experiments/README.md

# Geometry experiments

This directory evaluates one geometric hypothesis at a time without modifying the baseline
encoder or its result paths. Each file in `hypotheses/` is a complete replacement for
`encoder/geometric.py` and must expose `build_spatial_code()` and `dump_spatial_code()`.

The runner reads the existing SAM3 and DA3 caches directly and adapts them together in memory.
It may read an existing combined geometry cache when one is already present, but never creates
or modifies source inference caches.
It writes derived spatial codes below:

```text
experiments/caches/spatial codes/metric/tracking/uniform/<hypothesis>/<frames>/
```

Evaluation results belong below:

```text
experiments/results/symbolic/metric/tracking/uniform/<hypothesis>/<frames>/
```

Start with 64 frames. Run one scene with:

```bash
python -m experiments.run scene0699_00 \
  --hypothesis "Compute Gravity Before Building Object Instances" \
  --depth metric --tracking tracking --input uniform --frames 64
```

List available hypotheses with `python -m experiments.run --list`.

For a full cached batch, use the launcher. Its default `--workers 0` uses every CPU
available through the process affinity mask and budgets nested numerical threads across workers:

```bash
python -m experiments.launch \
  --hypothesis "Compute Gravity Before Building Object Instances" \
  --depth metric --tracking tracking --input uniform --frames 64
```

Evaluate the generated spatial codes with the existing symbolic scorer while keeping all
question-level results under `experiments/results/`:

```bash
python -m experiments.evaluate \
  --hypothesis "Compute Gravity Before Building Object Instances" \
  --depth metric --tracking tracking --input uniform --frames 64 \
  --quiet --errors
```