Production Launch Checklist and Architecture Review
A production launch is an evidence-based risk decision. Review the AI use case, source ownership, schema, embedding lifecycle, retrieval quality, authorization, privacy, prompt and tool boundaries, performance, recovery, observability, cost, and operations. Every checklist item needs an owner, artifact, result, exception process, and review date—not a confident verbal assurance.
What will you be able to do?
Run a structured architecture review.
Convert vague readiness into evidence.
Record risks, exceptions, and owners.
Define go/no-go and rollback gates.
Schedule post-launch validation.
What must be reviewed?
Product and data
User problem, non-goals, and safe no-result behavior are explicit.
Sources, versions, permissions, provenance, and ownership are mapped.
Chunking, model, dimensions, metric, and preprocessing are versioned.
Updates, corrections, revocation, deletion, and reconciliation pass.
Retrieval and AI quality
Exact baseline and held-out golden dataset exist.
ANN recall, task relevance, slices, and no-result cases meet gates.
RAG citations support claims and injection tests pass.
Model, prompt, and evaluation limitations are documented.
Security and governance
Runtime roles are least privilege; RLS/tenant denial tests pass.
TLS, secrets, network boundaries, encryption, audit, and retention pass.
External provider data use is approved.
Caches, logs, replicas, exports, and backups honor policy.
Agent tools have typed validation, approval, and idempotency.
Reliability and operations
Workload card, plans, capacity, and tail-latency load tests pass.
Pools, timeouts, backpressure, vacuum, indexes, and disk are budgeted.
Backups, PITR, failover, and restore drills meet RPO/RTO.
SLOs, alerts, runbooks, on-call, and drift checks are active.
Migrations, canary, stop conditions, rollback, and version trace pass.
Cost and organization
Low/base/high capacity and provider cost models are reviewed.
Unit cost and budget alerts are defined.
Owners cover database, ingestion, retrieval, model, security, and incident response.
Known risks have accepted owners and expiry dates.
What is the launch decision record?
Record release and dataset IDs, evidence links, passed/failed gates, residual risk, exceptions, approvers, canary scope, stop thresholds, rollback steps, launch window, and 24-hour/7-day/30-day reviews. Do not fabricate a sign-off in schema or documentation.
How do you verify it?
Choose five checked items at random and reproduce their evidence. Run one rollback, restore, cross-tenant denial, deletion, and quality regression. If evidence cannot be reproduced, the item is not complete.
What breaks in production?
The checklist becomes a document nobody can reproduce.
Quality gates have no frozen release artifact.
Residual risks lack owners or expiry.
Canary has no automatic stop conditions.
Launch succeeds technically but nobody watches user outcomes.
AI pair-work prompt
Facilitate a pgvector launch review using this checklist and my evidence index [paste]. Ask for artifacts, identify contradictions and missing owners, distinguish blocking from accepted risk, and produce a decision record. Do not mark any item complete without reproducible evidence.
Verification contract: An independent reviewer samples evidence and runs the five critical drills before the launch decision.