feat(examples): laser-cut glockenspiel — the SCS closed-loop demo part#402
Conversation
Eight free-free 6061-T6 bars tuned to C6-C7 major (lengths solved from the closed-form beam model f1 = 3.5608*(t/L^2)*sqrt(E/12rho), E and rho asserted against the kernel materials registry at build time) with cord holes on the fundamental's nodal lines at 0.2242*L, plus a folded 5052-H32 U-channel stand whose deck holes trace the converging nodal rows. Per docs/plans/2026-07-06-scs-closed-loop-demo.md (PR #397). build.mjs drives the full pipeline in-process: sheet_metal_create -> sheet_metal_check vs the SendCutSend catalog -> sheet_metal_suggest_fix (the naive fold fails 4 bend-relief checks; the fix lands in the design as parametric relief notches, not a skipped check) -> sheet_metal_unfold -> per-part DXF -> springback-compensated bend sequence -> folded STEP + GLB -> sheet_metal_cost, and emits receipt.json naming an oracle for every claim. frequencies.mjs reproduces the tuning table standalone (rounding to the 0.1 mm cut grid costs at most 1.2 cents). 6061 is deliberately the cut-only stock (SCS does not bend it); the stand uses their 0.125" 5052 bending stock at its fixed 3.18 mm radius and published K. All nine parts check shop-ready; outputs upload as-is. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
New MCP tool: strike a flat free-free bar in simulation and verify its
pitch before ordering. Modal frequencies from both the closed-form
Euler-Bernoulli model and a hole-aware 1-D Hermite-beam FEM (cord holes
enter A(x) and I(x) exactly, w_eff = w - 2*sqrt(r^2 - (x-x0)^2));
strike-excited synthesis (mode shape at strike point x half-sine mallet
spectrum, Q-based decay with cord-at-holes suspension damping) -> 16-bit
44.1 kHz WAV -> Hann/FFT with parabolic peak interpolation -> cents
verdict vs note/expect_hz. Takes a flat sheet_metal_create document_id
(dims, material, holes read from the session) or an explicit bar spec;
E/rho resolve from the kernel materials registry. The modes table
{hz, gain, q} doubles as the payload for client-side Web Audio synthesis
in the viewer widget later.
The FEM immediately earned its keep: the glockenspiel's nodal cord holes
flatten every bar ~5 cents (stiffness loss where mode-1 curvature is
nonzero) — bigger than the 0.1 mm rounding error. Bar cut lengths are
now hole-compensated (~0.2 mm shorter) and build.mjs gains an order-gate
stage: every bar's synthesized strike must FFT within +-5 cents of its
target note or the build fails. All 8 land within +-1.7 cents; a WAV per
bar ships in out/. Overtones sit at the non-harmonic free-free ratios
(2.76/5.40/8.93) and a center strike suppresses mode 2, like a player.
12 new tests: exact eigenvalue roots, FEM==closed-form on a uniform bar
to <0.5 cent, overtone ratios, synth->FFT round trip <1 cent, strike
position physics, doc-geometry extraction, bent-part rejection.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Added: New MCP tool that strikes a flat bar in simulation and verifies its pitch through the audio path: hole-aware 1-D FEM modal analysis → mallet-excited synthesis → WAV → FFT peak → cents verdict. The FEM immediately caught a real detune: the nodal cord holes flatten every bar ≈ −5 cents (stiffness loss where mode-1 curvature is nonzero) — bigger than the cut-grid rounding error. Cut lengths are now hole-compensated (~0.2 mm shorter), and Timbre physics is in there too: non-harmonic partials at 2.76/5.40/8.93 × f₁, center strike suppressing the antisymmetric mode 2, and cord-at-nodes suspension damping making the partials die in tenths of a second while the fundamental rings for seconds. Frequencies are first-principles; decay Q is a documented heuristic. The 12 new tests (439 total pass): exact free-free eigenvalue roots, FEM ≡ closed form on a uniform bar to <0.5 ¢, synth→FFT round trip <1 ¢, strike-position physics, session-doc geometry extraction, bent-part rejection. |
main moved server.ts to the assembler/ToolDef architecture, so simulate_strike's registration is ported to the new pattern: a toolDefs table exported from tools/acoustics.ts (pack sheet_metal, RW), spread in the server assembler, named in LIST_TOOL_ORDER, and titled in TOOL_METADATA — same three-point shape as the other sheet-metal tools. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Deliberate surface change: the strike-simulation tool joins tools/list after sheet_metal_nest. Regenerated via UPDATE_TOOL_SURFACE=1. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
What shippedYou can now hear a part before you order it. A new "Simulate Strike" tool takes a flat bar design, models how it vibrates when struck, synthesizes the resulting sound as an audio file, and checks whether the dominant pitch lands within a specified tolerance of the target note — all before any metal is cut. Alongside this, a complete laser-cut glockenspiel example (C6–C7 major scale, eight bars and a folded stand) demonstrates the full workflow: bar lengths derived from acoustic physics, suspension holes placed to preserve the fundamental, fabrication files checked against a real shop's capabilities, and every tuning claim verified by the simulated strike before the order goes out. Plain-English summary generated by Choji from this pull request. |
1 participant
The CAD half of the SCS closed-loop demo: a C6–C7 glockenspiel where every dimension is physics and every receipt claim has a household oracle — the loudest being a phone spectrogram.
What's in it
geometry.mjs— single source of truth. Bar lengths are solved from target pitches via the closed-form free-free beam modelf₁ = 3.5608·(t/L²)·sqrt(E/12ρ); E and ρ come fromvcad-kernel-sheet's materials registry (al-hard= 6061-T6) and are asserted against the live engine at build time. Cord holes sit on the fundamental's nodal lines (0.2242·L from each end). The stand's 16 deck holes sit directly under each bar's nodal holes — two converging rows, the tuning physics visible in the frame.build.mjs— the full pipeline in-process: create →sheet_metal_checkvs the SendCutSend catalog → unfold → per-part DXF → springback-compensated bend sequence → folded STEP + GLB → cost →receipt.json.frequencies.mjs— the tuning table standalone (no build needed): note, target Hz, as-modeled length, predicted Hz, error in cents (≤1.2¢ from rounding to the 0.1 mm cut grid).The DFM loop, exercised honestly
The stand's chamfered deck corners put material at the wall-bend ends: the naive fold fails
sheet_metal_checkwith 4 tear-out warnings.sheet_metal_suggest_fixanswersadd_bend_relief; the build applies it in the design (parametric notches, visible in the DXF) and re-checks to zero. The build asserts the naive design fails and the fixed one passes — the self-heal loop is part of the test.Material split is deliberate: SCS cuts 6061 but doesn't bend it (bars: cut-only, raw — coatings damp the ring); the stand uses their 0.125″ 5052-H32 bending stock at its fixed 3.18 mm radius and published K-factor (0.44).
Verification
sendcutsendcatalog (0 errors, 0 warnings).Depends editorially on #397 for the plan doc link; no code dependency.
🤖 Generated with Claude Code