AI User Feedback Loop — C5

The audit UI-024 deferred-comment problem

When PR #79 (2026-04-10) shipped the AI insight infrastructure, AIInsightCard.swift:230-235 carried this block at the bottom of recordFeedback:

// Note (audit UI-024): Local RecommendationMemory recording deferred —
// RecommendationMemory is owned per-AIOrchestrator instance, not a
// shared singleton. Wiring requires either an EnvironmentObject pattern
// or a dedicated DI container. Tracked separately; analytics signal
// already captures the feedback event server-side.

The analytics signal (home_ai_feedback_submitted) fired every tap. The on-device storage layer (RecommendationMemory) was fully implemented + tested. The UI surface (AIFeedbackView thumbs-up/down) existed. None of them were wired to each other.

Result: 47+ feedback events fired during the 2026-05-31 → 2026-06-01 GA4 window, but RecommendationMemory.totalCount on every user's device stayed at 0. The home screen kept surfacing the same recommendation patterns regardless of user response — "Your sleep quality could use a boost" could appear on day 50 even though the user dismissed it 49 times. No mechanical learning.

The three closures C5 ships

Closure A — UI-024 wire (the headline)

Promote RecommendationMemory from per-AIOrchestrator-instance to an app-lifecycle @StateObject via the new RecommendationFeedbackController facade. AIInsightCard.recordFeedback reads it via @EnvironmentObject and calls record(outcome:) directly. The deferred comment block is fully removed.

Closure B — Reinforcement loop in AIOrchestrator

AIOrchestrator.process(segment:jwt:overrideSnapshot:) gains an applyReinforcementLoop(recommendation:segment:) block between the finalRecommendation resolution and the latestRecommendations[segment] publish. Per-signal-per-segment confidence-tier-only adjustment:

For each signal in recommendation.signals:
  if dismissed >=3 times in last 30d for this segment:
    confidence = downgrade(confidence)   // high → medium → low
    fire home_ai_feedback_signal_suppressed
  else if acceptanceRate(segment) > 0.70 with >=5 outcomes:
    confidence = upgrade(confidence)     // low → medium → high
    fire home_ai_feedback_segment_boosted

Gated on AppSettings.aiFeedbackLoopEnabled (default ON). When OFF the orchestrator skips the block entirely — feedback is still recorded so re-enable doesn't lose history.

Closure C — Settings → AI Feedback management surface

New SettingsCategory.aiFeedback row in the Settings v2 dashboard. Detail screen has 4 sections:

1. Feedback Activity — total outcomes recorded (or empty-state copy)
2. Acceptance Rate by Topic — iterates AISegment.allCases:
   • ≥ 5 outcomes → "X% · N outcomes"
   • < 5 outcomes → "N of 5 (below quorum)"
   • 0 outcomes → "—"
3. Currently Suppressed (last 30 days) — collected from frequentlyDismissedSignals across all 4 segments
4. Controls — opt-out toggle + clear-history button (with confirmation dialog firing home_ai_feedback_history_cleared)

Bonus surface — dismiss-reason picker

When the user taps thumbs-down in the sheet's AIFeedbackView (NOT on the home card — keeps card tap latency at 1 tap per PRD OQ-2), a confirmationDialog surfaces 5 options: not-relevant / already-aware / disagree / repetitive / other. "Other" opens a secondary sheet with TextField capped at 80 chars (stored on-device, never sent to GA4 per PRD §"GDPR/Privacy").

Frozen algorithm constants

PRD §"FROZEN constants" locks these — changing requires re-Phase-1:

Three new analytics events

All screen-prefixed home_ per the 2026-04-08 project convention. Join the existing home_ai_feedback_submitted (unchanged).

Single-session lifecycle (5 phases in ~3 hours)

Total wall time: ~3 hours session-1 (Phases 0-5) + merge ceremony 2026-06-02 (Phases 6-8 with revert #579 sandwich — see below).

The Dependabot revert sandwich (2026-06-02 merge ceremony)

The merge ceremony surfaced a Dependabot bump that broke main mid-flight:

1. #575 E4 backlog drift correction → ✅ MERGED d774dc2
2. #577 Dependabot style-dictionary 3.9.2 → 5.4.2 bump → ✅ MERGED 91979f8 — BUT v5 API removed StyleDictionary.extend() and top-level .registerTransform(). sd.config.js uses both. make tokens started failing on every PR.
3. #579 Revert #577 → ✅ MERGED cee8801 (single-step git revert; main green again)
4. #572 C5 full lifecycle → ✅ MERGED ec5dff9 (after update-branch pulled in revert)

The C5 merge took 4 actions to land because of the intervening Dependabot incident. Followup queued: a proper style-dictionary v5 migration PR (rewrite sd.config.js for new StyleDictionary() constructor + instance .registerTransform()) shipped separately later.

What's deliberately not in C5

• Time-decay weighting — recent feedback weighted more than 30d-old. C5.b track (D1.a covers this).
• Cohort-level reinforcement — feedback from other users informing your recommendations. Privacy-impacting; D1.b (deferred to v8.0).
• Signal generation from feedback — AI-suggested replacement when suppressed. D1.c (deferred to v8.0+; mapping table vs LLM gateway open question).
• Suppression-transparency UX inside the home card — D1.d (deferred to v8.0+).
• Push notification reinforcement — C2 + C4 banner action_taken events route into RecommendationMemory. C5.c track (D1.c sibling).
• Server-side dismiss-reason aggregation — privacy boundary per PRD §"GDPR/Privacy".

Phase E discipline

C5 shipped + merged during the v7.9 Phase E 14-day soak (2026-05-21 → 2026-06-04). The release adds no enforcement gates and adds no new schema fields. Consumes existing v7.8.6 + v7.9 infrastructure exclusively. Phase E compliant — Day 13/14 at merge.

Cross-references

• Source case study — docs/case-studies/ai-user-feedback-loop-case-study.md
• PRD — docs/product/prd/ai-user-feedback-loop.md
• Phase 0 research — .claude/features/ai-user-feedback-loop/research.md
• Phase 2 tasks — .claude/features/ai-user-feedback-loop/tasks.md
• Sibling C2 (slot 40) — single-day readiness-aware training advisory
• Sibling C4 (slot 41) — sustained-trend HRV observer
• W26 CI concurrency fix (slot 42) — the previous merge-day sandwich
• Phase E context — slot 34 (framework-v7-9-promotion.mdx)
• Successor D1 (PR #576 draft) — adaptive-intelligence-next-pass; recommended v1 pair: D1.a time-decay + D1.d transparency UX