Add v030 paper data-only supplement: README.md

experiments/nav3r_v030_supplement_cpl_20260911/README.md

@@ -0,0 +1,34 @@
+# Nav3R v030 paper supplement — data-only, 2026-09-11
+
+Unified code, paper, metric definitions and reproduction instructions:
+https://github.com/AIGeeksGroup/Nav3R/tree/20260911/supplements/20260911_v030_cpl
+
+Code PR: https://github.com/AIGeeksGroup/Nav3R/pull/3
+
+This is a supplement to previous releases, not a replacement. The referenced v030 paper and its SHA are in the code release. Both repositories were private at publication; appropriate access is required.
+
+## Content
+
+|Archive group|Paper evidence|
+|---|---|
+|table_iii_primary320|Historical v191, 320 episodes × 3 conditions; STOP-aware Table III|
+|table_iv_full983_and_figure4_mechanism66|983 episodes × 2 conditions; paired 66 episodes × 10 controls; Tables IV and Figure 4|
+|section_vi_d_diagnostics|E1/E2 diagnostics, separate 14-scene 66-episode panel|
+|section_iv_e_navigation_training|Training labels, provenance, collection reports and training metadata; NOT full feature caches|
+
+`asset_manifest.json` gives every included file's original relative path, SHA256, bytes and source. `archives_data_only/*.tar.gz` contain new files. Existing files with the same verified SHA may be referenced at `existing_hub_revision`, avoiding duplication. `omitted_asset_inventory.json` is an inventory, NOT the omitted data itself.
+
+No new videos, large decision feature caches, or qualitative replay-frame arrays are included. Their remote originals are retained. Previously uploaded weights and videos remain untouched. The withdrawn first large shard is not part of this release.
+
+Use the GitHub release's `tools/materialize_assets.py`, with the fixed successful HF revision in `provenance/HF_RELEASE_STATUS.json`; then `tools/reproduce_paper.py` for CPU metric reproduction. Do not use a moving branch as an unrecorded dependency. If the successful receipt is not yet present, publication verification is still pending.
+
+## Boundaries
+
+- Recomputing metrics from original records is different from rerunning frozen navigation and different from reproducing exact training snapshots.
+- The 5,877 decision feature caches are not shipped. Their regeneration depends on the rollout policy and environment and is not guaranteed bit-identical. Training labels, splits and source hashes remain available.
+- Five-channel Binary reconstruction checkpoint with SHA a8150018a25bc572f50cfb7ba57d82b5cc8141746e679d1c7d48be36e57064ed was not found in the verified sources. No other checkpoint is substituted.
+- HM3D and public pretrained dependencies must be obtained under their own licenses. This supplement does not grant rights to third-party datasets/models or include all original reconstruction training data.
+- Validation scenes were used during development; deployment uses selected seed123. Same-state PTv3-zero logits equality is not independent closed-loop trajectory equality.
+- Failed/incomplete attempts are provenance only; use the explicit authoritative source manifests, not all files discovered recursively, for statistical analysis.
+
+Paper metrics use explicit STOP; missing STOP is not interpreted as false. See GitHub for thresholds, weights, paired bootstrap definitions and known limitations.