a physics the box has to match

the takeaway in one paragraph

Genesis-Embodied-AI's genesis-world is a published robotics physics engine: rigid bodies, MPM, SPH, PBD, and FEM solvers, all run through one stepped simulation loop. The lab pinned its exact revision, `6a2a1f0`, and vendored a checksum-locked subset of its source. It then wrote three of that engine's solvers over again in the box's own code: MPM, PBD, and SPH. Those reimplementations pass 9 of 9 tests in `research_worker/test/kernel.genesis-mpm.test.ts` and 11 of 11 in `research_worker/test/kernel.genesis-pbd-sph.test.ts`.

A third file, `research_worker/test/kernel.genesis-wgsl.test.ts`, passes 11 of 11 checks on the GPU shader source: nine against a frozen manifest, one on the pinned arithmetic, one on the atomics and statement order in the p2g and grid passes. That is 31 passing tests across the three files, 0 fail, 0 skipped: 20 solver reference tests and 11 shader checks. Nobody has run the pinned Genesis code and compared its numbers to the box's numbers step by step.

The program's plan calls that comparator `genesis_oracle_witness`. Its two carrier files, `ops/world/genesis_oracle.py` and `world/genesis/witness.ts`, are built and held at review on their own branch, unmerged, with a green validate receipt on file. Their validate runs the oracle's canonical-fixture helpers and grades the box's own reducers against hand-stated oracle output: `test_genesis_oracle.py` passes 5 of 5 and the TypeScript witness test passes 8 of 8, with no Genesis process in the loop. Neither has run against Genesis itself, because the runner that would call the real engine has never been pointed at it.

Tests you write yourself can only check that your code agrees with your own understanding of the math. They cannot check that your math agrees with what the original engine computes on a real input. Those are two different claims.

A draft law in the canon, `laws/hardware-correspondence`, states the general form of that demand: for a backend, the reference result and the backend result have to agree operator by operator. A self-consistent set of solvers with no outside referee does not satisfy that. It satisfies a weaker claim: the box's code does not contradict itself.

how the check works

The source is pinned to one exact commit of the Apache-2.0 `genesis-world` repository, vendored by `ops/world/vendor_genesis.sh` into `training/fixtures/genesis-6a2a1f0/`. Running `sha256sum -c MANIFEST.sha256` over that directory returns 0 FAILED lines. Every vendored file still matches the digest recorded at vendor time. That manifest was generated from this local copy after the clone, so this confirms no local drift since vendor time, not a match against the upstream repository.

`mpm.ts` implements one MLS-MPM step: quadratic B-spline weights, the APIC affine velocity term, and a stress update for four of Genesis's six MPM material models (elastic, liquid, snow, sand). `pbd.ts` runs one PBD substep over all seven stages the Genesis solver runs: stretch, bending, volume, density with S_Corr, viscosity, particle collision, then the boundary. `sph.ts` runs one WCSPH substep: density and pressure forces.

The GPU shader source lives under `wgsl/`, checked as literal text against a frozen manifest. `wgsl.ts` is the parser and schema that runs that comparison, not a second copy of the math itself. Nine of the eleven WGSL tests compare structure to the manifest and the source constants; one asserts the pinned MPM arithmetic directly; one asserts the atomics and statement order in the p2g and grid passes.

The program's own node grid, in `docs/GENESIS-WORLD-PARITY-PROGRAM.md`, names the missing runtime check directly: a node called `genesis_oracle_witness`, alongside a second node, `genesis_contracts`, whose carrier `world/genesis/contracts.ts` is held the same way on its own branch. Both are built, reviewed, and unmerged, each with a green validate receipt. The repo's default interpreter has no `genesis` module. The oracle venv the program pins does import it, and nobody has pointed the runner at it yet.

Listing `ops/world/` on this branch does not include `genesis_oracle.py`. Listing `research_worker/src/kernel/world/genesis/` does not include `witness.ts` or `contracts.ts`.

the numbers

31 of 31 written tests pass across the three files, 0 fail, 0 skipped: 20 are MPM, PBD, and SPH reference tests, 11 are shader checks. Most of those numbers are self-checks, the box's code against its own fixtures and frozen shader source. One is not: the PBD/SPH file also reads the vendored Genesis source at test time and asserts six box constants equal the values Genesis itself declares. None of the 31 reads a number the real Genesis engine produced by running it, because the runner that would produce that number has not been executed.

what this does not cover

• The 31 passing tests confirm the box's own math is internally consistent for the configurations they check: conservation of grid mass and momentum for one seeded 200-particle configuration with gravity off, translation equivariance for one one-cell shift, and a deterministic manifest parse. They do not confirm the box's numbers match a number the real Genesis engine produced on the same input.
• mpm.ts reproduces four of Genesis's six MPM material models (elastic, liquid, snow, sand), not elasto-plastic or muscle. sph.ts reproduces WCSPH only; the same Genesis file, sph_solver.py, also defines a DFSPH solver with no reference here.
• genesis_oracle_witness and genesis_contracts are both built and held at review, unmerged, each on its own branch. The repo's default interpreter has no `genesis` module; the oracle venv the program pins does import it, and nobody has pointed the runner at it yet, so no oracle fixture exists.
• The program's plan puts the rigid solver and coupler reference in wave 2. FEM, the stable-fluid solver, the tool solver, and the kinematic solver appear in the program's mechanism map and in neither wave.
• A fourth test file, training/tests/test_genesis_trace_data.py (19 tests), was not run for this note. The builder it covers reads a fixture directory, training/fixtures/genesis-6a2a1f0/oracle, that does not exist.
• The four WGSL shaders are checked as text against a manifest; they are never compiled or executed here. ops/world/genesis_wgsl_validate.py is the script that would run them under wgpu, and no result from it is reported.
• The word genesis names two different things in this lab's canon. This note's genesis is the vendored physics library, Genesis-Embodied-AI's genesis-world. The canon's ideas/genesis-superposition-edge is a separate, unrelated draft idea about a first-instantiation point in the lab's own ontology.

the next gate

The runner exists and is held at review. The engine is already importable from the pinned oracle venv; nobody has pointed the runner at it yet. Run the oracle in the pinned venv on one MPM step fixture, and diff the result against mpm.ts. Land genesis_oracle_witness and genesis_contracts, or narrow the claim to what the three solver test files actually cover.

• note 34 · the rest of the layer
• note 38 · fano surface collapse isoformer
• note 40 · informational collapse, topological time
• the canon