Operate predictions as controlled maintenance evidence.

01 / Setup

Deploy the contract first

1. 1
   Apply migration 011

   Create tenant-scoped readings, feature snapshots, predictions, feedback, validation, drift, and job evidence.

2. 2
   Deploy the analytics pipeline

   Set ANALYTICS_PIPELINE_SECRET, store it in Supabase Vault, deploy the Edge Function, then schedule the checked-in 02:20 UTC cron.

3. 3
   Choose the analytics destination

   Leave the export URL empty to score in Supabase only, or connect an HTTPS ingestion endpoint for the allowlisted feature batch.

4. 4
   Run a staging day

   Confirm row counts, prediction explanations, RLS behavior, and a deliberately rejected suggestion before production activation.

02 / Signal quality

Make failure labels explicit

Set work_orders.metadata.failure_event=true only for confirmed functional failures. Usage meters should be cumulative and non-decreasing. Sensor records need a stable metric key, unit, source, timestamp, and quality. Bad-quality readings are retained as evidence but excluded from baseline scoring.

03 / Daily review

Start with the reason, not the percentage

Review critical and high assets first. Compare the leading factors with the asset history, decide whether the suggested condition check is warranted, then accept, reject, or defer. Acceptance records intent; it does not create work or modify a PM template. Use the normal work-order flow to authorize maintenance.

04 / Model lifecycle

Candidate → shadow → active → retired

Keep a trained model in candidate until the artifact checksum, feature schema, training window, and backtest are recorded. Shadow mode compares predictions without operational effect. Activation requires an authorized Market President or Admin process, passed validation, a rollback target, and review of class-specific error rates. Drift creates a retraining review, never an automatic promotion.

05 / Troubleshooting

Follow the evidence chain