# Monday checklist — Decider 2B or Jev

Decider and Jev score a closed set. Amazon Bedrock writes sentences. AgentCore Gateway Cedar authorizes writes. Pin one checkpoint. Do not average two of them.

Read [Strands Decider 2B vs Jev](https://www.factualminds.com/blog/strands-decider-2b-vs-jev-2026/) first. If the column you picked is Jev, also run the [Jev Monday list](https://www.factualminds.com/examples/architecture-blog-2026/jev-typesafe-aws/monday-checklist.md).

## Day 1 — Residency

- [ ] List every field that would go into `state`
- [ ] PHI, PCI, or a named-region clause in that list → TypeSafe is out. Stay in the account.
- [ ] If TypeSafe stays in, name the egress (`api.typesafe.ai`) in the design note
- [ ] Legal has a sentence on the TypeSafe DPA, or the shadow week uses tickets you wrote

## Day 2 — Pin one ID

- [ ] Local path, small pin: `strands-decider-2B-qwen3.5-v1-2610` or `strands-decider-E2B-gemma4-v1-2610`, plus the Hugging Face revision you loaded
- [ ] Gemma step-up, only after a labeled miss: E4B, then `strands-decider-12B-gemma4-v1-2610`. 26B-A4B stays off the ticket until 12B is what missed ([15 Oct table](https://strandsagents.com/blog/strands-decider-gemma-4-models/): 12B is 200/231 at 136 ms H100; 26B-A4B is 207/231 at 245 ms)
- [ ] Local path, threshold already fit on the 1 Oct launch: stay on `strands-decider-2B-hobson-v19` until you reshadow
- [ ] Hosted path: `jev-1.13.0`. Not `jev-latest`. Not `jev-preview`.
- [ ] Re-read [TypeSafe Models](https://docs.typesafe.ai/models). On 16 Oct 2026, `jev-latest` pointed at 1.13.0 and `jev-preview` pointed at `jev-1.14.0` (released 14 Oct 2026, same $0.042/MTok).
- [ ] Rate limits come from your TypeSafe console. Do not copy 80 requests/s from the 11 Oct guide until the console shows it.
- [ ] Two questions for the first shadow. Choice lists include `other`.

## Day 3 — Latency on the box you will use

- [ ] You did not budget the v19 RTX 3090 median (115 ms, p95 299 ms) for a CPU path
- [ ] You know the README stamps that 3090 row onto v1 as "same architecture and size," which is not a new v1 timing run
- [ ] Gemma H100 medians are not that 3090 row. E4B at 139 ms and 12B at 136 ms are one release run, not a reason to call E4B slower
- [ ] A smoke timing exists for your hardware, or the SLO cell is explicitly blank
- [ ] A timeout fails closed into review. It does not become a default label.

## Day 4 — Shadow

- [ ] Labels were marked by a person on a held-out slice
- [ ] The comparison is rules or one Bedrock classifier against the single pinned checkpoint
- [ ] You did not send the same ticket to Decider and Jev and keep the higher score
- [ ] A 0.9 cutoff from v19's short-classification band (0.952 in `evaluation/results.md`) was not pasted onto v1 or onto long tickets

## Day 5 — The write is still someone else's job

- [ ] Decider or Jev confidence is not authorization
- [ ] AgentCore Gateway Cedar still permits or denies the tool
- [ ] A deterministic check compares amount and order id to the record
- [ ] Bedrock Guardrails still wrap the generative call
- [ ] Browser and Code Interpreter stay off unless that turn needs them

If you only do one thing: pin the versioned ID in the ticket before anyone discusses a threshold.