander's envelope follower, which
is outside the level calibration and was not swept. Restore it; the Expander
calibration owns that line.
An unpatched gate input is low on G1, so gated envelopes stayed at zero and
their probes compared only the OscA observer. A held Keyboard gate exercises
attack, decay and sustain; release is not observed. Keyboard is a singleton,
so a probe's own Keyboard source is reused.
Regenerate the parity matrix and baseline status from plan v24
All checks were successful
PR policy check / policy-gate (pull_request) Successful in 0s
Details
8c3efa8b6d 
Probe sweep v24: 842 of 1,489 probes pass (541 in v23); sweep-clean 17,
near 22, partial 23, divergent 44. Baseline midpoints carrying signal on both
engines rise from 92 to 104 of 109.
peterswimm deleted branch baseline-calibration 2026-10-21 20:36:07 +00:00
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
peter-swimm/nardwahl!9
No description provided.