jev is PortOS's local entailment scorer: give it a premise and a list of hypotheses, get back a per-hypothesis entailment distribution and an argmax with a confidence margin. It answers closed-set questions — "which of these options does this text entail?" — with no text generation, no tool surface, and no provider quota. See the ADR local jev decision service for why it exists and what was rejected.
It is an optional feature, off by default. Turn it on in Settings >
Features or Models > LLMs > jev. The management page and navigation stay
visible at /models/llms/jev even when integrations are disabled. Browsing the
page never installs or loads the model.
| Model | AlexWortega/openjev, subfolder qwen3.5-4b-nli, revision f8187e6e11d413d0771bcc7970b85f78e194264c |
| License | MIT, ungated — no Hugging Face token or model-card approval |
| Download | 4 files, ~9.1 GB (weights dominate). The repository's 35B variant, its trained MLP heads, and its code/ directory are never fetched and never executed |
| Runtime | A dedicated virtualenv at data/python/venv-jev (fallback ~/.portos/venv-jev) — separate from Prompt Guard's and from the image/video runtimes |
| Packages | torch, transformers, safetensors, huggingface_hub, pinned in server/lib/jev.js |
| Port | PORTS.JEV = 5566, loopback only, bound on demand |
Nothing above happens at boot, on a status refresh, or on first request. The download runs only from an explicit operator install.
Models → LLMs → jev shows the four install stages (python, venv,
packages, model), the failed runtime check, expected package versions, and
the last failed installation stage with its diagnosis. Overall ready is true
only when the pinned packages import and the pinned snapshot is cached.
For unattended local diagnosis, run from the PortOS checkout:
node scripts/setup-jev.js --status
node scripts/setup-jev.js --installBoth print JSON to stdout and exit 0 only when ready (1 otherwise). Install
progress goes to stderr. --status never installs, downloads, or loads a model.
Run these from the primary installation after updating it — CoS worktrees
deliberately use isolated data and ignore PORTOS_DATA_ROOT.
Install prefers an already-installed uv package installer and falls back to
Python pip, for machines where uv can reach the package index but pip cannot.
Repair retries the dedicated runtime without touching Prompt Guard or the image
and video environments.
Unlike Prompt Guard, which spawns a Python process per scanned item, jev runs a
persistent sidecar: scripts/run_jev.py binds 127.0.0.1:5566, loads the
checkpoint once, and answers JSON scoring requests. A 4B model is far too
expensive to re-import per call.
- It starts on the first scoring call, never at boot.
- Two concurrent first callers produce exactly one process.
- An idle timer unloads it after 10 minutes; Unload now on the panel frees it immediately.
- It refuses to start without its model directory, refuses a non-loopback peer, and refuses a premise that exceeds the model window rather than truncating and answering about the prefix.
- It never echoes the premise or a hypothesis back in an error, and its stderr is drained but never retained — a dependency traceback can carry local paths or the input itself.
decide() returns abstained: true whenever the margin between the top two
options' entailment probabilities falls below the threshold (default 0.15).
A caller must treat abstention as "ask something else" — never as "take the
top one anyway." The margin is the entire difference between "the model
distinguished these options" and "the model produced a number". A caller that
wants an additional floor beneath a clear-but-weak winner applies it on top of
the reported confidence; the service itself gates on margin alone.
The margin needs a runner-up, so decide() requires at least two options. A
yes/no question is expressed as two hypotheses ("…warrants a reply" /
"…warrants no reply"), not one.
Codes, never text. A Python traceback can carry local paths or the premise, so nothing from the sidecar's stderr ever reaches a payload or a log line.
| Code | Meaning |
|---|---|
jev-not-installed |
No dedicated virtualenv, or the pinned snapshot is not cached. Install from Models → LLMs → jev. |
jev-start-failed |
The sidecar could not be started, or died before reporting a loaded model. Check the install stages, then free memory — the checkpoint needs roughly its 9 GB on disk plus room to load. |
jev-timeout |
A scoring request exceeded its timeout (default 30 s). |
jev-request-invalid |
The premise or hypothesis list failed the bounds in server/lib/jev.js, or fewer than two options reached decide(). |
jev-response-invalid |
The sidecar's reply did not match the wire schema, or described a different hypothesis list than the one asked about. |
jev-premise-too-large |
The premise exceeded 32,000 characters, or the rendered pair exceeded the model window. The request is refused, not truncated. |
Install-stage failures reuse the shared Python diagnosis vocabulary —
package-missing, package-version-mismatch, network-failed,
certificate-failed, wheel-unavailable, dependency-conflict, disk-full
(server/lib/pythonRuntimeDiagnosis.js). For network-failed, check Python's
access to the configured package index; a browser or curl succeeding does not
prove Python can connect. For certificate-failed, repair certificate trust
rather than disabling TLS verification.
Declared once in server/lib/jev.js and mirrored into scripts/run_jev.py so
the boundary still holds if the script is invoked directly:
| Premise | 32,000 characters |
| Hypotheses | 32 per request, 512 characters each |
| Default margin | 0.15 |
| Request timeout | 30 s |
| Idle unload | 10 minutes |
It is not selectable as a chat provider. The descriptor is exposed beside
the catalog under its own decisionScorers key — the same guarantee
securityGuards gives Prompt Guard — and is never merged into models, so no
provider or model picker can offer it.
It is also not a replacement for the abuse guard. Prompt Guard screens external content before anything reasons over it; jev is one of the things that may then reason over it. See features/messages-security.md.
The Jev page separates installation, model residency, global integration enablement,
and source policies. Source off retains its legacy meaning: Shadow, scoring
alongside chat for comparison. Disabled (disabled) actually suppresses scoring
for that source. Prefer local uses chat on abstention/failure; Local only
skips unresolved items. Existing policies are unchanged until explicitly edited.
Issue replies and forge maintenance share the GitHub issue policy; message action and priority share email policy; Stacker News classification and risk share its policy and remain escalation-only. Scope adherence has a separate on-demand switch (default enabled under the global gate). It provides issue/PR advisories, not agent completion grades. Completion goal fidelity remains the separate chat reviewer in Models > Code Reviewers, with forge verification for merge objectives.
The page displays the shipped hypotheses directly from the decision registry and explains margin floors. Agreement is comparison with chat, not measured correctness; missing comparisons are not zero accuracy. Metrics and runtime status refresh while the page is visible. Trained-head evaluation remains a separate held-out comparison.