[Mattral]

Summary

Follow-ups 2 and 3 after run 1: a shoulder-midpoint pose anchor by default, and a guard that keeps one normalization scheme per study.

Two named schemes (worldgap.data.normalization.SCHEMES):

• shoulder_midpoint (default): pose anchored at landmarks 11/12.
• hip_midpoint: the spec 5.2 anchor (23/24), kept for framings where the hips are visible.

Both scale by shoulder width; hands are unchanged (wrist + bounding-box diagonal).

Why shoulders: in run 1's clean recordings MediaPipe's mean visibility was 0.005 for the hips vs 0.999 for the shoulders, so a hip anchor is an extrapolation. Spec 5.2 now records this as a deliberate, documented deviation, with that measurement as the justification.

The scheme is fixed per study, never per frame:

• It's a single parameter per call, recorded as one string in every rollout's metadata["normalization"]["scheme"]. Nothing switches anchors on visibility; test_anchor_never_switches_on_visibility pins that.
• Every loader takes normalization_scheme=. run_v1_real_data.py takes --normalization and writes it to a new study_settings.json next to the pre-registration, before any score.

The guard (same spirit as the temporal-provenance guard):

• GapAnalyzer.fit() refuses training rollouts with mixed schemes.
• compute_gap() refuses source/target rollouts whose schemes differ from each other or from the training scheme. Normalized vs raw counts as different. A gap across anchors includes the anchor change itself.
• The training scheme is saved in checkpoints. Pre-existing checkpoints load with it unknown, and only the source/target check applies to them.

The analyzer gains no modality branch; the check reads metadata only.

Tests: tests/test_normalization_scheme_guard.py (7). The invariance and reference-point tests in test_normalization.py now run under both schemes. The smoke test checks study_settings.json.

• Mutation check: with the compute_gap guard removed, 4 of the 7 guard tests fail. The other three cover the same-scheme path, fit's own guard and unnormalized sets, and pass as expected.

Note: re-scoring run 1's recordings under either scheme is a follow-up analysis; its numbers are not comparable with run 1's pre-registered result. This is stated in spec 5.2.

Verification

• pytest -q: 155 passed
• ruff check .: clean
• scripts/run_v2_actuation.py re-run (analyzer changed): every value identical
• Privacy greps clean
• CI green on 3.10 / 3.11 / 3.12

Added after review

• Conflict with docs: run 1's null is underpowered, not impossible #7 resolved (merge of main; the CHANGELOG keeps both sides).
• Prominent upgrade note, at the top of [Unreleased] and in the README: 0.2.0's loaders return unnormalized coordinates, main normalizes by default with a shoulder anchor, so the same video yields different states. 0.2.0 stores, checkpoints and scores are not comparable, and the guard refuses to mix them.
• The guard's error now names the remedy: re-extract with one scheme, or apply normalize_rollout(r, scheme) to rollouts saved unnormalized. The second is exact, because normalization is per frame. Verified on run 1's real data: normalizing the saved clean/take0 store vs re-extracting the video gave a maximum state difference of 0.0 across all 46 windows. A unit test pins that equivalence, and another asserts both refusals name the fix (9 guard tests).
• Docs: docs/v1_first_run_results.md and spec 5.2 note that run 1's stores are unnormalized, so a re-score must normalize first, and that a re-score is a follow-up analysis not comparable with ρ = −0.224.

pytest -q: 157 passed · ruff check . clean.

🤖 Generated with Claude Code