GeoEvidence component utilities and paired statistics: research source archive
Seven exact source/config/test/initializer files are preserved from Git commit 7b0efa2cda306040fccab8e0620923a2d44c8428, specifically the inner active GeoEvidence tree. This is not the older top-level implementation, not another upload of the previously published factual summaries, and not a complete training reproduction.
Included and excluded
component_advantage.py: NumPy component normalization, validity masks, variation and quota utilities, with the original11 synthetic pytest tests unchanged.paired_statistics.py: standard-library paired/image-group bootstrap analysis, plus newly written synthetic-only tests identified in SOURCE_MANIFEST.json.- Minimal original package initializers, and exact
paired_cf_4b_step20.yamlas a frozen reference only. - No private dataset, real evaluation row, feedback, credential, machine path, personal identifier, model weight, optimizer, original research report or training launcher is included.
Do not execute the training configuration. It intentionally preserves historical upload_every_step: true, public: true and a repository target; those fields are not an instruction or authorization to start training, upload, modify a repository or change visibility. The referenced patched runtime is not included. Tests only parse this YAML.
Synthetic CPU validation
With Python3.12, NumPy2.2.6, pytest8.3.5 and PyYAML6.0.3 available:
PYTHONPATH=src PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 PYTHONDONTWRITEBYTECODE=1 python -B -m pytest -p no:cacheprovider -q tests/test_component_advantage.py tests/test_paired_statistics_archive.py
There are11 unchanged original component tests and6 additional synthetic statistics/config tests. No test consumes private inputs, loads a model, invokes training or uses a Hub API. Statistical CLI tests use only temporary synthetic questions/pairs and a small20-replicate bootstrap, not a real experiment. This verifies software behavior, not policy quality, real-data metric validity or full paper reproduction.
The analysis tool's normal CLI can write per-question CSV and include caller-selected identifiers. Do not automatically publish its outputs from real data. Audit output privacy separately. The synthetic tests and package do not include those real outputs. All role strata required by the current CLI are populated in synthetic fixtures; unsupported real-input shapes are not certified by this smoke suite.
Provenance and rights
Read SOURCE_MANIFEST.json, DEPENDENCY_CLOSURE.json, THIRD_PARTY_NOTICE.md and RESEARCH_ARCHIVE_TERMS.md. The operator has authorized public research archival of owned code; this is not selection of a new open-source license or an extension of verl's Apache license. Historical configuration bytes are unchanged. UPLOAD_MANIFEST.json pins all other release files. Local test-tool wheels and receipts are deliberately outside the public package.