# First graph chooser study: integration evidence

Workbench [PR #26](https://github.com/kimjooyoon/gooo-ecosystem-workbench/pull/26),
tested head `cf66ac049e929af1efb3a57b0d70ae859ddf05c3`, merged main
`96eeafbb308c68c760d86a3a6677ced02e7a4958`. The tested and merged trees are equal.

PR CI [37753453807](https://github.com/kimjooyoon/gooo-ecosystem-workbench/actions/runs/37753453807)
and push CI [37753448302](https://github.com/kimjooyoon/gooo-ecosystem-workbench/actions/runs/37753448302)
succeeded. The exact PR artifact ZIP is retained here with its GitHub metadata
and verified digest. It includes full Linux training outputs, exported models,
row predictions and Gooo assessment observations.

`verification.json` states the actual comparison scope. All 12 models' weights
and metadata, four split files, all train/evaluation/ablation predictions and the
complete Gooo feature audit agree with the original local artifacts. The full
race log and a separate local repeat-training receipt are retained. Each native
model test makes a fresh construction, native execution, saved replay and four
budget probes for the three program families. Published archive tests also
recount all 72 constructions, 72 replays and 288 budget observations.

## Additional diagnostic after observing results

`shared_projection.go` groups each shared field judge's 256 float32 values by
exact bits, across all 1,728 field observations. Run it with Go 1.27.1 from the
workbench repository root. There are 108 distinct inputs and no groups with
contradictory binary labels. Its upper bound is 1,728/1,728 fields. This narrows
the next investigation beyond exact shared-input collisions; it does not show
that the small network can learn the mapping or generalize to new programs.
This diagnostic was performed after the frozen study and changed no training
settings or published weights.