# Model weights

This directory stores the exact fitted models used in the reported comparisons.

## `our_model/`

Selected LatticeSpike checkpoints. Each file name includes the dataset version,
protocol, and seed.

## `native_baselines/`

Paper-faithful baseline models. These keep each baseline's native training and
inference protocol. They are not the primary apples-to-apples comparison.
The included MVA files are clearly marked five-trial pilots. They are useful for
replay and inspection but are not the final 200-trial native reproduction.

## `fair_baselines/`

Baseline models retrained under the causal matched protocol used for the main
comparison with LatticeSpike. The adapted MVA method is named MVA-Causal. It
uses one validation-selected hyperparameter lock for every evaluation seed and
must not be presented as native MVA.

## External bundle

The complete shared bundle is stored in the `LatticeSpike_Model_Weights`
OneDrive folder. The upload was verified at 33 files and 999,849,787 bytes. It
includes the five final MVA-Causal checkpoints. It does not include the final
native 200-trial MVA checkpoints, which remain separate from the fair
comparison.

Every populated directory must include a `manifest.json` with the source commit,
dataset SHA256, split SHA256, model revision, seed, command, artifact SHA256, and
reported result files. EPR has no trainable weights, so its manifest entry records
that it is deterministic.

Weights represented in this repository use Git LFS. The final MVA-Causal
checkpoints are distributed through the external bundle and are recorded in
`fair_baselines/manifest.json` with exact sizes and SHA-256 hashes.