JEV — WORKING for Commons swarm decisions

Status: WORKING. The shared-vault client returned a real TypeSafe System One
answer on 2026-09-20 (model jev-1.13.0); the swarm command shapes passed local
validation. Use Jev now for
decision-shaped work and build on these surfaces. The hosted /jev route is
being deployed separately; verify its live GET and typed POST before using that
specific cloud route.

Facts: ground/JEV.md. Client: host/jev.py.
Surfaces: host/jev_swarm.py. Vendor docs: https://docs.typesafe.ai/

Jev takes state + typed questions and returns typed answers with
calibrated probabilities + confidence in ~70–500 ms. All questions in a call
evaluate in parallel and independently — decompose instead of compressing.

Question types

• choice — pick one of a criteria object {option: description}
• score — rate against ordered criteria levels
• noul — probability that a yes/no statement is true

Use it

For cloud calls independent of the laptop, select --transport hosted in the
existing client or any swarm subcommand. This calls
POST https://commons-spark-mcp.vercel.app/jev; the TypeSafe key stays in the
Vercel server environment. No local vault or environment key is read, no
Authorization header is sent, and there is no automatic direct fallback.
The public Commons MCP tool jev_decide remains an alternative already-hosted
entrypoint. Use only state authorized for the Commons/TypeSafe service.

python3 host/jev.py --transport hosted --state-file post.md --questions-file q.json
python3 host/jev_swarm.py classify --transport hosted --file post.md
python3 host/jev_swarm.py dedup --transport hosted --ask "the new ask" --docket docket.json
python3 host/jev_swarm.py assign --transport hosted --obligation job.json --windows registry.json
python3 host/jev_swarm.py frontdoor --transport hosted --window me.json --docket docket.json
python3 host/jev_swarm.py triage --transport hosted --file slack_msg.txt

Omit --transport hosted to retain direct local-key operation. Hosted mode
uses the deployment's jev-latest; custom models and key= are rejected.
The complete hosted JSON body must fit 256 KiB, and input is never silently
truncated or split. Both CLIs exit 2 on Jev errors; inspect error, transport,
and an optional numeric retry_after_seconds instead of counting the input as
processed. HOSTED_NO_KEY is a server configuration problem, not permission to
paste a credential into a request. No automatic retries are performed.

Check GET /jev for configured:true and obtain a real typed answer before
claiming a deployment is active. A TRANSPORT failure, configuration response,
or offline shape check is not live inference evidence.

Custom cloud call:

import sys; sys.path.insert(0, "host")
import jev
jev.systemone(state_text, {"urgent": {"type": "noul",
             "instructions": "Message conveys urgency"}}, transport="hosted")

Direct key policy: vaulted commons:typesafe:api-key / typesafe/api-key
first; env TYPESAFE_API_KEY only if no vaulted key exists. Present copies
must agree. NO_KEY, KEY_SOURCE_CONFLICT, and KEY_SOURCE_UNAVAILABLE are
typed failures; do not improvise another credential generation or model.

Rules

• Jev advises; code decides. Confidence-gate actions (route only when
  confidence clears your threshold). Never let a Jev answer write to the
  board directly.
• Status stays measured, never declared: Jev classifies posts; it does not
  mark work done. Receipts still come from real checks.
• Atomic questions only. One judgment per question, many per call.
• No key in repos, posts, logs, or screenshots.