# Execution record — selected CPU tests and analytic geometry probe

## Source and environment

- Source repository: [`koulakis/manifold-microscope`](https://github.com/koulakis/manifold-microscope)
- Pinned commit: [`564652bbdfa9ca6b905a3891a52b878a2af3a18e`](https://github.com/koulakis/manifold-microscope/tree/564652bbdfa9ca6b905a3891a52b878a2af3a18e)
- Git tree: `1da7572f290b2b8c9be1492bc00e587a47e4a61c`
- Source licence: [BSD-3-Clause](https://github.com/koulakis/manifold-microscope/blob/564652bbdfa9ca6b905a3891a52b878a2af3a18e/LICENSE)
- Runtime: CPython `3.13.14`, PyTorch `2.13.0`, macOS arm64, CUDA unavailable.

The source checkout was recorded clean before and after the bounded work. The
full resolved package list is retained at
[evidence/environment-lock.txt](../../evidence/environment-lock.txt).

## Commands and outcome

The following commands were run from an isolated audit workspace against the
pinned source:

```sh
./.venv/bin/python -m pytest -q \
  source/microscope/manifold_examples/tests/test_symbolic_computations.py \
  source/microscope/tests/test_patches.py
./.venv/bin/python audit/run_toy_finite_difference.py
```

The selected unit-test subset passed twice: `3 passed in 12.24s`, then
`3 passed in 4.86s` during artefact verification. The deterministic probe
passed its three embedded acceptance assertions:

| Fixture and measurement | Observed value | Probe threshold |
| --- | ---: | ---: |
| Circle total-volume relative error | `9.96160370350763e-05` | `< 1e-3` |
| Torus total-area relative error | `0.0032086359550390874` | `< 5e-3` |
| Torus scalar-curvature maximum absolute error | `0.008539875376420536` | `< 0.1` |

See [the sanitised run record](../../evidence/run-record.md),
[test output](../../evidence/unit-tests.txt), and
[machine-readable result](../../evidence/toy-finite-difference-result.json).

## Deliberate exclusions

The source's broader finite-difference tests require `cuda:0` and large grids;
they were not run on this CPU-only host. No data download, dSprites, COIL-20,
GPU, manifold fitting, VAE, model/checkpoint, generalisation-bound, paper
benchmark, table, figure, or external service was used.

## Interpretation

This execution demonstrates only that the selected released finite-difference
path completed on tiny analytic fixtures under the recorded environment. It is
not a performance evaluation, paper reproduction, or verification of either
official claim.