genesis-eval

The external measurement rig for Genesis

Layout

• validation/ — the probe suite. Substrate (exact-math mechanism checks), capability (end-to-end behavior), dynamics (in-vivo ablation), breakthrough (stimuli a lookup table cannot answer). rust-probes/ is a sibling crate that exercises the production genesis crate through its public API only. See validation/README.md for tiers, methodology, and the run commands.
• eval_heldout/, eval_transfer/ — ARC held-out and transfer sweeps: results.tsv, per-step telemetry/, and summary.txt. The negative results are in these files on purpose.
• eval_arc_worldplay_20260924/ — per-step telemetry from an external ARC training/eval harness, preserved because it is the only record of those episodes.
• arc_adapter.py, world_play.py, tests/ — the external harness and its tests.

What is not here

environment_files/ and environment_files_heldout/ are excluded on purpose: ARC environments are distributed by the arcengine SDK, and republishing the held-out corpus would contaminate the very evaluation it exists to keep honest. Run the SDK to fetch your own copies.

Run

Requires a side-by-side checkout of genesis — see validation/README.md.