InterfaceRL.{hpp,cpp,tpp} + InterfaceRLFileFormat.hpp are the upstream
reference implementation of the entire NISPS feedback subsystem —
nisps/ml/{geo_push,replay,feedback,jolt,ou_noise}.hpp are all ports of it,
several still carrying `// upstream InterfaceRL.hpp:NNN` line references.
The Phase-4 vendoring dropped examples/ because nothing compiled it. That
was correct for the build and wrong for the codebase: with the source of
truth out of tree, upstream redesigned the geometric dislike (deleted the
/(1+len) taper, doubled kGeometricPushScale, tripled the negative-LR base,
moved to batch training over all negatives every tick) and we did not notice
for months.
Copied verbatim from memllib @ e291192 — the same commit the rest of the
vendored tree pins — into lib/memllib/reference/, which sits OUTSIDE src/
and is therefore never compiled: PlatformIO's LDF only recursively builds an
Arduino-format library's src/ folder. Verified: slpworkshop still builds
(RAM 28.8%, flash 2.1%).
reference/README.md states the two rules (never compiled, never edited — a
divergence from upstream is a recorded decision, not an edit here) and
VENDORED.md's "what was dropped" section now tells the truth.
Resolves ALIGNMENT defect 6c.
Upstream memlp (github.com/MusicallyEmbodiedML/memlp @ ea777502, the commit
upstream/main pins) applies gradients with RMSProp everywhere: Layer.h:239
ApplyAccumulatedGradients, the m_sq_grad_avg running average at Layer.h:601,
StaticMLP.h:268. nisps/ml/training.hpp shipped SGD only and filed the
difference as an optimiser-choice research question. It was not one.
RMSProp divides each step by the running gradient magnitude, so an upstream
lr is a NORMALISED step; under SGD the same number multiplies the raw
gradient. Every learning rate ported from upstream therefore landed in an
optimiser that reads it differently — most visibly feedback.hpp's
`geo_lr_ = 0.001f // upstream InterfaceRL.hpp:312`, an RMSProp LR pasted
into a single SGD step.
rmsprop_step() ports Layer.h:239 exactly: clip at +/-10, sq = min(0.9*sq +
0.1*g^2, 1e6), adj = min(lr/(sqrt(sq)+1e-6), 1.0), w -= adj*g. The
adjusted-LR clamp stays one-sided as upstream's std::min is, so the negative
lr used by train_targets' "train away from this target" path behaves as it
does upstream. The per-weight squared-gradient average is new persistent
state and lives in the storage policies (FixedStorage arrays /
DynamicStorage arena) so nisps/ stays allocation-free and the firmware's
zero-heap contract holds. It is optimiser state, not model state: excluded
from weight_count()/get_weights()/set_weights(), matching upstream, and
cleared by MLPCore::reset_optimizer_state() (upstream ResetOptimizerState).
draw_weights() deliberately does NOT clear it — upstream's DrawWeights
doesn't either.
Measured with tests/cpp/ml_bench.cpp:
D1 one geometric dislike moves the mapping 1.6e-2, up from 5.3e-5 (~295x),
and repeated presses now CONVERGE on the intended 0.5 push (0.12 at 10,
0.56 at 100) instead of creeping linearly forever.
A4 geometric-vs-Diffuse gap narrows from ~4100x to ~14x in one press.
U4 the upstream-LR positive path actually trains now (range_util 0.71 at
100 ticks/gesture, was 0.016 — it was inert under SGD).
Not fixed by this, and now tracked as ALIGNMENT defect 6d: the dose
asymmetry. lurch_max is still ~1.08 against a [0,1] output range.
Golden vector stages 2 and 3 re-captured; stages 0 and 1 are pre-training
and did not move, which is the cross-check that only the update rule
changed. manifold/public/nisps.wasm rebuilt so parity-check compares like
with like — it FAILED at up to 5e-2 against the stale artifact and PASSES at
2.4e-7 against a fresh one. parity-check.sh only builds the WASM when it is
missing, never when it is stale; noted in MAP.md and filed separately.
ALIGNMENT defect 6 resolved (moved to Recently resolved); 6b's optimiser
cross-reference updated; new defect 6d for the positive-training dose.
Gates: build-cpp-tests 138 tests / ctest 4/4, parity-check PASS, lint-cpp
clean, manifold typecheck clean.
NISPS is a controller, not a synth: the object of study is the mapping
f: control-space -> parameter-space and how a musician's gestures deform
it. Loss measures fit to points the user dictated, which is the one thing
they never experience. So this measures geometry and gesture-response.
tests/cpp/ml_bench.cpp 61 scenarios, REPORTS never asserts (same
discipline as engine_bench.cpp). Shape-
agnostic via MLPCore<DynamicStorage>
(--shape, default 2,16,16,16,8), seeded
RNG throughout, branch points replayed
from scratch rather than snapshotted.
tests/cpp/test_ml_behaviour.cpp 20 asserting invariants, wired into
nisps_core_tests.
scripts/bench-ml.sh native + WASM from one source; --compare,
--sweep-shape, --smoke, --scenario, --seed.
Documents two contracts that fail SILENTLY (both now pinned by tests):
a thumbs-up must call BOTH mlp.add_example() and fb.store_positive(),
since dislike_geometric k-NNs the replay buffer and not the MLP dataset;
and placed_output() is valid only while state == Placing, after which an
empty span whose l2() is 0 scores a broken lifecycle as a perfect place.
ALIGNMENT defect 6 re-ranked (SGD-vs-RMSProp is not a research axis - it
silently invalidated every ported hyperparameter) and split into 6b (the
geometric dislike was ported from a superseded upstream design) and 6c
(InterfaceRL, the reference impl, is not in the tree).
Gates: build-cpp-tests (138 tests, ctest 4/4), parity-check PASS,
lint-cpp clean, bench-ml.sh --smoke runs end to end.
Four items from one workflow, committed together because their build and CI
wiring genuinely interleaves — nisps/CMakeLists.txt, run-all-tests.sh and
ci.yml each carry hunks from two of them, and the stage renumbering (1/5 ->
1/6) touches every line. Splitting would produce commits that do not build,
which is worse than a commit that does four things and says so.
S26 part 2 — the curve declaration now matches reality. params[].curve stays
the mode-wide DEFAULT; a voice_spaces entry may now be {name, curve_overrides}
declaring only the slots where THAT voice space deviates. The 6 modes with one
voice space are byte-identical. The values were derived MECHANICALLY by a new
codegen/curve-audit.ts that models the four idioms a p[N]*p[N] regex misses
(alias form, memlcelium's implicit-counter sq() lambda, loop-generated indices,
smooth_params_), inlines helpers, and RAISES rather than guessing when it
cannot reduce an expression. A drift gate cross-checks 1179 (voice space x
param) slots against engine source on every run and was proved to fail loudly
on three drift classes. Application stays in the engine: nisps/engines,
nisps/pipeline and nisps/core are untouched, generated output is pure insertion
(755 insertions, 0 deletions), and the rebuilt nisps.wasm was byte-identical.
S4 / 7.2 — firmware reads the active mode's driver config at mode start, and
mic/line is real. My brief assumed the engine owns this; the code disagreed and
the code was right. sound_analysis_midi's EngineT is NoOpEngine — the mic lives
on a separately-composed AnalysisEngine member — so engine-level wiring would
have compiled, passed every gate, and left the one mic mode on line input.
Hence a mode-level seam defaulting to engine().driver_config(). Separately,
DriverConfig's defaults (line_level 0, output_volume 1.0) had drifted from
memllib's actual 3/0.8 because nothing had ever read them; wiring them as-is
would have made every silent mode louder and its line input maximally
insensitive — a behaviour change disguised as plumbing. Now pinned by a test.
Also: GetSysClockSpeed() panic()s on unsupported sample rates and runs on the
first line of setup(), so sample_rate needed a fallback ahead of clock setup.
CI's firmware env list gains soundanalysismidi — it is the only mic variant and
nothing else compiles that path.
Plan 5e — telemetry is real. A loss_history C-API entry across the full 5-layer
chain lets the browser read the per-iteration loss the core already records.
The audit named one fabrication site; there were two — wasm-iml.ts's
synchronous train() published lossHistory: [loss] as well. A third, ctx.loss,
was not merely dead but actively synthetic (fallbacks of prev * 0.82 and a
literal 0.5, rendered by nothing) and is deleted. The firmware buffer stays
untouched, per the L25 call. EngineApi.lossHistory() reads spine state rather
than the MLP handle, because trainAsync() fits on the worker's mirror net and
the handle would give a subtly-wrong second answer.
Plan 5f — engine throughput is measurable. One source compiled twice (CMake
natively, emcc for WASM) so the targets compare directly and no WASM export is
added. Sequencers are driven into a working state, and every row prints its own
working-state evidence so a number produced by an idle engine is visible rather
than plausible. Reports, never asserts: a wall-clock threshold on shared
hardware is meaningless or flaky, same call as the firmware size job.
ALIGNMENT: the telemetry defect is deleted (built, not deferred); the
performance defect is rewritten to what is actually left — these are HOST
numbers, and nothing measures the RP2350 at 150 MHz, which is the target the
mission's constraint is about. Q4 (memllib ownership) and Q5 (legacy feedback
modes) are closed.
Corrections to my own earlier claims, both found by agents contradicting the
brief: manifold/ONBOARDING.md was NOT "now accurate" — its primitives list
still named five deleted primitives and cited a seededGradient() that does not
exist. And the parity harness misses the sequencer engines because it runs 128
frames while their sequencers evaluate every 400-500 samples, NOT because
all-params-0.5 fails to trigger them (it does trigger: 0.5 maps to ratio 2,
firing three times per bar). The fix is a longer window, not different params.
Gates: run-all-tests.sh ALL GREEN — 4/4 ctest, parity PASS, lint clean, curve
drift 1179 slots ok, 39 e2e (was 33). Firmware: 5 envs built including the mic
variant.
One cut, no dual path. Closes ALIGNMENT defect 3 ("Arduino-CLI build
machinery is actively hostile") and vision bullet 4.
platformio.ini carries 16 [env:], one per variant, each passing
-DMEMLNAUT_MODE_TYPE; selftest passes -DNISPS_SELFTEST=1 instead. The env list
IS the registry now — the .ino comment-registry and the NISPS_ST_* token-paste
table are deleted rather than migrated. L12 noted that table was already
silently missing the currently-shipped SLPWorkshop variant, which is the whole
argument against having a second list.
Also deleted: the Python/sed machinery that rewrote the COMMITTED .ino on every
build, the sketch symlink forest, the global TFT_eSPI User_Setup.h mutation
(now -D flags — TFT_eSPI's own documented PlatformIO recipe), the UF2
boot-mount detection stack (upload_protocol=picotool talks to the bootloader
directly), and build-firmware-arch.sh entirely. Scripts 683 -> 435 lines.
memllib is vendored at lib/memllib/ from upstream e291192; no submodules
remain. VENDORED.md records provenance and the re-sync procedure.
S9: a firmware-build CI job compiles three representative envs against a cached
toolchain and reports per-variant flash/RAM. Firmware is in an automated gate
for the FIRST time. The old ci.yml comment justified excluding it as "low
verification value" — an assessment that did not survive contact, since the
SelfTest variant sat broken for an unknown period calling a DisplayDriver
method that did not exist at the pinned memllib commit, and nothing noticed
because nothing built it.
Verified: all 16 envs build from an empty cache, each within ~520 bytes of the
arduino-cli binary it replaces, flash and RAM. Measured as .text+.rodata /
.data+.bss+vector+uninitialized — NOT PlatformIO's console line, which
double-counts .data on this board. This does not prove the hardware boots; no
flash+smoke test was possible and that stays an operator chokepoint.
slpworkshop 248232/145028 pafsynth 256880/149716 selftest 216228/17960
(all 16 in the CI log format; none exceeds 2% of a 16 MB flash)
Two traps recorded so nobody rediscovers them: vendoring memllib's subdirs
without a src/ wrapper makes PlatformIO's library builder silently compile
NOTHING while still linking; and project build_flags land BEFORE the
framework's own -std=gnu++17 -Os, so build_unflags is required.
CORRECTION carried in this commit: the firmware sizes in c19d846's message and
the first version of the memllib recon doc were wrong — SLPWorkshop 145348,
PAFSynth 145300, SelfTest 141840. They came from building variants in sequence
through a SHARED incremental arduino-cli build directory, which reused stale
objects and under-reported by ~75 KB. Clean-cache rebuilds of the identical
commit give 216736/18492 for SelfTest. The real cost of the memllib upstream
bump is +216 bytes flash, not +316. Never measure firmware size through a
reused build dir.
HISTORY NOTE: this commit and the docs commit before it were rebuilt (force-push,
2026-07-21) so that each contains only what its message describes. The first
versions had the firmware deletions stranded in the docs commit by a shared-index
race between concurrent agents; content is byte-identical to the originals.
Gates: run-all-tests.sh ALL GREEN (nisps/ untouched by this change beyond
include paths); 16/16 pio envs build.
L25: keep the firmware loss-history buffer. With §7.3's "make it real, browser
only, behind a feature flag", the telemetry decision set is now complete and
plan 5e has no judgement calls left in it — only work.
Also corrects the ALIGNMENT entry, which described four live fragments: the
fabricated gradient-health UI (L1) was already deleted in Phase 1, so three
remain. Fragment inventory re-verified against the tree rather than copied
forward from the audit.
The §7.5 decision was "rebase our commits onto upstream, then vendor". On
inspection there was no rebase to do, and the inventory that recommended it
(7a30da9) was wrong on two counts — both corrected in this commit.
All three fork commits touch only `examples/`. That directory is not in the
sketch symlink forest (firmware/MEMLNaut-NISPS/src/ links exactly audio,
hardware, interface, synth, utils, PicoDefs.hpp), so it is never compiled.
They existed to let the RL code build against nisps/core *while it was being
ported*, and that port is done — nisps/ml/{jolt,ou_noise,feedback,geo_push}.hpp
cite the upstream sources directly. Two of the 31 commits I flagged as
"work we are missing", `9fcd459 jolts` and `d0d8a72 noise`, are precisely the
ones already absorbed.
So the question was never rebase-or-not but which snapshot to vendor. Current
upstream, because the staleness was already costing us:
- upstream has DisplayDriver::NavigateToView; the pin does not. The SelfTest
variant called it and had been failing to compile (worked around in
b953681). It was written against a newer memllib, not against nothing.
- e291192 "l r input swap" is a hardware bug fix: the physical L/R input
sockets are wired to opposite codec ADC channels. Every mode on the old
pin sees its stereo input backwards.
Verified by building all three variants with arduino-cli:
SLPWorkshop 145348 flash (+320) 87388 RAM (+4)
PAFSynth 145300 flash (+312) 107060 RAM (+4)
SelfTest 141840 flash (+320) 12028 RAM (+4)
Exactly one compile error: the .ino used kSampleRate in a constexpr, and
upstream 1997699 made it a runtime `extern size_t` so a mode can pick its own
rate. constexpr -> const; it is a once-per-second diagnostic print. The uniform
+316 is the AudioDriver/DisplayDriver deltas — the bulky new upstream code
(GrainDelayI16, ReverbI16, ModFXI16, CCSelectView, RLView, VUMeterView,
PSRAMManager) is header-only and unreferenced, so the linker drops all of it.
.gitmodules points at upstream again: with the pin on an upstream commit, the
fork holds nothing the firmware compiles. Phase 0 pointed it at the fork only
because b37fc53 existed on no remote. The fork's feat/nisps-core-swap branch
stays pushed; nothing is destroyed.
Refs: plan §5, §7.5; ALIGNMENT Q4.
Plan §3 marked burned down, noting where execution departed from the written
plan and why: L35 deleted rather than relocated (zero consumers), and S35's
describe ABI widening forcing two out-of-scope buffer fixes. Also records the
package.json test-glob bug found en route — new unit tests were silently not
being run.
ALIGNMENT defect 5's trailing sentence updated: the behaviour bugs it listed as
pending are fixed.
The three top-level orienting docs are each falsified by several of the eight
preceding commits, so they land here as one sync rather than being split
across commits that would each leave them half-true. Same phase, same push.
MAP.md — removed fixed_buffer.hpp, voice_space.hpp, the src/daisysp entry and
its symlink from the sketch-tree list, input_router.hpp/wire_inputs (inputs are
wired by bind_peripherals now), test_fixed_buffer.cpp, SplitStage/ReadoutStrip/
InputMini, BackendAdvanced.tsx, and feedback/rng.ts; corrected the primitives
count 12 -> 7; replaced the "engine LIFTED from playground/src" provenance;
rewrote the perf-attribute convention and deleted the NISPS_AUDIO_FUNC gotcha
(both macros and gotcha are gone).
docs/AGENT-REFERENCE.md — the memory-section-attribute instruction now names
only the macros that exist; input_router.hpp dropped from the firmware tree;
symlink list and the submodule-init note no longer mention daisysp.
ALIGNMENT.md — defect 5 (dead mass and registry sprawl) rewritten: the deletion
half is done, so the entry now scopes to what actually remains, which is the
registry/dual-truth half (Phase 3) plus the stale specs (docs pass). Per the
ALIGNMENT convention this is a rewrite-to-current, not a checkbox.
Phase 0 of the 2026-07 simplification audit (plan §1). CI has been 100% red on
main since 2026-07-13 and every "gates green" claim since rested on local runs.
- S7 / critic gap 2: push memllib `feat/nisps-core-swap` (3 commits incl. the
pin b37fc53) to monkey-w1n5t0n/memllib and repoint .gitmodules at the fork.
Those commits existed on exactly one disk; `git ls-remote` now resolves the
pin, so `submodules: recursive` checkout and fresh clones work again. Drops
the compensating unreachable-pin error paragraph in build-firmware-arch.sh.
- S24 / S31: the manifold-tests job regenerates from schemas/, runs the codegen
golden test, and fails on a dirty diff — the "schema changes ship with both
generated outputs" rule is now enforced rather than assumed.
- S32: a WASM freshness gate runs the parity harness against the *committed*
manifold/public/nisps.{js,wasm} before the CI rebuild overwrites it. That
artifact is what the webhook ships to production, so a stale commit now fails
loudly instead of shipping.
- critic gap 3 / operator decision §7.4: the VPS webhook
(~/.config/webhooks/meml-deploy.sh, not in this repo) waits for the `CI`
workflow to conclude success on the pushed SHA before building. Fail-closed;
MEML_SKIP_CI_GATE=1 for an emergency hand-deploy. Verified the gate query
returns `failure` for fa37047, i.e. it would have blocked that deploy.
- S31: corrected run-all-tests.sh's false "single command CI invokes" header.
Docs moved with the code: ALIGNMENT defect 1 deleted (resolved) and the rest
renumbered; MAP.md's unreachable-pin warning replaced with the fork pin and a
pointer to the §7.5 vendoring decision; ONBOARDING documents the deploy gate
and the tracked-WASM-ships-to-prod hazard; plan §1 marked burned down.
Gates: scripts/run-all-tests.sh ALL GREEN (ctest 4/4, parity 1273 floats within
1e-5, lint, typecheck, 33 Playwright specs).
Schema-backed modes in console/model.ts are now DERIVED from the codegen
schemas in src/modes/generated/ (source of truth): real param names/groups/
count, plus each mode's ml net shape (MFMode.ml) and schema engine_id. A thin
manifold OVERLAY supplies only label/glyph/ModeClass/input/ordering. New
browser-viable modes xiasri + slp_workshop get derived entries; schema-less
visualizer + c15 stay hand-written on DEFAULT_MODE_ML. Schema min/max/default/
label/curve surface as engine-unit metadata (schemaMin/... on MFParam) without
touching the 0..1 routing semantics.
Switching instrument mode reshapes the runtime-shaped WASM net to the mode's
schema ml config (ConsoleApp effect keyed on [engine, modeId]; no confirm
modal). Boot lands paf_synth dims (4->[10,10,14]->33) once WASM is ready. The
P2.3 axis-count reshape offer still reads the engine's live inputSize and does
not spuriously prompt on a mode switch.
Adds schema-modes.spec.ts (P5 gate): drives switches via a new window.__mf
debug seam and asserts describe() dims, getWeights count, output length/bounds,
and UI param count FROM the imported schemas; spot-checks trainAsync after a
switch. Updates reshape/probe-api/geo-dislike specs to assert from the boot
mode schema instead of the retired fixed 32/126 shape.
All gates green: typecheck, unit (9), build, e2e (33).
- plan P2 marked landed (hardware timing spot-check deferred to the
chokepoint-B session on physical MEMLNaut)
- ALIGNMENT: defect #3 resolved by the runtime-shaped browser MLP
- MAP/AGENT-REFERENCE/nisps-wasm-README: describe the honoured-dims
create + reshape ABI; note per-mode dims become schema-real at P5
- fresh wasm artifacts from the gate run
