Reference suites: twelve modules reach reference-compared #10

Merged

peterswimm merged 5 commits from reference-suites into main

2026-10-22 02:14:10 +00:00

peterswimm commented

2026-10-21 23:41:25 +00:00

Owner

Scope

Phase 3 of the parity roadmap: private G1 reference suites for the modules the parity sweep found clean (excluding modules owned by the v25-calibration and input-fix work). Twelve modules now complete their reference contract and are marked reference-compared in the catalog. Cases and evidence stay private; this PR carries only model fixes, tests, catalog status and a short roadmap note.

Model fixes, each driven by a failing reference case, applied to the Faust export and the Python mirror with unittests:

Reference evidence: approval is now derived only from retained per-case reports that complete the current contract. A stored approved flag is no longer trusted.
Overdrive (62): the measured G1 soft-clip law replaces the tanh curve. The drive knob is value/128 (127 is full). The mod input is held on the 4-frame control grid.
Shaper (83): the five modes follow the measured polynomials (1−(1−m)³, 1−(1−m)², linear, m², m³).
CompareLev (59): the input is sampled on the 4-frame G1 control grid, so edges land where the G1 puts them.
FilterBank (32):
- The bottom band is a 4-pole lowpass and the top band a 4-pole highpass; they were bandpasses before.
- Adjacent bands alternate polarity, so they add at the crossovers instead of cancelling.
- The Python mirror now uses TPT state-variable sections, matching fi.svf. The old Chamberlin sections were unstable near Nyquist at 48 kHz.
Preflight: contract_topology_violations compared the integer section with a string, so it flagged every module in every patch. It now checks the contract's allowed_sections.

Testing

python3 -m unittest discover -s tests: 131 tests OK. tools/export_public.py into an empty directory: public tree verified.

Each case renders on the private G1-Emu target and on the Faust candidate, under the harness's default tolerances (reference-compare --candidate-engine faust). Any pass whose RMS error was more than 10% of the reference RMS was additionally held back from import. Parameter-extremes always has at least two reports. Event cases use a single Keyboard note-on, start at analysis time 0, and use transient alignment.

Module	Scenarios	Result
2Output (4)	extremes, midpoints (incl. selectors), ports, modulation, midi-events ×2	all pass → reference-compared
1Output (5)	extremes, midpoints (incl. selectors), ports, modulation, midi-events ×2	all pass → reference-compared
FilterBank (32)	extremes, midpoint, ports, modulation, reset-retrigger ×2	all pass after the fix (static within ~1.5%) → reference-compared; retrigger-2 held by the 10% guard (10.2%), scenario covered by retrigger-1
Mixer(8) (40)	extremes, midpoint, single-input port probes, modulation	all pass → reference-compared
Overdrive (62)	extremes, midpoint, ports, modulation	pass after fix (failed 5–18% before) → reference-compared
OnOff (76)	extremes, midpoint, ports, modulation, reset-retrigger ×2	all pass → reference-compared
Delay (78)	extremes, midpoint, ports, modulation	all pass (rel < 5%) → reference-compared
Shaper (83)	extremes, midpoints (selectors), ports, modulation	pass after fix (mode 4 was 26% off) → reference-compared
LevMult (111), LevAdd (112), 1to2Fade (113), 2to1Fade (114)	extremes, midpoint, ports, modulation	all pass → reference-compared
Compressor (21)	extremes, midpoint, ports, modulation	all 25 cases pass tolerances; modulation (20 Hz) is 10.6% low, held → stays approximate
CompareLev (59)	extremes, midpoint, ports, modulation	static cases pass (~1%); modulation fails edge timing → stays approximate
EventSeq (17)	—	not run; needs a multi-edge clock → stays approximate

Documented blockers (also in docs/PARITY_ROADMAP.md):

Compressor:
- Measured release is linear in dB, about 105·e^(−0.0344·knob) dB/s; the model's exponential decay is much faster.
- The detector looks peak-referenced, yet the output is flat from 8 to 330 Hz. That is unresolved.
CompareLev: modulation edges drift −5 to +3 samples over 1 s with OscA fine-pitch error, which belongs to OscA calibration.
EventSeq: every fixture clock source (ClkGen, or CompareLev fed from OscA) is still being calibrated.

preflight-disting now accepts patches built only from the twelve modules: Delay→2Output, Overdrive→2Output and FilterBank→2Output all report ok. A patch with any non-reference-compared module (OscA, Compressor, CompareLev) is still refused as unverified. Before the topology fix it refused every patch. Nothing was built: the Disting NT toolchain is not installed.

Impact Statement

Twelve modules move from approximate to reference-compared, so Disting preflight accepts them.
Four model laws change audibly toward the G1:
- Overdrive soft clip
- Shaper curves
- CompareLev control-grid timing
- FilterBank edge bands and band summation
Patches using those modules will render differently, closer to the hardware.
No tolerances were widened.
The parity matrix was not regenerated.
No cases, captures, ROM material or local paths are included.

## Scope Phase 3 of the parity roadmap: private G1 reference suites for the modules the parity sweep found clean (excluding modules owned by the v25-calibration and input-fix work). Twelve modules now complete their reference contract and are marked `reference-compared` in the catalog. Cases and evidence stay private; this PR carries only model fixes, tests, catalog status and a short roadmap note. Model fixes, each driven by a failing reference case, applied to the Faust export and the Python mirror with unittests: - **Reference evidence**: approval is now derived only from retained per-case reports that complete the current contract. A stored `approved` flag is no longer trusted. - **Overdrive (62)**: the measured G1 soft-clip law replaces the tanh curve. The drive knob is value/128 (127 is full). The mod input is held on the 4-frame control grid. - **Shaper (83)**: the five modes follow the measured polynomials (1−(1−m)³, 1−(1−m)², linear, m², m³). - **CompareLev (59)**: the input is sampled on the 4-frame G1 control grid, so edges land where the G1 puts them. - **FilterBank (32)**: - The bottom band is a 4-pole lowpass and the top band a 4-pole highpass; they were bandpasses before. - Adjacent bands alternate polarity, so they add at the crossovers instead of cancelling. - The Python mirror now uses TPT state-variable sections, matching `fi.svf`. The old Chamberlin sections were unstable near Nyquist at 48 kHz. - **Preflight**: `contract_topology_violations` compared the integer section with a string, so it flagged every module in every patch. It now checks the contract's `allowed_sections`. ## Testing `python3 -m unittest discover -s tests`: 131 tests OK. `tools/export_public.py` into an empty directory: public tree verified. Each case renders on the private G1-Emu target and on the Faust candidate, under the harness's default tolerances (`reference-compare --candidate-engine faust`). Any pass whose RMS error was more than 10% of the reference RMS was additionally held back from import. Parameter-extremes always has at least two reports. Event cases use a single Keyboard note-on, start at analysis time 0, and use transient alignment. | Module | Scenarios | Result | |---|---|---| | 2Output (4) | extremes, midpoints (incl. selectors), ports, modulation, midi-events ×2 | all pass → reference-compared | | 1Output (5) | extremes, midpoints (incl. selectors), ports, modulation, midi-events ×2 | all pass → reference-compared | | FilterBank (32) | extremes, midpoint, ports, modulation, reset-retrigger ×2 | all pass after the fix (static within ~1.5%) → reference-compared; retrigger-2 held by the 10% guard (10.2%), scenario covered by retrigger-1 | | Mixer(8) (40) | extremes, midpoint, single-input port probes, modulation | all pass → reference-compared | | Overdrive (62) | extremes, midpoint, ports, modulation | pass after fix (failed 5–18% before) → reference-compared | | OnOff (76) | extremes, midpoint, ports, modulation, reset-retrigger ×2 | all pass → reference-compared | | Delay (78) | extremes, midpoint, ports, modulation | all pass (rel < 5%) → reference-compared | | Shaper (83) | extremes, midpoints (selectors), ports, modulation | pass after fix (mode 4 was 26% off) → reference-compared | | LevMult (111), LevAdd (112), 1to2Fade (113), 2to1Fade (114) | extremes, midpoint, ports, modulation | all pass → reference-compared | | Compressor (21) | extremes, midpoint, ports, modulation | all 25 cases pass tolerances; modulation (20 Hz) is 10.6% low, held → stays approximate | | CompareLev (59) | extremes, midpoint, ports, modulation | static cases pass (~1%); modulation fails edge timing → stays approximate | | EventSeq (17) | — | not run; needs a multi-edge clock → stays approximate | Documented blockers (also in `docs/PARITY_ROADMAP.md`): - **Compressor**: - Measured release is linear in dB, about 105·e^(−0.0344·knob) dB/s; the model's exponential decay is much faster. - The detector looks peak-referenced, yet the output is flat from 8 to 330 Hz. That is unresolved. - **CompareLev**: modulation edges drift −5 to +3 samples over 1 s with OscA fine-pitch error, which belongs to OscA calibration. - **EventSeq**: every fixture clock source (ClkGen, or CompareLev fed from OscA) is still being calibrated. `preflight-disting` now accepts patches built only from the twelve modules: Delay→2Output, Overdrive→2Output and FilterBank→2Output all report ok. A patch with any non-reference-compared module (OscA, Compressor, CompareLev) is still refused as unverified. Before the topology fix it refused every patch. Nothing was built: the Disting NT toolchain is not installed. ## Impact Statement - Twelve modules move from `approximate` to `reference-compared`, so Disting preflight accepts them. - Four model laws change audibly toward the G1: - Overdrive soft clip - Shaper curves - CompareLev control-grid timing - FilterBank edge bands and band summation - Patches using those modules will render differently, closer to the hardware. - No tolerances were widened. - The parity matrix was not regenerated. - No cases, captures, ROM material or local paths are included.

peterswimm added 5 commits

2026-10-21 23:41:26 +00:00

Derive reference approval only from retained reports that complete the contract 257395e2e0

Model Overdrive and Shaper with their measured G1 transfer curves 976248d408

Hold CompareLev input on the 4-frame G1 control grid 60aa93438e

Model FilterBank edge bands as 4-pole shelves and alternate band polarity c952496924

Mark twelve modules reference-compared and fix preflight section placement check

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

Details

25711dfc1f

peterswimm merged commit abb9e9b223 into main

2026-10-22 02:14:10 +00:00