# Public Data Provenance

Distribution Statement A. Approved for public release; distribution is unlimited.

## Dependency Graph

The dataset supports the following reconstruction when combined with the
separately maintained experiment implementation:

```text
public FLAC + vehicle GPS + sensor coordinates
  -> 10 ms audio mean power and timestamp alignment
  -> 200 ms resampling and WGS84-to-UTM distance calculation
  -> gdf_cleaned + normalized_cleaned + valid_indices
  -> five-step contexts and all node subsets of size 1, 2, or 3
  -> 964,017 ordered context/action/utility examples
  -> attach the released 24-dimensional learned_embeddings.npy rows
  -> fit the declared alpha=1 ridge reward model
  -> linearized bandit instance
  -> 5-second modular circular replay
  -> PCA(10) + KMeans(20) state process and transition estimate
  -> delayed contextual-bandit experiment
```

The intermediate `examples_with_embed.pkl` is a dataframe cache. It contains
timestamps, sensor subsets, subset strings and sizes, 85-dimensional contexts,
42-dimensional raw action vectors, utility and distance diagnostics, nearest
sensor labels, and 24-dimensional learned embeddings. It is not source data.

The intermediate cyclic bandit-instance pickle is also a cache. It contains
the ordered feature/action arrays, observed and linearized rewards, oracle
values, replay indices, clustered state sequence, transition counts and
matrix, stationary distribution, total-variation diagnostics, and mixing-fit
parameters. The release stores its portable replay/state/config components,
but not the pickle object.

## Active Construction

- Raw sequence: `20250812_1050`
- Active nodes: `[1, 2, 3, 4, 5, 7, 8, 9, 10]`
- Excluded node: `6`
- Observation interval: 200 ms
- Context history: 5 observations
- Action subsets: all active-node subsets of size 1, 2, or 3 (129 actions)
- Utility weights for second/third selected nodes: `0.45` and `0.20`
- Realized utility distance scale on the valid chain: `10.316986163806751` m
- Chronological time split: 60% train, 20% validation, 20% test
- Embedding seed: `22`
- Embedding architecture: two towers, hidden width 512, tower width 8,
  multiplicative combined representation of width 24
- Reward fit: ridge regression with alpha `1.0`
- Replay trim: inclusive row positions 265 through 7,472
- Replay stride: 25 observations; period 7,208; horizon 1,000,000
- State model: PCA dimension 10 and KMeans with 20 states,
  `random_state=0`, `n_init=20`
- Mixing fit: lags 1 through 80 from a TV-distance curve computed through lag
  150; fitted beta `0.8075480035772411`

## Clean-Rebuild Validation

An isolated validation rebuilt the processed tables from the released raw
files. Shapes, columns, row order, valid indices, and every numeric value
matched the retained processed artifacts exactly (`max_abs_diff = 0`). The
964,017 example rows were then reconstructed without an old example cache.
Direct model inference differed from the retained embedding array by at most
`7.2e-7`; attaching the portable array produced exact scientific equality.
The resulting cyclic instance, including rewards, replay indices, states,
transition matrix, and mixing estimates, matched the prior cache exactly.

No original/private scientific file was read during the clean build. The
public data therefore contain sufficient scientific information to
reconstruct the experimental instance when combined with the experiment
implementation. Private validation logs and source filesystem paths are not
part of this release.