AquaSynth/state/evidence.jsonl

{"kind":"migration","lesson":"The C# repo must own reusable authoring, analysis, preset, scoring, and Faust-generation behavior; only the Rust renderer/runtime and experimental compiler branch are intentionally left behind.","source":"user direction and migration pass through commit 3246672"}
{"kind":"realtime","lesson":"Runtime-facing audio code must avoid unbounded operations: allocation, locks, I/O, console output, OS calls, and context switches do not belong in audio callbacks.","source":"PortAudio callback documentation"}
{"kind":"dsp_graph","lesson":"Treat patch scripts as deterministic dataflow graphs with explicit rate and block semantics; terse syntax is acceptable only when lowering remains inspectable.","source":"Pure Data block/subpatch model and Faust block-diagram model"}
{"kind":"sfxr_parity","lesson":"SFXR normalized filter values are not linear cutoff frequencies. In Faust export, small SFXR high-pass values must remain very low; mapping hpf directly to thousands of Hz removes the body of pickup, jump, powerup, and blip presets while still leaving them audible enough for weak tests to pass.","source":"AquaSynth engine native Faust parity test against C# SFXR reference renderer"}
{"kind":"dsl_reference_corpus","lesson":"BFXR, SFXR, 808, FM, and wobble presets are reference patches for testing and DSL design, not terseness trophies. Built-ins should be readable enough to teach the patch language, and the parser should support multi-line commands instead of forcing all expressive patches into one-line golf.","source":"user clarified preset purpose for patch DSL iteration"}
{"kind":"advanced_reference_patches","lesson":"The synth reference catalog needs complex examples that combine defaults, templates, multiple voices, modulation buses, FM, formants, noise layers, repeat, and soft clipping. These patches expose DSL readability and composition pressure better than short one-shot SFX presets.","source":"user asked for sophisticated patch references"}
{"kind":"reference_synth_roadmap","lesson":"AquaSynth should use external synth and DSP corpora as reference targets, not just internal demo patches. Start with DX7/Dexed SysEx for structured FM, then ZynAddSubFX, Faust examples/libraries, Surge XT, and Cardinal as progressively broader pressure tests. Keep provenance and license state explicit before vendoring fixtures.","source":"user clarified any synth can be broken down and rebuilt via AquaSynth DSL"}
{"kind":"runtime_patch_parameters","lesson":"Synth patches must expose expressive runtime parameters. Faust supports this through UI controls and hosted parameter APIs, so AquaSynth should compile a graph once and vary declared parameter values without recompilation. Recompile only for structural graph changes.","source":"user asked whether Faust can vary one number without recompiling"}
{"kind":"synth_test_ownership","lesson":"AquaSynth owns reference-patch correctness. AquaSynth should only smoke-test the engine synth host and bundled Faust boundary; SFXR/BFXR/reference parity belongs here so the engine can keep moving while the synth library is under active surgery.","source":"user questioned synth tests living in AquaSynth"}
{"kind":"package_boundary","lesson":"AquaSynth consumes AquaSynth.Dsl through a pinned package, not a live project reference. Breaking changes in AquaSynth are allowed during synth construction as long as a new package version is packed and AquaSynth is updated deliberately when the boundary should move.","source":"user wanted AquaSynth builds insulated from synth-agent work"}
{"kind":"runtime_patch_parameters","lesson":"AquaSynth patch scripts now have first-class `param` declarations and patch/reference model contracts. Faust emission exposes declared parameters as smoothed hslider controls, but target binding is intentionally still absent and warned about; do not claim parameters shape sound until binding is designed.","source":"parameter contract implementation on 2026-05-13"}
{"kind":"runtime_patch_parameters","lesson":"AquaSynth parameter binding uses Option B: numeric field values may reference declared parameters with `@/path`, producing exact field bindings such as `/voices/0/filter/lpf`. The field owns the meaning; the parameter owns the runtime value. Avoid target-routing parameter semantics unless a new authority is deliberately mapped.","source":"user rejected target binding as a landmine; field-site binding implemented on 2026-05-13"}
{"kind":"dx7_reference_path","lesson":"DX7 reference elaboration starts with structural extraction, not translation. The repo now parses 155-byte voice edit buffers, 128-byte packed voices, and 4104-byte packed 32-voice SysEx banks into neutral DX7 records with checksum validation and ReferencePatch features. Next pressure point is explicit algorithm topology metadata for carriers, modulation edges, and feedback edge location.","source":"DX7 SysEx parser slice on 2026-05-13"}
{"kind":"dx7_reference_path","lesson":"DX7 algorithm topology is now derived from the reverse-engineered OPS algorithm ROM table, not hand-drawn diagrams. Topology exposes carrier operators, modulation edges, feedback-register writers, self-feedback operators, delayed-feedback targets, and raw ROM steps. Use this as pressure evidence before adding an AquaSynth operator graph layer.","source":"DX7 algorithm topology slice on 2026-05-13; Ken Shirriff OPS ROM table"}
{"kind":"dx7_reference_rebuilds","lesson":"Two DX7-style AquaSynth rebuilds now make the pressure explicit: algorithm 32 maps cleanly to layered additive voices, but algorithm 8 exposes missing graph structure (operator modulation edges, summed modulators, and operator-local self-feedback). Do not patch this with more Voice fields; decide whether an operator graph model owns this invariant.","source":"DX7 rebuild catalog slice on 2026-05-13"}
{"kind":"operator_graph_model","lesson":"DX7 algorithm-8 pressure justified a first-class OperatorGraph abstraction. OperatorGraph owns FM topology: operators, modulation edges, carriers, and feedback. Do not hide this authority inside Voice fields. Remaining DX7 pressure is exact feedback-register timing and rate/level envelopes.","source":"operator graph implementation slice on 2026-05-13"}
{"kind":"operator_graph_authoring","lesson":"Operator graph authoring should use readable `operator`, `route`, and `carrier` declarations with `env=ad:...` or `env=adsr:...`. Compact `ops=`/`edges=` tuples are acceptable as importer/interchange scaffolding only, not the face of the DSL.","source":"user rejected number-soup operator syntax; readable operator graph syntax slice on 2026-05-13"}
{"kind":"note_envelope_model","lesson":"AquaSynth's general envelope model is ADSR: attack, decay, sustain level, release. Note timing belongs to `Note` records: frequency, gate duration, and host/MIDI source. SFXR sustain maps to note gate duration; SFXR punch maps to lower sustain level plus voice gain compensation so ADSR peak remains implicit 1 and the held level sits below the transient.","source":"user identified SFXR sustain as note duration and Punch as transient above sustain; punch mapping correction on 2026-05-13"}
{"kind":"faust_polyphony","lesson":"Prefer Faust-managed MIDI polyphony over AquaSynth graph inflation. AquaSynth should emit a single voice graph with standard `freq`, `gain`, and `gate` controls plus `[midi:on][nvoices:n]`; Faust architectures own MIDI decoding and voice allocation unless a concrete target proves the boundary insufficient.","source":"Faust official MIDI/polyphony docs and user preference for Faust support on 2026-05-13"}
{"kind":"operator_graph_parameters","lesson":"Readable operator graph fields bind runtime parameters at exact `/opgraphs/...` field paths: graph gain/note fields, operator ratio/level/feedback/envelope fields, and route index fields. Keep compact `ops=`/`edges=` as importer scaffolding and avoid graph-wide macro routing.","source":"operator graph field-site binding slice on 2026-05-13"}
{"kind":"dx7_envelopes","lesson":"DX7 four-rate/four-level operator envelopes can be lowered to labeled AquaSynth ADSR approximations plus gate duration, but this is not exact DX7 EG execution. Reference rebuilds should record approximation as matched pressure and exactness as still missing.","source":"DX7 envelope approximation slice on 2026-05-13; Yamaha DX7 manual describes EGs as four rates and four levels"}
{"kind":"render_parity_harness","lesson":"AquaSynth Faust output can now be compiled to Faust-generated C# and rendered into mono float samples for AudioAnalyzer comparisons. This proves the candidate side of parity; DX7 parity still requires a real Dexed/reference render or captured licensed fixture, not self-comparison.","source":"Faust render parity harness slice on 2026-05-13"}
{"kind":"reference_parity_doctrine","lesson":"External synth references exist to exercise AquaSynth DSL expressiveness, not to clone foreign internals. Consume only references that can be backed by parity tests proving AquaSynth can reproduce the behavior in terse, readable syntax; parsing a format is inventory, not evidence.","source":"user clarified reference-synth purpose and parity-test requirement on 2026-05-13"}
{"kind":"dx7_reference_rendering","lesson":"DX7 reference rendering now uses dexed-py as optional test tooling against a public-domain SysEx fixture. Keep fixtures and renderer helpers under the test project, preserve provenance and SHA-256 beside vendored banks, and keep AquaSynth.Dsl.nupkg free of SysEx/Python/test assets.","source":"dexed-py public-domain fixture harness slice on 2026-05-13"}
{"kind":"dx7_rendered_parity","lesson":"The first thresholded rendered DX7 parity target is analog1.syx voice 13, MC-MM 5-3. AquaSynth matches the rendered behavior with a short sine-like patch and explicit score/log-mel/envelope/duration/RMS/zero-crossing/centroid thresholds. Treat this as proof of the parity loop, not proof of exact DX7 operator execution; next target should audibly require operator routing.","source":"first thresholded DX7 rendered parity slice on 2026-05-13"}
{"kind":"operator_feedback","lesson":"A harder DX7 algorithm-8 target exposed direct operator self-reference as non-renderable Faust. Operator feedback now emits as delayed Faust recursion, which keeps the render path alive and is closer to DX7 feedback pressure; exact DX7 feedback scaling remains unsolved.","source":"PRC SYNTH1 parity probe and delayed feedback render fix on 2026-05-13"}
{"kind":"dx7_operator_levels","lesson":"DX7 operator output level is nonlinear and includes key/velocity offsets before envelope gain. Added a reference-import helper to approximate effective AquaSynth operator levels, but PRC SYNTH1 still only improved to ~0.34 with static levels and ~0.42 with envelope tweaks; hard parity now points at DX7-style operator envelope/gain evolution rather than static level mapping alone.","source":"DX7 effective operator level probe on 2026-05-13"}
{"kind":"staged_operator_envelopes","lesson":"Operator graphs now have a readable staged envelope surface, env=rl rates=... levels=..., backed by RateLevelEnvelope and Faust rl4_env rendering. This exercises the DSL with DX7-style multi-target contours without changing the general ADSR voice envelope. Direct DX7 rate/level lowering on PRC SYNTH1 peaked around ~0.316 in quick probes, so the next parity pressure is calibrated DX7 EG timing/gain rather than more syntax.","source":"staged operator envelope slice and PRC SYNTH1 probe on 2026-05-13"}
{"kind":"dx7_prc_listening_artifacts","lesson":"The hard PRC SYNTH1 parity probe now writes ignored WAV artifacts for ear checks under artifacts/parity/dx7-prc-synth1. A calibrated readable operator-graph candidate with staged env=rl envelopes scores ~0.458: duration is close, while RMS and timbre remain off. Use listening plus metrics to choose the next calibration target instead of blindly adding DSL surface.","source":"PRC SYNTH1 listening WAV probe on 2026-05-13"}
{"kind":"dx7_prc_calibration","lesson":"PRC SYNTH1 improved to a modest passing score of ~0.606 without new DSL surface. Removing soft clipping, using a 392 Hz graph base, and boosting the op3 carrier branch helped; global detune and per-edge route damping worsened metrics and were cut. Remaining pressure is DX7 output/index scaling or algorithm output compensation, not more envelope syntax.","source":"PRC SYNTH1 calibration pass on 2026-05-14"}
{"kind":"dx7_prc_logmel","lesson":"Listening and log-mel agreed that the aggregate PRC SYNTH1 score was misleading. The Dexed reference has strong low/body partials around 131 Hz and 393 Hz; AquaSynth's earlier candidate lacked the low body despite a better aggregate score. The probe now gates on log-mel <=0.255 with a loose aggregate floor, and remaining mismatch points at DX7 output/index scaling and algorithm compensation.","source":"PRC SYNTH1 perceptual/log-mel calibration on 2026-05-14"}
{"kind":"dx7_output_level_curve","lesson":"A project-authored Dexed single-carrier sweep shows DX7 carrier amplitude follows 2^((outputLevel - 99) / 8). The previous internal-level proxy made mid-level operators far too loud. `Dx7SysEx.OperatorOutputAmplitude` now owns the measured curve; next calibration rung is two-operator modulation index.","source":"DX7 single-carrier calibration rung on 2026-05-14"}
{"kind":"dx7_modulation_index","lesson":"Isolated two-operator Dexed FM fits full-level modulation at about 12.55 radians. Under AquaSynth's current Faust phase formula this maps to route index 6.275 for a full-level modulator. Apply this to isolated two-op branches, but do not blindly apply it to cascaded algorithm stacks before algorithm output compensation is modeled.","source":"DX7 two-operator modulation calibration rung on 2026-05-14"}
{"kind":"dx7_feedback_scale","lesson":"Isolated Dexed feedback sweeps fit a nonlinear AquaSynth feedback table: 0, .01, .02, .05, .10, .19, .38, .66 for DX7 feedback values 0-7. PRC feedback value 5 was already accidentally close via feedback*0.04; the table moves that knowledge into Dx7SysEx.OperatorFeedbackAmount.","source":"DX7 feedback calibration rung on 2026-05-14"}
{"kind":"dx7_detune_and_com","lesson":"PRC exposed two separate DX7 facts: ratio-mode detune must lower into operator frequency ratios, and ROM COM output compensation must be applied relative to the algorithm-32 six-output baseline used by the carrier-amplitude calibration. This cut the PRC carrier scale table; the remaining cascaded-route constant is still a fenced hard-target scar needing summed-modulator calibration.","source":"DX7 PRC parity rung on 2026-05-14"}
{"kind":"dx7_topology_aware_route_index","lesson":"The PRC hard target proved DX7 route index cannot key only on ROM edge kind: a direct edge feeding a later summed branch behaves like part of the cascaded/summed stack, not like the isolated two-op branch. `Dx7SysEx.OperatorRouteIndex(topology, edge)` now owns this distinction; PRC no longer has a private route function.","source":"DX7 summed/cascaded route lowering on 2026-05-14"}
{"kind":"dx7_project_authored_patch_rendering","lesson":"Project-authored Dexed patches can be generated directly inside tests from explicit operator specs. Use this for lawful calibration rungs instead of squeezing every DX7 fact out of public-domain hard patches.","source":"DX7 project-authored summed-stack parity on 2026-05-14"}
{"kind":"dx7_envelope_curve_cut","lesson":"Reusing the operator output amplitude curve for DX7 EG levels is not ready: it can improve isolated envelope shape metrics with shorter timing, but it failed the hard PRC log-mel target. Keep EG levels linear until isolated envelope parity passes without regressing PRC.","source":"DX7 envelope level experiment on 2026-05-14"}
{"kind":"dx7_envelope_trace","lesson":"Direct DX7 EG tracing shows the current `env=rl` abstraction is the wrong shape for some DX7 envelopes: the traced EG can spike above unit gain and drop below .02 within about 9 ms while `env=rl` is still near 1. A runtime `env=dx7` surface was attempted and cut because matching the local trace did not yet match Dexed plugin audio.","source":"DX7 envelope trace artifact on 2026-05-14"}
{"kind":"dx7_envelope_applied_gain","lesson":"Dexed-style operator audio does not use the raw DX7 EG state directly. It samples the EG once per 64-sample block and linearly interpolates applied gain across the block. A DX7-capable AquaSynth envelope must model that applied-gain timebase; matching raw EG trace alone can make a candidate far too quiet or otherwise wrong.","source":"DX7 interpolated envelope trace artifact on 2026-05-14"}
{"kind":"curved_rate_level_envelope","lesson":"AquaSynth staged operator envelopes now support per-segment curves and deliberate levels above unity. Keep this authority in `RateLevelEnvelope`: it protects the invariant that multi-target operator contours can model exponential synth envelopes without inventing one-off DX7 syntax.","source":"Curved rate/level envelope implementation on 2026-05-14"}
{"kind":"dx7_applied_envelope_lowering","lesson":"`Dx7SysEx.ApproximateAppliedRateLevelEnvelope` improves the hard PRC target by tracing Dexed-style block-interpolated EG gain, normalizing it against AquaSynth operator output level, and emitting curved `env=rl`. Latest PRC metrics: log-mel 0.21931955, envelope distance 0.10271241, RMS ratio 1.0003664, score 0.5026183. Zero-crossing worsened to 0.28642866, so contour parity exposed a remaining oscillator/phase mismatch rather than solving DX7 parity outright.","source":"DX7 applied-envelope PRC probe on 2026-05-14"}
{"kind":"dx7_applied_envelope_peak_picker","lesson":"Listening caught a real lowering bug: the first applied-envelope approximation treated the first 64-sample block as the attack peak, which made slower rising op2 in PRC swell late (`0.014 -> 0.613`) instead of hitting early and decaying. Picking the first near-peak before gate gives op2 `0.981 -> 0.843 -> 0.660` and improves PRC to log-mel 0.17186502, RMS 1.0000212, zero-crossing 0.7815517, score 0.6501268.","source":"PRC SYNTH1 listening-guided envelope fix on 2026-05-14"}
{"kind":"dx7_summed_route_scale","lesson":"Listening identified missing harsh stacked-modulation brightness. Raising `Dx7SysEx.SummedOperatorModulationRouteIndex` from 1.6 to 2.4 improves PRC to log-mel 0.16794802, envelope distance 0.09510617, centroid ratio 1.1254044, RMS ratio 1.00095, score 0.68491656, and also tightens the project-authored algorithm-8 summed-stack log-mel when graph gain is normalized. Keep isolated direct branches at 6.275; the stronger 2.4 scale belongs only to summed/cascaded DX7 branches.","source":"DX7 stacked modulation brightness calibration on 2026-05-14"}
{"kind":"dx7_summed_route_scale_hotter","lesson":"Follow-up listening and band analysis showed the previous summed/cascaded scale still sounded underwater. Raising `Dx7SysEx.SummedOperatorModulationRouteIndex` to 6.0 and normalizing PRC graph gain to 0.25 improves PRC to log-mel 0.13874426, envelope distance 0.067291506, RMS ratio 0.9912135, zero-crossing ratio 0.95325506, centroid ratio 1.1156512, score 0.7546445. A hotter 7.0 scale did not buy enough log-mel improvement and worsened RMS/score, so 6.0 is the current coherent stop.","source":"DX7 stacked modulation brightness hot sweep on 2026-05-14"}
{"kind":"patch_library","lesson":"Repo-owned AquaSynth .aqua stock now lives under patches/ with library.yaml; every .aqua must parse/export in tests and the patch library must remain out of AquaSynth.Dsl.nupkg unless intentionally promoted.","source":"Patch library folder slice on 2026-05-14"}
{"kind":"dx7_algorithm8_split_route_parity","lesson":"The DX7 summed/cascaded route scale 6.0 is now backed by split project-authored algorithm-8 rendered probes, not only PRC or one combined stack. Latest real-Dexed run: combined stack log-mel 0.04029555 score 0.8664546, summed pair log-mel 0.042429477 score 0.870778, cascade log-mel 0.03811625 score 0.9157809. Keep isolated direct branches at 6.275 and summed/cascaded branches at 6.0 until broader evidence contradicts it.","source":"Split algorithm-8 parity probes on 2026-05-14"}
{"kind":"patch_library_taste","lesson":"Not every parity probe belongs in the patch library. Harsh project-authored DX7 calibration probes should stay in tests and ignored listening artifacts unless they are musically useful stock; the navigable .aqua library is for reusable inspiration, not measurement torture signals.","source":"User listening correction on 2026-05-14"}
{"kind":"dx7_community_voice_parity","lesson":"Actual public-domain analog1.syx community voices are useful broader probes, but listening rejected the first loose gate. ANLGSYN 1, RES SYNTH1, DX1 LEAD B, and MELLOWSOLO should remain pressure until their timbre/zero-crossing mismatch is understood. Only Mooger and fixed Piano Bass are currently defensible rendered community parity targets.","source":"Community DX7 analog1.syx survey on 2026-05-14"}
{"kind":"dx7_voice_transpose","lesson":"Piano Bass in analog1.syx proved imported DX7 voices must lower the voice transpose byte into graph frequency. Transpose 24 is neutral; transpose 12 lowers MIDI note 60 to 130.8128 Hz. Hardcoding 261.6256 made Piano Bass a full octave too high. After the fix, Piano Bass improves to log-mel 0.16180529, zero-crossing 0.905511, score 0.64267576. Do not call the broader analog1.syx survey passing; only Mooger and fixed Piano Bass are currently defensible.","source":"Piano Bass listening correction on 2026-05-14"}
{"kind":"dx7_fixed_frequency","lesson":"ANLGSYN 1 proved DX7 fixed-frequency operators must lower as absolute Hz divided by the graph base frequency, not as note ratios. Dexed fixed mode uses a log-frequency formula; op1 fixed 0/0 is about 1 Hz and op3 fixed 0/8 detune 14 is about 1.207 Hz, yielding ratios 0.007615 and 0.009227 against the transposed 130.8128 Hz graph base. This restores the low buzzing modulation and improves ANLGSYN 1 to log-mel 0.19427659, zero-crossing 0.94662774, score 0.659334.","source":"ANLGSYN 1 listening correction on 2026-05-14"}
{"kind":"dx7_max_feedback","lesson":"ANLGSYN 1 still sounded too 'woooh' after fixed-frequency lowering because high-band buzz was missing. Route-index sweeps did not help; increasing only DX7 feedback value 7 from 0.66 to 2.2 restored much more 1.2-5 kHz energy while leaving feedback value 5 at 0.19 for PRC. Latest ANLGSYN 1: log-mel 0.17553389, centroid 0.9519112, zero-crossing 0.90904, score 0.6448498. Envelope/RMS worsened, so this is a timbre correction, not final ANLGSYN parity.","source":"ANLGSYN 1 max-feedback listening correction on 2026-05-14"}
{"kind":"dx7_output_nonlinearity","lesson":"ANLGSYN 1 listening caught a remaining 'deeooh' failure: the max-feedback candidate had a close attack but lost sustained high-band punch halfway through. Enabling AquaSynth's existing output soft clip only for max-feedback self-feedback DX7 voices improved ANLGSYN to score 0.70946133, log-mel 0.18456098, envelope 0.10628832, RMS 0.9816304 while PRC stayed passing with soft clip disabled at score 0.7546429 and log-mel 0.13874851. Do not apply the nonlinearity globally; PRC proves that would be a compensator.","source":"ANLGSYN 1 sustained-buzz parity correction on 2026-05-14"}
{"kind":"dx7_max_feedback_envelope_floor","lesson":"Output soft clip alone did not fix ANLGSYN 1's mid-note drift. The self-feedback source operator was under-driven after the attack; preserving a 0.9 applied-envelope sustain floor for max-feedback source operators lifted the sustained 2.5-5 kHz band ratio from about 0.175 to above 1.0 and improved ANLGSYN 1 to score 0.7245976, log-mel 0.18460114, envelope 0.119950555, zero-crossing 0.97149533. Add sustained band gates when broad log-mel allows audible drift through.","source":"ANLGSYN 1 sustained high-band parity gate on 2026-05-14"}
{"kind":"dx7_com_compensation_cut","lesson":"ROM COM carrier compensation was louder but harmonically wrong for the analog1 community pass. User listening preferred the no-COM renders across patches; the real issue was that they were too quiet. Cut hidden per-carrier COM boosts: `OperatorOutputCompensation` returns unity and graph gain owns loudness recovery. Passing community gains with no COM are Mooger 0.75, Piano Bass 0.90, RES SYNTH1 0.75. DX1 LEAD B and MELLOWSOLO sound closer by ear with no COM but remain numeric pressure targets.","source":"community DX7 listening correction on 2026-05-14"}
{"kind":"dx7_effective_note_scaling","lesson":"DX7 operator key-scaling and rate-scaling must lower against the effective played MIDI note after voice transpose, not hardcoded MIDI 60. `Dx7SysEx.OperatorRateScaling` now feeds the block-interpolated applied envelope trace, and imported operator levels use the effective note. This improves RES SYNTH1 attack/log-mel and Piano Bass harmonic metrics; Piano Bass graph gain rebalances from 0.90 to 0.72. MELLOWSOLO remains the clearest harmonic mismatch and now has a pressure artifact test rather than a fake parity pass. A trial output-level-dependent applied-envelope trace was cut because it worsened MELLOWSOLO and ANLGSYN 1.","source":"community DX7 listening correction on 2026-05-14"}
{"kind":"dx7_modulator_peak_headroom","lesson":"Listening found ANLGSYN 1 and RES SYNTH1 blowing out into a drone at the modulator-envelope peak while Piano Bass was spot on. The surviving DX7 probe change caps only algorithm-2 non-carrier applied-envelope Level1 at 0.92, preserving Piano Bass and improving RES SYNTH1 while keeping ANLGSYN judged by log-mel and sustained high-band evidence rather than aggregate score. A wider summed-route headroom experiment for Mooger fan-in was tested and cut because it worsened Mooger metrics and did not isolate the overtone-emphasis issue.","source":"community DX7 listening correction on 2026-05-14"}
{"kind":"dx7_mellowsolo_tail_window","lesson":"MELLOWSOLO pressure artifacts need a longer render window than 1.0s because the AquaSynth candidate emits a 0.6s release from a 0.65s gate, so the old WAV cut the tail before it could finish. Extending only the MELLOWSOLO pressure render to 1.25s improves log-mel from about 0.429 to 0.347 and makes the tail audible without changing the synth lowering or passing parity targets. A global traced-release-duration experiment was tested and cut because it shortened releases too aggressively and broke PRC, Piano Bass, RES SYNTH1, and ANLGSYN 1. RES SYNTH1 has oscillator sync enabled, so its phase-like attack issue is more likely relative operator attack/peak shape than free-running oscillator phase.","source":"community DX7 listening correction on 2026-05-14"}
{"kind":"dx7_graph_pitch_lfo","lesson":"ANLGSYN 1's missing beat is DX7 pitch LFO, not amplitude LFO: raw analog1.syx has AMD=71 but all operator AMS values are 0, while PMD=32/PMS=1/sine speed 38/delay 33 are active. AquaSynth operator graphs now own graph-level vibrato (`vibrato`, `vibrato_hz`, `vibrato_delay`) and the DX7 probe lowers that LFO to about depth 0.003694 at 5.993 Hz with a 0.233s fade-in. A hard delay gate was tested and cut because it pushed log-mel over the ANLGSYN gate; fade-in keeps ANLGSYN passing at log-mel about 0.19497 and score about 0.6979.","source":"ANLGSYN 1 pitch-LFO beat correction on 2026-05-14"}
{"kind":"zyn_reference_start","lesson":"The ZynAddSubFX phase should start with structural feature extraction, not translation. Upstream Zyn instruments live in a separate bank submodule and the preset provenance is not clear enough to vendor casually, so the first parser slice uses project-authored .xiz fixtures. `ZynInstrumentReader` reads plain or gzip-compressed XML and classifies active ADD/SUB/PAD engines, kit layering, envelopes, free envelopes, LFOs, filters, formants, and effects into ReferencePatch features.","source":"ZynAddSubFX feature extraction kickoff on 2026-05-14"}
{"kind":"zyn_reference_rebuilds","lesson":"The first Zyn-style AquaSynth rebuilds can sketch additive lead, PAD texture, and vocal/formant layer fixtures with existing voices, macros, formants, filters, noise, and modulation, but the missing abstractions are now concrete: additive harmonic-bank syntax, PAD/spectral source generation, free envelopes for normal voices, named kit/layer routing with per-layer effect sends, and richer formant/vowel morphing. Keep these as named pressure before adding new DSL surface.","source":"ZynAddSubFX rebuild catalog slice on 2026-05-14"}
{"kind":"zyn_worst_case_survey","lesson":"Scanning the upstream ZynAddSubFX instrument bank at commit e9f64a9 found 1358 .xiz files and the hardened reader now parses all of them. The nastiest target is `olivers-100/0032-Drum Kit.xiz`: 16 enabled kit items, mixed ADD/SUB/PAD engines, 106 envelopes, 25 free envelopes, 73 LFOs, 37 filters, and 3 formant filters. The broader top list says named kit/layer routing should come before syntax golf for additive banks, PAD sources, or formant motion; otherwise new abstractions will have no coherent owner.","source":"ZynAddSubFX pressure survey on 2026-05-14"}
{"kind":"layer_language_golf","lesson":"The first Zyn-driven language-golf abstraction is `layer`: named layer declarations own engine tags, optional MIDI key range metadata, default gain, and effect-send labels, while lowering to ordinary AquaSynth voices for now. The abstraction generalizes across additive, PAD-texture, and vocal/formant rebuilds and gives future additive-bank, PAD-source, free-envelope, and formant-motion syntax a coherent owner. It deliberately does not implement those later features yet.","source":"Zyn named layer DSL slice on 2026-05-14"}
{"kind":"harmonic_bank_language_golf","lesson":"Additive `harmonics` syntax now records authored ratio/gain partial banks under named layers and lowers them to ordinary AquaSynth voices. This generalizes additive lead and future drawbar/partial-bank targets without hiding Zyn ADDsynth phase, bandwidth, oscillator shaping, or free-envelope semantics behind a fake parity claim.","source":"Zyn harmonic-bank DSL slice on 2026-05-14"}
{"kind":"voice_rate_level_envelopes","lesson":"The existing readable `env=rl rates=... levels=... curves=... gate=...` contour now applies to normal voices and therefore to named layer defaults, not only operator graphs. This gives PAD and vocal Zyn rebuilds staged non-ADSR onset/release shape while preserving the invariant that Note owns gate timing and envelope data owns level shape. Arbitrary Zyn free-envelope point curves remain missing pressure.","source":"Zyn staged voice envelope DSL slice on 2026-05-14"}
{"kind":"pad_spectral_cloud","lesson":"PAD-style layers now have `spectrum layer=... root=... spread=... partials=ratio:gain,...`, recorded as `SpectralBank` and generated as PADsynth-style FFT tables: frequency-domain harmonic spreading, deterministic random phase, one inverse FFT, normalization, then Faust waveform/rdtable playback before ordinary voice treatment. This gives PAD texture rebuilds a real spectral source authority without inflating the voice graph; roots/spreads/partials are authoring-time table shape, not runtime controls. It still does not claim full Zyn PAD harmonic-profile/randomness/pitch-zone behavior.","source":"Zyn PAD spectral-cloud DSL slice on 2026-05-14"}
{"kind":"zyn_reference_oracle","lesson":"ZynAddSubFX source is pinned as GPL test/development material under `external/zynaddsubfx` at 3ab608c432996ba4d582176572c0b0f82328c825. Treat it like dexed-py: a parity oracle for tests and artifacts, not AquaSynth implementation, runtime dependency, or NuGet package content. Before further Zyn DSL golf, build a test-only PAD renderer/wrapper from this source and compare AquaSynth rebuilds against actual Zyn behavior.","source":"Zyn source pinning for parity on 2026-05-14"}
{"date":"2026-05-14","area":"zyn-reference","lesson":"Pinned Zyn can build on Windows through MSYS2/MinGW when GUI/plugins are disabled, tests are enabled, Werror is off, and the stale non-async MiddleWare Part constructor call is patched. Apply Zyn compatibility patches to ignored artifact source copies, not the clean pinned submodule.","evidence":"tools/zyn-reference/build-zyn-reference.ps1 copies external/zynaddsubfx to artifacts/zyn-source-msys, applies tools/zyn-reference/patches/0001-msys-part-constructor.patch, builds PadNoteTest, and ctest -R PadNoteTest passes."}
{"date":"2026-05-15","area":"zyn-pad-parity","lesson":"The Zyn PAD oracle now emits real reference material through Zyn PADnoteParameters/PADnote, and the first project-authored PAD comparison is not close: log-mel 0.552296, envelope 0.774501, RMS ratio 0.234763, centroid ratio 1.025735, score 0.236117. Treat this as evidence to isolate source table generation, playback/gain, envelope, and fixture mapping before adding more Zyn DSL syntax.","evidence":"tools/zyn-reference/ZynPadReference.cpp plus tests/AquaSynth.Dsl.Tests/ZynReferenceParityTests.cs write artifacts/parity/zyn-pad-reference/pad-texture-zyn.wav, pad-texture-aqua.wav, and report.txt when the Zyn artifact renderer is built."}
{"date":"2026-05-15","area":"zyn-pad-parity","lesson":"User listening correctly identified the first AquaSynth PAD rebuild as too noisy. The fix was architectural cleanup before new syntax: cut unearned noise/drive/soft-clip from the rebuild, interpolate spectral table reads, and use Zyn-sized 131072-frame PAD tables. Score improved from 0.236117 to 0.487325, with RMS ratio 1.008516 and centroid ratio 1.083082. Remaining mismatch is source profile/table generation and envelope semantics, not decorative layers.","evidence":"artifacts/parity/zyn-pad-reference/report.txt after the de-bee pass; full dotnet test passed 96/96."}
{"date":"2026-05-15","area":"zyn-pad-parity","lesson":"The octave-low Zyn PAD mismatch came from conflating spectral table root with note playback frequency. `spectrum root` must own table/base frequency, while `freq` owns the played note. Using Zyn's sample basefreq 77.7813 with playback freq 261.6256, and removing gratuitous spread/vibrato, improved the PAD report to log-mel 0.353365, envelope 0.348752, RMS ratio 0.97283, centroid ratio 0.999243, score 0.612832.","evidence":"PatchScript now preserves spectrum freq as voice note frequency; PatchScriptTests.SpectralBankSeparatesTableRootFromPlaybackFrequency covers the invariant; artifacts/parity/zyn-pad-reference/report.txt records the improved score."}
{"date":"2026-05-15","area":"zyn-pad-parity","lesson":"Aqua's remaining detuned impression was caused by stacking same-note PAD spectral voices. Collapse body/high content into one coherent spectral table before adding layers for richness. A simple gain trim then brought loudness into line: RMS ratio 1.001704, centroid ratio 1.004072, score 0.6211.","evidence":"BuiltInScripts.ZynStylePadTexture now uses one active PAD spectrum with partials 1, 1.5, and 2 plus patch gain 1.1; artifacts/parity/zyn-pad-reference/report.txt records the focused render."}
{"date":"2026-05-15","area":"zyn-pad-fixtures","lesson":"More PAD fixtures immediately exposed which pressure is real: `pad-pure` and old `pad-texture` render identically through Zyn, so the filter/category XML in `pad-texture` was not live oracle behavior. `pad-harmonic` changes table and note output through explicit OSCIL harmonics and should be the next PAD parity rung.","evidence":"tests/AquaSynth.Dsl.Tests/Fixtures/ZynAddSubFX/ProjectAuthored/pad-pure.xiz and pad-harmonic.xiz plus artifacts/parity/zyn-pad-fixtures/report.txt."}
{"date":"2026-05-15","area":"zyn-upstream-gpl-bank","lesson":"GPL Zyn presets are acceptable as test-only feedstock. The upstream instrument submodule at `external/zynaddsubfx/instruments` commit c5c912131b31df5fdf372d2f06a25aaf2375837f provides 1318 `.xiz` files; survey and PAD render tests write ignored reports/WAVs while the NuGet boundary still excludes `.xiz`, `external/`, and artifacts.","evidence":"ZynInstrumentTests.SurveysUpstreamGplInstrumentBankWhenAvailable and ZynReferenceParityTests.ZynPadReferenceRendererSurveysUpstreamGplPadFixturesWhenBuilt; focused Zyn tests passed 14/14."}
{"date":"2026-05-15","area":"zyn-artifact-layout","lesson":"Zyn PAD listening artifacts now follow the DX7 parity shape: per-synth timestamped run folders with Zyn reference WAVs, source-table WAVs, `report.txt`, and `candidate.aqua`. Upstream GPL PAD runs write `candidate-aquasynth.wav`; this was initially the current ZynStylePadTexture baseline and is now superseded by generated per-preset PAD rebuilds.","evidence":"ZynReferenceParityTests focused run passed 4/4 and wrote folders such as artifacts/parity/zyn-upstream-pad-fixtures/0121-ghost-ensemble-20260515T072330036."}
{"date":"2026-05-15","area":"zyn-pad-translation","lesson":"Zyn PAD artifacts now use per-preset generated `.aqua` instead of the generic baseline. `ZynInstrumentReader.RebuildFirstPadAsAquaSynthScript` translates the first enabled PAD kit item into one named PAD layer plus a `spectrum` bank using OSCIL harmonic magnitudes, Zyn oracle sample `basefreq` as table root, PAD volume as layer gain, and simple rate/level envelope scaffolding. Reports explicitly list missing Zyn semantics: bandwidth/profile spreading, harmonic profile shaping, and filter envelope/LFO lowering.","evidence":"ZynInstrumentTests.RebuildsProjectAuthoredPadAsAquaSynthSpectrumScript plus focused Zyn tests passed 13/13; latest upstream PAD reports include generated candidate scripts and log-mel scores from ~0.276 to ~0.637."}
{"date":"2026-05-15","topic":"zyn-pad-profile-table-semantics","lesson":"Zyn PAD bandwidth/profile/position fields belong on SpectralBank as table-generation data, not in parity side channels. Mapping those fields and Zyn's exponential PAD volume curve improved gain/envelope evidence, but profile math alone did not solve log-mel parity.","evidence":"Focused Zyn tests passed 13/13 after adding PadSpectrumProfile and zyn_* script fields; upstream PAD batch still has log-mel roughly 0.34-0.69 depending on fixture."}
{"date":"2026-05-15","topic":"zyn-oscilgen-approximation","lesson":"Expanding a subset of Zyn OscilGen base_function plus spectrum_adjust is conceptually the right owner for one-harmonic PAD presets, but the first approximation is mixed and must not be treated as solved parity.","evidence":"Soft Pad log-mel improved to 0.342738 and Organ Choir Pad2 to 0.573216, while sin2x pad regressed to 0.524421, DoublePadBass to 0.689315, and Ghost Ensemble to 0.482155. Next cut needs more OscilGen semantics or gating."}
{"date":"2026-05-15","topic":"zyn-oscilgen-stage-order","lesson":"Zyn OscilGen partial extraction should only run the sampled waveform pipeline when oscillator processing stages are live; otherwise direct harmonics stay readable and capped. A broad base-function expansion made some presets worse without earning its complexity.","evidence":"After adding sampled filter/waveshape/modulation stage order and restoring a capped direct path, sin2x pad improved to log-mel 0.455193, while Soft Pad stayed 0.342738 and Ghost stayed 0.40922. DoublePadBass and Organ Choir remain poor, so this is an incremental rung, not parity."}
{"date":"2026-05-15","topic":"zyn-pad-kit-layering","lesson":"Do not treat first enabled PAD kit item as a Zyn instrument when kit_mode is multi-kit. Multi-kit PAD presets need per-kit table roots and layered Aqua spectrum banks, or parity artifacts compare against an amputated preset.","evidence":"DoublePadBass has kit_mode=1 with enabled PAD kit items 0 and 1. Rebuilding/summing both moved log-mel from about 0.655 to 0.493 and generated doublepadbass_0 plus doublepadbass_1. Raising the volume cap to 3.0 worsened log-mel to 0.521 and was cut."}
{"date":"2026-05-15","topic":"zyn-ghost-base-function","lesson":"A direct-harmonic readability gate must not erase source identity. Ghost Ensemble needed power-base OscilGen expansion even without downstream oscillator stages; otherwise it collapsed toward a generic two-partial pad.","evidence":"Ghost Ensemble has base_function=4 with two explicit harmonics. Expanding power-base multi-harmonic sources changed the candidate from two partials to 32 partials and improved score to 0.446977, with log-mel 0.402789."}
{"date":"2026-05-15","topic":"zyn-pad-static-filter","lesson":"Zyn PAD static global filter frequency is audible layer-balance authority, not optional decoration. Read nested FILTER params under FILTER_PARAMETERS and lower old Zyn cutoff values through the exponential FilterParams curve, not a linear freq/127 guess.","evidence":"DoublePadBass kit0 filter freq=12 and kit1 freq=61. Linear mapping gave lpf=0.094488/0.480315 and still sounded like all kazoo. Zyn's curve maps them to lpf=0.003325/0.047225, improving log-mel to 0.334546 and making the generated patch deep-bass-first. Filter envelopes/LFOs and level recovery remain missing pressure."}
{"date":"2026-05-15","topic":"zyn-pad-detune-and-filter-order","lesson":"Zyn PAD note frequency is not just the requested note plus table root; FREQUENCY_PARAMETERS coarse/fine detune are live playback authority, and analog LP2 needs a steeper low-pass than AquaSynth's old one-pole default.","evidence":"DoublePadBass has coarse_detune=15360, which Zyn getdetune wraps from octave 15 to -1 octave. Emitting spectrum freq=130.8128 instead of 261.6256 and lpf_order=2 moved DoublePadBass log-mel from 0.334546 after static cutoff mapping to 0.259298; remaining gap is RMS/envelope."}
{"date":"2026-05-15","topic":"zyn-pad-filter-envelope","lesson":"Zyn PAD FILTER_ENVELOPE is audible transient authority. Static cutoff parity can get the tone right while still missing the opening hit; AquaSynth needs explicit low-pass cutoff envelopes with nonzero start levels.","evidence":"DoublePadBass sounded tonally right but softer at onset. Mapping Zyn ADSR-filter octave offsets into lpf_env=rl with lpf_start improved log-mel from 0.259298 to 0.241993, envelope distance from 0.631427 to 0.554658, RMS ratio from 0.390177 to 0.456173, and score from 0.370218 to 0.404382."}
{"date":"2026-05-15","topic":"zyn-pad-output-calibration","lesson":"After PAD table roots, detune, filter order, filter envelope, and logarithmic envelope times are mapped, remaining Zyn PAD loudness can be a generated-rebuild output calibration rather than another per-layer fixture tweak.","evidence":"Current PAD fixture RMS ratios before calibration were consistently low: DoublePadBass 0.511669, Soft Pad 0.619347, Organ Choir 0.588127, Ghost 0.730155, sin2x 0.775653. Emitting patch gain=1.6 moved DoublePadBass candidate WAV peak to 0.979 against clipped Zyn reference peak 1.0, envelope distance to 0.226563, RMS ratio to 0.818669, and score to 0.631168 while keeping batch RMS plausible."}
{"date":"2026-05-15","topic":"zyn-pad-filter-q","lesson":"Zyn PAD analog filter Q is real filter damping authority and belongs as a general AquaSynth filter surface (`lpf_q`), not as a fixture-specific cutoff/gain fudge.","evidence":"Ghost Ensemble has FILTER freq_track=97 and q=28. Adding note-tracked cutoff lowered Ghost lpf from 0.369706 to 0.282776 and adding explicit lpf_q=0.499021 moved log-mel from 0.38424 to 0.375637 and RMS ratio from 1.155288 to 1.09995. DoublePadBass remained acceptable and improved to log-mel 0.247729 / score 0.650069 after Q."}
{"date":"2026-05-15","topic":"zyn-pad-q-order-cascade","lesson":"Explicit low-pass Q must not erase low-pass order. When a reference filter has both damping and slope, Faust lowering should cascade resonant stages according to `lpf_order` rather than replacing the filter with one generic resonant low-pass.","evidence":"Ghost Ensemble still sounded bright in the middle after lpf_q. Cascading resonlp stages for lpf_order=4 moved Ghost log-mel from 0.375637 to 0.369402 and RMS ratio from 1.09995 to 1.054875 while keeping DoublePadBass near its accepted score: log-mel 0.247748, score 0.650029. ZynPadReference harmonic dumps showed Ghost's first 32 imported OscilGen partial ratios already matched Zyn, so this was filter slope, not first-page harmonic extraction."}
{"date":"2026-05-15","topic":"zyn-oscil-filter-index","lesson":"Zyn OscilGen oscillator filter functions use one-based harmonic indices. Do not translate them through a zero-based AquaSynth index; that moves spectral features to the wrong partial and can make a preset sound like the wrong overtone is being emphasized.","evidence":"sin2x uses oscillator filter type 13. With zero-based indexing, AquaSynth put the spike on h3 and had centroid ratio 1.227953. After preserving Zyn's one-based index, candidate partials are h2-dominant, centroid ratio improves to 0.998145, log-mel moves from 0.491437 to 0.479782, and UpstreamSin2xOscilFilterKeepsSpikeOnSecondHarmonicWhenAvailable asserts h2 > h3 * 10."}
{"date":"2026-05-15","topic":"zyn-organ-highpass-selected-root","lesson":"Organ Choir's coherent octave blast needs real Zyn PAD high-pass ownership and selected table roots for OscilGen shift-base sources, but selected roots must not become the global default. The source rule is narrow: keep sample-0 roots unless the PAD OscilGen base function is the shift source (`base_function=7`) and the oracle provides the selected note table root.","evidence":"Organ Choir kit0 has analog HP2 with stages and FILTER_ENVELOPE, while kit1 uses base_function=7. Mapping hpf_order/hpf_env and using the selected 622.25 Hz table root only for kit1 moved the candidate's dominant peak to 522.67 Hz with 261 Hz at -1.29 dB, close to Zyn's 523.33 Hz lead and 261 Hz at -2.23 dB. Latest Organ log-mel is 0.499122 / score 0.51966, while accepted Soft Pad, Ghost, and DoublePadBass remain on sample-0 roots; global selected roots had regressed DoublePadBass."}
{"date":"2026-05-15","topic":"zyn-organ-basefunc-mod-adaptive","lesson":"Organ Choir's violin-vs-wind mismatch was OscilGen source-table ownership, not pitch, envelope, or gain. Zyn `base_function_modulation` warps the phase before base-function sampling, and `adaptive_harmonics` remaps the prepared oscillator spectrum at the PAD table root; both must happen before PAD bandwidth/profile spreading.","evidence":"Before the fix, Aqua kit1 partials formed a smooth ramp (h3 -3.5 dB, h4 -5.1 dB, h8 -10.2 dB relative to h1) while Zyn's harmonic dump was sparse (h3 -27.1 dB, h4 -70.3 dB, h8 -50.2 dB). Applying base_function_modulation=1 and adaptive_harmonics=1 moved kit1 table energy to 400-800 Hz 0.2713 vs Zyn 0.2780 and 8-16 kHz 0.1471 vs Zyn 0.1355; Organ log-mel improved to 0.471483 and score to 0.541933 without changing accepted DoublePadBass, Ghost, or Soft Pad."}
{"date":"2026-05-15","topic":"zyn-pad-ear-vs-logmel","lesson":"For sustained Zyn PAD fixtures, log-mel distance can remain pessimistic after the perceptual target has been reached. Keep it as a regression and triage signal, but do not keep grinding accepted fixtures solely to chase a DX7-like numeric threshold when listening and structural source-table evidence agree.","evidence":"After mapping selected roots, HP filter motion, Q/order cascades, one-based OscilGen filter indices, base-function phase modulation, and adaptive harmonics, the user accepted DoublePadBass, Ghost Ensemble, Soft Pad, Organ Choir, and the other recent community PAD fixtures by ear. Organ Choir still reports log-mel 0.471483 despite direct kit1 table energy matching the Zyn body/air balance and the listener judging all fixtures equivalent."}
{"date":"2026-05-16","topic":"pad-syntax-generalization","lesson":"Reference-synth names must not become canonical AquaSynth syntax or public patch-model vocabulary. Importers may translate Zyn controls, and parser aliases may keep old parity artifacts readable, but generated/authored `.aqua` should use AquaSynth-native `pad_*` terms for general spectral-table intent.","evidence":"Zyn PAD parity had produced generated scripts with `zyn_mode`, `zyn_bandwidth`, `zyn_bwscale`, `zyn_profile`, and `zyn_position`, plus public model names like `ZynHarmonicProfile`. The generalization pass changed generated scripts to `pad_mode`, `pad_bandwidth`, `pad_bwscale`, `pad_profile`, and `pad_position`, renamed the public profile/position records and enums to PAD-native names, preserved legacy `zyn_*` parser aliases, and passed `dotnet test AquaSynth.slnx --no-restore` with 108 tests."}
{"date":"2026-05-16","topic":"project-rename-aquasynth","lesson":"The C# synth repo is now AquaSynth and the Rust reference branch is AquaSynth-rs. Keep repo wiring, package metadata, namespaces, solution/project names, env vars, docs, and state aligned to AquaSynth; do not reintroduce AquariumSynthCSharp/AquariumSynth.Dsl names.","evidence":"Renamed solution to AquaSynth.slnx, projects to src/AquaSynth.Dsl and tests/AquaSynth.Dsl.Tests, namespaces/package to AquaSynth.Dsl, package version suffix to 0.1.0-aquasynth.1, DX7 env var to AQUASYNTH_DX7_PYTHON, and origin to https://github.com/GameCult/AquaSynth.git. Verified `dotnet build AquaSynth.slnx --no-restore`, `dotnet test tests\\AquaSynth.Dsl.Tests\\AquaSynth.Dsl.Tests.csproj --no-build -v minimal` with 108 passed, and `dotnet pack src\\AquaSynth.Dsl\\AquaSynth.Dsl.csproj -c Release --no-restore` producing AquaSynth.Dsl.0.1.0-aquasynth.1.nupkg."}