GeoEvidence answer-training recipe facts
A narrow factual preservation record of declared numerical configuration and source-file fingerprints. This is not GR1 source code, a reproduction bundle, a training dataset or an evaluation claim. Original project documents/scripts were inspected locally but are not redistributed here. Environment defaults can be overridden; no actual run completion or performance is inferred from this record.
TRAINING_FACTS.json records the declared answer-only rollout settings,42-pair batch default, one-epoch/seed42 recipe and two scope names. Relative source filenames and hashes permit later correlation without publishing private machine paths or samples. The project sync commit is a documented provenance pointer, not an independently authenticated private repository read.
GR1 source rights remain unresolved. The public upstream verl Apache license is not extended to this project. No source-license permission, image/question/annotation rights, base-weight license or private-feedback disclosure is granted by these factual records. The already-published model weights are not copied here. See TERMS.md and UPLOAD_MANIFEST.json.
VERSION_LICENSE_PARTITIONS.json separates the older top-level snapshot from the actual inner Git working copy and separately licensed verl runtime. V1_ALIGNMENT_AGGREGATES.json contains numeric-only historical483-question/2509-pair metrics, not the later v2 protocol; original data/file paths were excluded. These aggregates were already present in local mixed archives; the new contribution is a clean usable factual subset, not a claim that the underlying bytes were never backed up.