[rsasaki0109]

Summary

• What changed? The readback record (docs/use_cases/records/gsdc2023_official_readback_20261024.json) gains a taroz_kaggle_probes section, and the goal doc gains a short section on it. The section records:
  • the Kaggle score of the archived taroz test output;
  • four phone-group swaps;
  • 40 single-drive swaps on top of the official recipe file 57678818.
• Why? The remaining gap to the taroz post-deadline score (Private 0.024 m, Public 0.020 m) could not be located on train. Train with the current recipes is level with taroz or better.

Scope

• Single feature / single operational improvement / single user-visible value
• Adds or tightens one regression, sign-off, dogfood, or artifact-schema check
• No unrelated refactor mixed in
• No docs-only noise mixed into solver/protocol work. Docs only.
• Protocol ingest, parser work, runtime knobs, solver behavior, and docs are split unless this PR explicitly needs the boundary crossing

External Reference

• Source repo used for comparison: taroz/gsdc2023. Its archived test result_gnss_imu.mat files were assembled as submission.m does.
• What was adopted here: nothing. The probes are diagnostic, and the recipe is unchanged.
• What was intentionally not adopted: taroz rows in any submission the recipe pins.

Behavior Boundary

• Runtime default changed? No.
• Env-gated or config-gated? No.
• Artifact/schema changed? Yes. The readback record gains a taroz_kaggle_probes section.
• Dataset or credentials required? No.

Validation

• python3 tests/test_cli_tools.py
• python3 tests/test_benchmark_scripts.py
• python3 tests/test_packaging.py
• python3 tests/test_python_bindings.py -v
• python3 tests/test_ros2_node.py
• ctest --test-dir build --output-on-failure

Focused checks run:

python3 tests/test_gnss_reproduce.py   # OK (the record still loads)
git diff --check

Sign-off / Benchmark Impact

• Added or updated regression/sign-off: none.
• Affected dataset(s): GSDC 2023-2024 test (Kaggle readback only).
• Findings:
  • Taroz archive: Public 0.796 / Private 0.782 m, the README score exactly (57683971).
  • Private split size: about 20 drives. A 2-epoch A205U misalignment (about 470 m) scored Private 24.117 m.
  • Private: most of the gap is on lax-n Pixel5 (-0.010), sjc-we1 Pixel7Pro (-0.007) and sjc-r A505U (-0.006). Native is better on 5 Private drives.
  • Public: mtv-m Mi8 (-0.018) and the ebf-xx start (-0.017) dominate. Native is better on ebf-x Mi8 (+0.014).
  • Where it shows: taroz wins through slowly varying 0.1-0.4 m differences, often at drive starts. Train shows no such start deficit (first 5 minutes 0.506 vs 0.516 m median error).
• Expected metric impact: none.

Notes

• Risk: none (docs only).
• Follow-up PRs: possibly the ebf-xx start (Public) and the lax-n start (Private).
• Explicit non-goals: using taroz output in the recipe.

---

Generated by Claude Code