Implement Outlier Detection Service
epic-periodic-summaries-core-logic-task-007 — Build the service that classifies peer mentors as outliers based on configurable thresholds for activity volume. Must support underactive detection (below minimum activity threshold), overloaded detection (above maximum threshold), organisation-scoped evaluation runs, and structured outlier reports surfaced to coordinators.
Acceptance Criteria
Technical Requirements
Execution Context
Tier 1 - 540 tasks
Can start after Tier 0 completes
Implementation Notes
Model the detection algorithm as a pure function `List
Store persisted outlier reports as JSONB in Supabase for flexibility, with a typed Dart model for deserialization. The `OutlierDetectionConfig` thresholds represent activity event counts, consistent with how HLF tracks 380 individual registrations per year.
Testing Requirements
Unit tests with flutter_test using mocked ActivityAggregationRepository and mocked Supabase client: test underactive detection (activity count < min); test overloaded detection (activity count > max); test boundary values (count == min and count == max are not outliers); test that empty organisation (zero peer mentors) returns empty report; test missing threshold config returns empty report with warning. Integration tests: seed a Supabase test project with 20 peer mentors at varying activity levels and verify outlier classification is correct. Minimum 95% branch coverage on detection logic.
Supabase pg_cron or Edge Function retries could trigger multiple concurrent generation runs for the same period and organisation, producing duplicate summaries and sending multiple push notifications to users — a serious UX regression.
Mitigation & Contingency
Mitigation: Implement a database-level run-lock using an INSERT … ON CONFLICT DO NOTHING pattern keyed on (organisation_id, period_type, period_start). Only the first successful insert proceeds; subsequent attempts read the existing lock and exit early. Test with concurrent invocations in a Deno test suite.
Contingency: If duplicate summaries are detected post-deployment, add a deduplication cleanup job that removes all but the most recent summary per (user_id, period_type, period_start) and sends a corrective push notification.
FCM and APNs have different payload structures and size limits. An oversized or malformed payload could cause silent notification drops on iOS or delivery failures on Android, meaning mentors never learn their summary is ready.
Mitigation & Contingency
Mitigation: Build the PushNotificationDispatcher with separate FCM and APNs payload constructors, enforce a 256-byte body limit on the preview text, and run integration tests against the Firebase Emulator and a test APNs sandbox.
Contingency: Fall back to a generic 'Your periodic summary is ready' message if personalised preview text construction fails, ensuring delivery even when the personalisation pipeline encounters an error.
Outlier thresholds that are too tight will flag most mentors as outliers (alert fatigue for coordinators), while thresholds that are too loose will miss genuinely underactive mentors — directly undermining HLF's follow-up goal.
Mitigation & Contingency
Mitigation: Implement thresholds as configurable per-organisation database settings rather than hardcoded constants. Provide sensible defaults (underactive < 2 sessions/period, overloaded > 20 sessions/period) and document the tuning process for coordinators in the admin portal.
Contingency: If coordinators report threshold miscalibration after launch, expose a threshold configuration UI in the coordinator admin screen and allow real-time threshold adjustment without requiring a code deployment.
The app may not have 12 months of historical activity data for all organisations at launch, making year-over-year comparison impossible for most users and rendering the comparison widget empty, which could disappoint users expecting Wrapped-style insights.
Mitigation & Contingency
Mitigation: Design the generation service to gracefully handle missing prior-year data by setting the yoy_delta field to null rather than zero. The UI must treat null as 'no comparison available' with appropriate placeholder copy rather than showing a misleading 0% delta.
Contingency: If historical data import from legacy Excel/Word sources becomes feasible, add a one-time backfill Edge Function that populates prior-year activity records from imported spreadsheets. Until then, explicitly communicate the data-availability limitation in the first summary each user receives.