COMPLETED PRODUCTION READINESS EXAMPLES Use these examples as calibration for the production readiness worksheet. A useful readiness record names evidence, owner, release decision, blocking gaps, and the exact next action. Do not copy these answers into your own release record. --- COMPLETED EXAMPLE: SUPPORT REFUND ASSISTANT service: support-refund-assistant owner: support-platform framework: deterministic workflow with bounded investigation loop, policy gate, approval gate, and tool gateway release: internal pilot candidate; draft-only for real traffic until finance approval integration is live | Gate | Evidence | | --- | --- | | State owner | refund_workflow owns request state, order evidence, recommendation, approval status, and stop reason. | | Tool manifest | read_order, search_refund_policy, draft_refund_request; issue_refund remains blocked outside finance approval. | | Policy enforcement point | policy_gate validates policy version, refund threshold, account status, and evidence completeness before any recommendation can move forward. | | Approval boundary | finance approval required for high-value refunds and policy exceptions; approval binds exact action, amount, account, expiry, and idempotency key. | | Memory rules | no long-term customer memory writes in pilot; run evidence stays in trace with retention and redaction policy. | | Trace schema | trace includes intake, route, evidence retrieval, recommendation, policy decision, approval decision, tool proposal, and stop reason. | | Eval suite | support refund eval report covers missing policy, high-value approval, denied approval, unsupported direct refund, and stale policy version. | | CI gate | npm test and npm run capstones:evidence must pass before release candidate. | | Rollback path | disable draft_refund_request and route all refund cases to human support queue. | | Kill switch | feature flag refund_assistant_enabled=false blocks model-mediated recommendation path. | | Incident process | every wrong recommendation becomes an eval fixture with trace ID, policy version, and reviewer decision. | Readiness: green: read-only evidence gathering and draft recommendation for internal reviewers. yellow: limited internal pilot when approval UI, trace dashboard, eval report, and rollback owner are verified. red: automatic payment execution. The payment tool remains unavailable until live approval, idempotency, and finance audit checks pass. release decision: Yellow. Safe for internal pilot as draft-only. Not approved for autonomous refunds. blocking gaps: - live finance approval integration; - production trace dashboard link in release evidence; - on-call runbook for payment-tool incident. next action: Complete approval integration test and attach the trace for one approved and one denied refund. --- COMPLETED EXAMPLE: RESEARCH RAG ASSISTANT service: research-rag-assistant owner: knowledge-platform framework: retrieval workflow with answer synthesis, citation check, refusal path, and eval gate release: read-only beta for approved source collections | Gate | Evidence | | --- | --- | | State owner | research_workflow owns query, source collection, retrieval set, citation check, answer draft, and refusal reason. | | Tool manifest | search_index, read_source, check_citations; no write tools or external messaging tools in beta. | | Policy enforcement point | source_access_gate checks tenant, collection, source license, and freshness before retrieval result enters context. | | Approval boundary | no human approval for read-only answers; approval required before adding a new source collection. | | Memory rules | no personal memory writes; source snippets are trace references, not durable user memories. | | Trace schema | trace records query, retrieval filters, source IDs, omitted sources, citation decisions, answer/refusal, and stop reason. | | Eval suite | missing-evidence, stale-source, conflicting-source, citation-faithfulness, and refusal-quality cases. | | CI gate | retrieval eval must pass minimum faithfulness and refusal thresholds before deployment. | | Rollback path | pin previous index version and disable new source collection route. | | Kill switch | disable answer synthesis and return retrieval-only source list. | | Incident process | unsupported answer becomes a regression case with source IDs and citation check output. | Readiness: green: retrieval-only source browsing for approved collections. yellow: answer synthesis when citation checks pass and missing-evidence cases refuse. red: answers over unapproved sources, private documents without access trace, or stale indexes without freshness labels. release decision: Yellow. Release as read-only beta for approved collections. blocking gaps: - broader stale-source fixtures; - dashboard panel for citation-check failure rate; - source owner signoff for new collections. next action: Add three stale-source regression fixtures and require citation-check output in release evidence. --- COMPLETED EXAMPLE: MULTI-AGENT DELIVERY WORKFLOW service: delivery-workflow-assistant owner: operations-platform framework: coordinator workflow with specialist workers, merge policy, transcript eval, and human escalation release: demo only until operations approval and live-system sandboxing exist | Gate | Evidence | | --- | --- | | State owner | delivery_coordinator owns shared goal, task assignments, worker outputs, merge decision, escalation state, and final status. | | Tool manifest | read_delivery_plan, read_inventory_status, draft_customer_update; no live dispatch or driver-contact tool in demo. | | Policy enforcement point | coordinator validates worker outputs against route policy, customer communication policy, and escalation rules. | | Approval boundary | operations approval required before customer-visible message or dispatch change. | | Memory rules | worker transcripts stay in run trace; no cross-run memory until retention and correction rules are approved. | | Trace schema | trace records task split, worker input, worker output, dissent, merge decision, policy check, escalation, and stop reason. | | Eval suite | transcript eval checks role drift, missing dissent, bad merge, unsupported customer message, and failed worker handling. | | CI gate | transcript eval must fail on bad merge or customer-visible action without approval. | | Rollback path | disable worker dispatch and return to single human-owned operations workflow. | | Kill switch | delivery_multi_agent_enabled=false disables coordinator fan-out. | | Incident process | failed merge becomes a transcript eval with expected dissent and accepted final decision. | Readiness: green: offline transcript evaluation and internal planning demo. yellow: internal operations assistant after sandboxed tools, approval UI, and runbook are verified. red: live dispatch, driver contact, or customer messaging without approval and sandboxed side effects. release decision: Red for production. Keep as demo and evaluation harness. blocking gaps: - sandboxed operations tools; - approval UI for customer-visible messages; - live runbook and escalation owner; - baseline comparison against single-workflow design. next action: Add sandboxed tool adapter and compare coordinator output against the deterministic workflow baseline on ten incidents.