# Paired feedback rule catalog

Catalog version: `1-pending-review`. Analyzer version: `1`.

A [provisional draft assuming Damien is the reference](feedback-rule-draft.md) is enabled at James's request as a separate provisional coaching evaluator. It uses inside/outside labels, a 10° elbow attention filter, and separately reviewed visual knee-position rules (“toward the turn” / “up close to the tank”). Knee positioning is never inferred from included-angle measurements. Each provisional bullet carries `user-assumed-reference` provenance and `damien-reference-draft-3` version; none claims coach approval.

The coach-reviewed production catalog remains empty. The app displays **Provisional coaching** for Rider and Reference rider, with the assumption that the reference position is correct. Source identities and names in authored prose remain unchanged. Identity checkboxes were removed at James's request; explicit session roles identify the riders. A source-bound `photoCoaching` review takes precedence over the angle-rule fallback and supplies up to three authored whole-body points with visual-review provenance. The direction is assessed from those photos unless a saved user selection overrides it. A conflicting override withholds the authored points. This is not a general automatic image classifier.

## Live wheel lean evidence

Wheel lean is a separate image-plane evidence row, not an elbow/knee coaching rule and not part of the authored three-point budget. Draft wheel/reference proposals must be reviewed before pair approval. Approval binds to both current annotations and source evidence; edits withdraw it. Image-vertical and assumed-level differences compare acute magnitudes; a shared stationary feature compares signed relative orientation. Conservative annotation sensitivity spanning zero leaves ordering unresolved. Unknown perspective and steering error remain explicit. Similar speed is user-reported context and cannot enable comparison, risk/traction scoring, prescribed angles, physical lean, or causal body-position advice.

For pairs without an applicable authored review, elbow comparability and turn-direction gates remain. Knee placement is an optional source-bound visual comparison saved in the pair context; changing turn direction clears it. Explicitly obscured knees suppress its controls and advice. Knee/foot/pelvis placement are considered together, without assuming identical knee flexion or peg contact in every photo. Coach review and utility validation remain unmet.

| Candidate topic | Measurements | Threshold | Approved wording / provenance | Status |
|---|---|---|---|---|
| Anatomical left elbow | Reviewed shoulder–elbow–wrist included angles | Not established | None | Inactive |
| Anatomical right elbow | Reviewed shoulder–elbow–wrist included angles | Not established | None | Inactive |
| Anatomical left knee | Reviewed hip–knee–ankle included angles | Not established | None | Inactive |
| Anatomical right knee | Reviewed hip–knee–ankle included angles | Not established | None | Inactive |

Activation requires a versioned rule ID, anatomical feature and difference direction, reviewed positive meaningful-difference threshold in degrees, priority tier, conditional suggestion, explicit limitation, and nonempty review provenance. Both rider identities and that feature's views must be reviewed as comparable. Inside/outside language additionally requires explicit corner direction. Feature comparability is a human review decision, not camera calibration.

Rules are evaluated separately from measurement extraction. Priority tier and stable rule ID determine order; raw delta size does not rank anatomical concerns. Feature/direction duplicates collapse, and at most three focuses are emitted. Each focus carries its rule version, pair ID, exact frame measurements, and anatomical joint IDs. The enclosing report carries source evidence, input revision, context, and catalog/analyzer versions.

`FeedbackRulesTests` constructs synthetic catalogs solely to prove threshold boundaries, unsupported-rule suppression, ordering, deduplication, directional gating, and the cap. Those fixtures are never imported by the app or production catalog.

All initial real suggestions require independent coach support/actionability review and James's evidence-identification/utility assessment. Keep pair-specific review records and derived artifacts in ignored local storage. No review request was sent to another person during implementation.

The current evidence does not establish foot pressure, peg contact, seat/pelvis contact, grip force, causal dependencies between these topics, or hidden depth. No 2D edit or angular difference validates a 3D posture or promises a riding benefit.