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.
Phase 2, S35. Two real defects from one root cause, both confirmed by trace
rather than taken from the audit:
1. Divergence. WasmIML built its TS Dataset mirror with a cap of 100 while
every addExample() ALSO pushed into the C++ FIFO ring, capped at 128. Since
train() reads the C++ ring and trainAsync() reads the TS mirror, past 100
examples the two trained on different datasets — silently.
2. Latent OOB read. nisps_ml_train sizes its sample-weight span by the C++
side's example_count() (up to 128), but wasm-iml.ts allocates that heap
buffer from the TS dataset's size (<=100). Once the ring exceeds the mirror,
the span reads past the end of the caller's allocation.
Fix: name the capacity ONCE as nisps::ml::kDefaultMaxExamples = 128, used by
FixedStorage's default template arg, DynamicStorage's default ctor arg, and the
MLP<> alias (which is the only real FixedStorage instantiation path and carried
its own independent 128 literal — the last copy of this dual truth). Expose it
through nisps_ml_describe and have the TS side read it instead of hardcoding.
Dataset's constructor default is removed entirely: a default was what invited
this bug class, and the sole call site now always supplies the describe() value.
ABI NOTE: this extends nisps_ml_describe from a 6-int to a 7-int descriptor.
nisps_ml_describe always writes 7 ints regardless of the caller's buffer, so
every call site had to grow in the same change or it would overflow the WASM
heap by 4 bytes per call. All five sites updated: three in wasm-iml.ts (init
defaults, init per-instance, reshape re-describe — the finding said there were
two), one in wasm-worker.ts, one in tests/cpp/parity_wasm.mjs. The parity
harness's expected-dims check now also pins the new max_examples slot.
Regression test: tests/cpp/test_mlp_storage_defaults.cpp — pins the two storage
policies to one constant, and drives MLPCore<DynamicStorage> exactly as
bindings.cpp does past the old TS cap, asserting it saturates at 128 and not at
100. Fail-before/pass-after confirmed by temporarily setting the constant to
100: 2 failures, named. Reverted: green.
Audit correction: the cited dataset.ts:81 is the FIFO eviction check; the
hardcoded default was at dataset.ts:45.
Gates: run-all-tests.sh ALL GREEN, parity PASS.
Phase 1 group 2 (L27, L26, L28, S21, L13, ST6, S20).
- L27: fixed_buffer.hpp + its test + the CMake entry — no consumers.
- L26: dislike_multiplier_ and its doubling/halving bookkeeping — upstream
InterfaceRL residue that drove nothing. The audit pointed at the wrong test
file for the surviving reference; the actual assert was in
test_mlp_geo_dislike.cpp:211, removed here.
- L28: added copy_weights_to(std::span<float>) to FixedStorage and
DynamicStorage and switched feedback.hpp's take_snapshot/push_undo/nudge to
it. Drops the permanent whole-net flat_ scratch buffer from FixedStorage and
the per-gesture double copy. Behaviour-identical: same source values, same
write order, same RNG draw order in nudge().
- S21 + L13: deleted NISPS_AUDIO_MEM / NISPS_APP_SRAM / NISPS_AUDIO_FUNC —
zero use sites outside perf.hpp and comments — and rewrote midi_io.hpp's one
misshapen NISPS_AUDIO_FUNC use as a plain `inline void`. perf.hpp now
documents only the inlining/hotness macros that actually exist, and
audio_driver.hpp no longer claims an SRAM discipline the code never had.
- ST6: feedback.hpp's header now describes the four current modes and the
Geometric default, dropping the retracted "geometric push NOT ported" claim.
S20 — OPERATOR DECISION (§7.1): the four legacy feedback behaviours
(RandomiseOutputs, RandomiseMlp, AvoidStyle::Diffuse, the RandomiseMlp branch of
on_drag) are KEPT, not deleted. They are wanted as building blocks for
experimenting with how different instruments feel under different behaviours.
Each is now marked at its definition as deliberately-retained research reserve
so future audits stop flagging it as dead code.
L25 (the 16 KB firmware loss-history buffer) is NOT done here — see the phase
report; it turned out to be coupled into the shared mlp.hpp, and its fate
belongs with the browser telemetry build (§7.3 / plan §6.5e).
Gates: run-all-tests.sh ALL GREEN.
Operator-approved ABI change (P2 stop-point). The WASM MLP is now
MLPCore<DynamicStorage>:
- nisps_ml_create(input, output, hidden[3], n, seed) honours its args;
non-positive/null fall back to the historical 32→[10,14,18]→126, so
every pre-P2 caller (manifold, worker, parity harness) stays
bit-identical. Invalid/oversized dims (>4096) → null.
- NEW nisps_ml_reshape(ml, in, out, hidden, n, spread): fresh net at the
new dims, warm-started via nisps/ml/warm_start.hpp (overlapping region
copied; rest keeps spread init); feedback controller re-created (state
resets — reset-on-reshape modal is the front-end contract). Failure
leaves the old net untouched.
- nisps_ml_describe(ml, out): takes the handle; null reports defaults.
- FeedbackController got the same storage split: algorithms in
FeedbackControllerCore<FbStorage>; FixedFeedbackStorage keeps firmware/
tests source-identical via the old alias; DynamicFeedbackStorage (one
arena) sizes to the runtime net. Firmware .text unchanged (122692).
- MLHandle: per-instance scratch vectors; dropped the dead 2MB
batch_out_scratch.
- TS: types.ts decls (+_nisps_ml_reshape), wasm-iml re-describes the
created instance, worker carries a shape-contract note for P2.3.
Verified: ctest 4/4 incl. new warm-start grow/shrink test; reshape ABI
smoke (dims honoured, overlap survives, invalid rejected, outputs
bounded); parity PASS unchanged (2.4e-7); lint clean; manifold 9 unit +
20 e2e green; firmware .text 122692 (+0.30% vs pre-P2 baseline).
Algorithms (forward, backprop/SGD, init, move_weights, diagnostics) now live
once in MLPCore<Storage> (nisps/ml/mlp.hpp). Storage models:
- FixedStorage (storage.hpp): template-sized std::array, zero heap. The
classic MLP<NIn,H1,H2,H3,NOut,...> is an alias preserving kInput/kHidden*/
kOutput/kNumLayers/weight_count() constexpr — firmware + bindings + modes
compile unchanged.
- DynamicStorage (dynamic_storage.hpp): runtime dims, ONE arena allocation
at construction, nothing per-call. #error under NISPS_TARGET_EMBEDDED
(new macro in core/perf.hpp); sole lint-cpp.sh heap-allowlist entry, plus
a lint check that fails if the #error guard disappears.
Verification:
- new ctest test_mlp_storage_parity: fixed↔dynamic BIT-identical across
init/draw/inference/train(FIFO)/move_weights(pin mask)/eval_loss/
layer_stats/set_weights/infer_batch/reset; invalid+moved-from inert
- golden ML vectors (pre-refactor constants) pass → bit-stable refactor
- native↔WASM parity PASS, max delta unchanged (2.4e-7)
- chokepoint B compile: PAFSynth .text 122324→122692 (+0.30%, ±1% budget);
RAM +416B (eval scratch)
- fix: firmware-common.sh used bare 'python' (absent here) → ${PYTHON:-python3}
Part of one-core-engine-refactor P2. nisps_ml_create ABI untouched (P2.2 is
an operator stop-point).
monkey-w1n5t0n