# Monday checklist — Jev as a System-1 sidecar on AWS

Jev classifies, routes, scores, and verifies. Amazon Bedrock generates. AgentCore Gateway Cedar authorizes writes. Do not invert those layers.

## Day 1 — Residency and payload

- [ ] List every field that would go in TypeSafe `state` (ticket body, policy snippet, order JSON)
- [ ] If any field can be PHI, PCI PAN, or a named-region customer clause → **stop**. Use Bedrock in-account instead
- [ ] Confirm there is **no** VPC endpoint for TypeSafe; NAT / flow logs will show `api.typesafe.ai`
- [ ] Legal has a sentence on TypeSafe DPA / ZDR **or** you are restricted to synthetic tickets

## Day 2 — Pin the model and write atomic questions

- [ ] Model ID is **`jev-1.13.0`**, not `jev-latest` / `jev-preview`
- [ ] Five questions max for the first shadow week: two Noul, two Choice (include `other`), one Score
- [ ] Instructions name JSON paths with backticks (`ticket_message`), not implied intent
- [ ] Arithmetic, date compare, and counts stay in code ([jev-1.13 jaggedness](https://docs.typesafe.ai/model-jaggedness/jev-1.13.md))
- [ ] Choice cardinality ≤ 255, or you have a documented two-stage score-then-choose

## Day 3 — AWS wiring (still no production writes)

- [ ] `TYPESAFE_API_KEY` in Secrets Manager; Lambda / Runtime role `GetSecretValue` only
- [ ] Timeout ≥ 3s (vendor claim is 70–500 ms E2E from West Coast; you are probably not there)
- [ ] Retry on 429 honoring `Retry-After` (published limits 1,200 rpm / 250k tok/s **can change without notice**)
- [ ] CloudWatch EMF: `noul` or `confidence`, `choice`, `model` (versioned ID from the response)

## Day 4 — Shadow against an in-account classifier

- [ ] Same labels on **Amazon Nova Micro** (or current Bedrock classifier) and Jev
- [ ] Log disagreements; a human labels a held-out slice
- [ ] Promote Jev only if you can explain the misses — not because TypeSafe published 193.6×
- [ ] Bedrock Guardrails still wrap the **generative** call ([Guardrails guide](https://www.factualminds.com/blog/how-to-set-up-amazon-bedrock-guardrails-production/))

## Day 5 — Writes still on Cedar

- [ ] High-confidence Jev Choice is **not** authorization
- [ ] AgentCore Gateway Policy LOG_ONLY then ENFORCE for refunds / CRM / catalog edits
- [ ] Pydantic AI: Jev may pick a tool; Bedrock `FallbackModel` writes args; Gateway still allows or denies
- [ ] Browser / Code Interpreter **off** unless the turn requires them
- [ ] Model remaining AgentCore sessions on the [pricing calculator](https://www.factualminds.com/tools/amazon-bedrock-agentcore-pricing-calculator/)

## Hard stops

- [ ] No Jev on a HIPAA write path with raw clinical or patient mail in `state`
- [ ] No treating “cannot hallucinate” as “cannot be wrong”
- [ ] No Score interpolation as a calculator
- [ ] No Agents Classic for net-new builds (maintenance for new customers after 30 Jul 2026)
- [ ] No invented Jev latency in a design review — we have not called the API