This directory contains compact, public result records for Milestone 26. Raw datasets, licensed body models, checkpoints, per-sample predictions and demo caches are intentionally not redistributed.
m26_relation_grounder_protocol.json: protocol frozen before training, including scene-level folds, protected-scene audit and training budget.m26_relation_grounder_cv_summary.json: complete three-fold development cross-validation summary, per-fold metrics and pooled results.m26_pooled_metrics.csv: README-ready pooled comparison table derived from the JSON summary.
M26 uses 2,254 QA records from 35 ScanNet scenes and 1,014 answer-changing counterfactual pairs. Every pooled prediction is out-of-fold with respect to its scene. Neither the protected M21 final set nor the legacy 668-QA final set was read during this experiment.
The labels are deterministically derived from explicit surface-distance and contact-fraction relations supplied to the model. Near-ceiling M26 performance therefore validates relation extraction, body-part identity, transfer of the frozen rule across development scenes, and counterfactual direction handling. It is not evidence of open-world physics, force, friction or support reasoning. An independently labeled external/OOD evaluation remains future work.