Sanitized research workspace for the Kaggle competition
neurogolf-2026. The project treats each ARC-style task as compact program
synthesis and compiles candidate rules into cost-conscious ONNX graphs.
This snapshot contributes first-party ONNX/DSL tooling, compact program- synthesis experiments, tests, and reproducibility notes. Competition inputs, generated submission bundles, and third-party artifacts remain at their original sources.
This is a working research record: small, inspectable tools, experiments, and notes for studying compact ONNX programs for NeuroGolf tasks. It is designed to be useful to people comparing ideas, extending a tool, or tracing how a claim was tested—not as a mirror of the competition's files or of other participants' artifacts.
The publication boundary is intentional. First-party code and original research writing are shared here; competition data, leaderboard exports, submission bundles, and third-party artifacts stay with their source. See DATA_SOURCES.md for retrieval and provenance, and RELEASE_MANIFEST.md for the release boundary and validation record. PUBLICATION_BOUNDARY.md records the detailed review of retained research evidence versus excluded source files.
- first-party ONNX/DSL tooling and tests;
- task-analysis scripts and small reproducible diagnostics;
- research notes, score ledgers, and handoff documentation;
- build helpers that describe the local submission workflow.
- competition inputs, raw task data, and downloaded leaderboard exports;
- submission archives and generated ONNX bundles;
- copied public kernels and third-party artifacts;
- virtual environments, downloads, caches, profiler dumps, and multi-hundred-MB trace JSON files;
- credentials and machine-specific absolute paths.
For a source release, obtain data through the official competition after
accepting its terms, then pass its local directory to the relevant tool's
--comp-dir option. Do not commit that directory or its generated outputs.
Please share corrections, small reproductions, and alternative implementations with clear provenance. Attribute upstream work, keep access-controlled material out of issues and pull requests, and distinguish a local result from an official leaderboard result. This archive keeps negative results on purpose: they are often the shortest path to a better next experiment.
GitHub main is the verified remote copy of this sanitized archive. A separate
Kaggle Dataset upload was attempted, but it never appeared in the account's
dataset listing, so it is not counted as a verified backup.
For this source snapshot, cite CITATION.cff or
CITATION.bib and use the tagged
snapshot-2026-09
source state. Cite the separate ONNX artifact dataset independently when using
those reviewed model bytes.
- Snapshot:
snapshot-2026-09 - Boundary and verification:
RELEASE_MANIFEST.md