Jev reference architecture
A reference architecture here is an unofficial box diagram in words: clients → your API → questions module → one TypeSafe door → traces → action / HITL. It is not TypeSafe’s unpublished internals.
Independent box diagram in words — not TypeSafe’s unpublished internals, not a key desk. Confirm the public loop on docs.typesafe.ai. Replicas and “local Jev” binaries are not Jev and not endorsed.
Independent angle (cover, do not clone)
Win delta (architecture-known-unknowns): reference boxes you can implement; no fake internal diagrams.
Boxes
- Clients — product apps; they see enums, not raw Nouls.
- Decision service — auth, timeouts, idempotency, door adapter.
- Questions module — git-versioned instructions/criteria/floors.
- Door — first-party HTTP/SDK or a documented gateway (matrix).
- Trace store — model, usage, probabilities, action, question SHA.
- HITL — review UI.
- Eval job — gold JSONL, pinned model.
Non-boxes (do not add)
- A fictional TypeSafe orchestrator
- A local Jev binary
- A “rationale generator” calling itself Jev
Failure domains
- Door outage — fail to HITL or deterministic default; do not infinite-retry 529.
- Bad questions deploy — 422 spike; revert the questions SHA, not “the model.”
- Alias move —
response.modelchanges; floors may be wrong. - PII leak — too much
state; fix the filter, not the vendor first.
None of these require knowing unpublished model architecture.
Failure domains you can operate without internals
You do not need a mermaid of TypeSafe’s training stack. You need domains you can page:
| Domain | Symptom | First move |
|---|---|---|
| Door outage | 529 / timeout cluster | Fail to HITL or deterministic default; official SDKs retry 429/529 only — do not infinite-loop |
| Bad questions deploy | 422 spike | Revert the questions SHA, not “the model” |
| Alias move | response.model ≠ pin; review rate jumps |
Pin jev-1.13.0 if floors exist; jev-latest can move |
| PII leak | Too much state |
Fix the filter; vendor is not the first ticket |
| Policy mapping bug | Autos look “accurate” and expensive | The enum lives in your decision service |
| Replica confusion | SemIf/OpenJev traces in the same store | Tag processor; those interfaces are not Jev |
Unknown internals stay unknown. We will not draw a fictional TypeSafe orchestrator, a local Jev binary, or a “rationale generator” that calls itself Jev. Patterns that are yours live on architecture patterns; the wiring punch list is implementation checklist; platform ownership is for platform teams.
What each box actually owns
Clients see enums, not raw Nouls. The decision service owns auth, timeouts, idempotency, and the door adapter (POST https://api.typesafe.ai/v1/systemone or a documented gateway). The questions module is git. The trace store holds model, usage, probabilities, action, and SHA. HITL is a UI you build — review templates. Eval is a job on gold JSONL with a pinned model.
LangChain, if present, is an optional harness in the client or service — not a TypeSafe runtime. Multi-region is your service plus whatever residency TypeSafe will state in writing. We will not invent regions.
Vendor list price and RPM figures stay labeled vendor. Schema-safe responses leaving the door are not a proof that the architecture “cannot be wrong.”
When not to use this page’s advice
Do not paste this as an official TypeSafe reference architecture. Do not add boxes we listed as non-boxes. Do not treat an open-weight clone as box 4. Do not skip fail-closed because the diagram looks complete.
What this page does not claim
- Not an official reference architecture from TypeSafe.
FAQ
Where does LangChain sit? Optional harness in the client or service. See LangChain harness.
Multi-region? Your service + vendor residency. Ask TypeSafe; we will not invent regions.
Do we need their unpublished model diagram? No. Operate the door, the SHA, and the enum.
Can we run Jev offline in the same boxes? Hosted Jev is not on-device. A local clone is a different product.
Disclaimer
This is an independent unofficial site and is not affiliated with TypeSafe AI; official documentation is available at https://docs.typesafe.ai. Never treat jev.pro as TypeSafe official documentation. We do not sell, issue, or proxy API keys.
Open-cluster pages are independent field-guide notes. Replicas and third-party interfaces mentioned anywhere on jev.pro are not Jev and not endorsed. Hub: Open. Siblings: architecture patterns, implementation checklist, for platform teams. Canonical: https://docs.typesafe.ai.