Reference model corpus

Build, test, and measure InvariantCAD against a small set of real mechanical models.

The repository reference corpus is a deliberately small outcome harness. Each model is authored with the same public APIs available to a library consumer, has one named part output, and records analytical default-volume and bounds expectations. It does not stand in for features that InvariantCAD does not yet support.

Model	Construction	Default output
Electronics enclosure	Box subtraction plus four cylindrical mounting bosses	enclosure
Six-bolt flange	Overlapping cylinders, a through bore, and a six-hole bolt circle	flange
Hollow stepped shaft	Three overlapping shaft stages and a continuous axial bore	shaft

All three models are parameterized, material-backed parts and use only the common Manifold and stock OCCT modeling subset: boxes, cylinders, transforms, Booleans, parts, and materials. The corpus therefore exercises mesh-first and exact evaluation without pretending that unsupported sketch, surfacing, or assembly workflows exist.

Deterministic outcome tests

tests/reference-models.test.ts rebuilds every document twice, checks canonical serialization, validates the document, and evaluates compatible cases on both shipped kernels. The assertions cover analytical volume, world bounds, non-empty mesh structure, and density-derived mass. Manifold receives a small relative tolerance for segmented circles; stock OCCT is checked against the analytic values.

Run the focused test directly:

pnpm exec vitest run tests/reference-models.test.ts

Measurement-only benchmark

Benchmark schema v2 emits JSON and intentionally has no pass/fail performance thresholds. It launches one fresh child process for every model/kernel pair. Each case has a 120-second process timeout; a timed-out worker is killed and reaped with its kernel/model context included in the failure. The first run in that process is labeled fresh-runtime-first-run: it creates a new kernel and evaluator, builds a fresh document, and performs the first evaluation. It is not a claim that the machine, filesystem cache, WebAssembly compiler cache, CPU frequency, or surrounding host is cold.

Every same-runtime-repeat builds a new document with byte-identical canonical serialization, reuses that case’s kernel and evaluator, and starts only after the previous result is disposed. No evaluator artifact cache is enabled, so a repeat is not described as a cache hit.

pnpm exec tsx scripts/benchmark-reference-models.ts \
  --kernel manifold \
  --repeat-runs 3

mkdir -p .artifacts
pnpm exec tsx scripts/benchmark-reference-models.ts \
  --kernel all \
  --repeat-runs 1 > .artifacts/reference-models-v2.json

Use one or more --model <id> options to measure a subset. The complete one-repeat corpus also executes in Node 26 CI as a protocol and correctness check, without comparing performance values.

Each run records canonical document bytes and SHA-256, node and parameter counts, measured geometry and mass, tessellation counts and buffer sizes, binary STL size, and STEP size when the kernel advertises STEP export. OCCT uses the explicit 0.1 linear deflection, 0.3 angular deflection, absolute tessellation profile recorded in the report. Manifold records that its mesh is controlled by authored primitive segmentation and its kernel default; the report does not imply triangle-count parity between backends.

Timings remain separated into document construction, canonical serialization, evaluation, measurement, physical mass properties, tessellation, STL serialization, optional STEP serialization, result disposal, and total workflow time. Repeat samples also receive minimum, median, and maximum summaries.

Memory evidence is Node’s process.resourceUsage().maxRSS high-water value in KiB at named lifecycle boundaries inside the dedicated case process. It includes Node.js, loaded modules, JavaScript, WebAssembly, and native kernel state. It is process-wide, cannot attribute allocations to the model, and is not instantaneous RSS. The public kernel protocol does not expose native allocation counters, so current reports explicitly record native-handle telemetry as unsupported; private adapter bookkeeping is not mislabeled as a native-handle count.

Compare reports produced from the same source revision on equivalent hardware, operating system, and runtime versions. Do not use this script as a cross-machine leaderboard.