TECHNOLOGY BRIEF / 19 SEP 2026Fast Decisions.
Not More Chat.
A specialist for the small judgments inside a bigger AI workflow.
Give Jev context and a fixed set of questions. It returns choices, scores, and probabilities—not paragraphs. Hermes keeps the planning, writing, and execution.
ClassificationRoutingRelevance checks
Where It Fits
A sanitized ticket, log summary, or candidate document.
Which queue? How urgent? Is this relevant?
Validate the answer; retain existing approval rules.
Proposed use pattern. The installed plugin is explicitly invoked; it does not automatically intercept messages or route background work.
Choice
Pick from options supplied by the application.
Example question: engineering, sales, administration, or needs review?
Noul
Return the probability that a yes/no statement is true.
Example question: does this incident need prompt attention?
Score
Rate the input against an ordered rubric.
Example rubric: no interruption → some users affected → everyone blocked.
Why The Speed Claim?
Jev produces its bounded outputs in parallel instead of writing a response token by token.
TypeSafe reports 70–500 ms and 40–200× faster responses on suitable System One tasks. Its headline workflow comparison reports 193.6×; the company says that is near the high end of likely real-world gains.
Vendor results, not a ForgeFX benchmark. Network time, fallback calls, and the fraction of work moved to Jev determine actual savings. This does not mean the whole agent becomes 100× faster.
Why “Jev”?
The name refers to William Stanley Jevons: improving efficiency can increase total consumption by making more uses worthwhile.
TypeSafe’s bet: cheaper decisions make it practical to put intelligence into many more software steps.
Typed output does not mean infallible judgment. A valid label can still be wrong. Confidence must be evaluated on the actual workload.
Our Verified Setup
- Hermes: v0.21.3; meets the plugin’s declared minimum.
- Installed and enabled: community Jev plugin v0.1.2, pinned to
28a4ea39. - Local check passed: discovery, import, and registration of one tool; no hooks.
- Live inference blocked: no dedicated TypeSafe/Jev credential found. Example invocation returned
not_configured.
Next requirement: a TypeSafe API key from an account with access. Setup selects the jev-latest model alias; the exact serving version is known only after a successful response.
Cloudflare is an alternative, requiring a dedicated Workers AI token plus AI Gateway/Unified Billing configuration. Existing general Cloudflare credentials are not a substitute.
No main-model change, paid inference, or automatic routing was enabled. Current Slack tool catalog does not expose Jev; the installed CLI is available. Native gateway exposure needs a restart and catalog check.
Enable Live Examples
Run setup in a private interactive terminal on the Hermes host. It prompts for the key without echoing it and makes one small, potentially billed validation request.
hermes jev setup --backend typesafe
Do not paste credentials into Slack. Direct TypeSafe usage is separate from the Codex subscription. Published pricing is $0.042 per million input tokens, with output tokens free; confirm current account pricing before ongoing use.
Prepared Example
A synthetic login outage asks all three questions in one call. No real customer content is included.
hermes jev evaluate --file \ /Users/forgebot/.hermes/artifacts/\ jev-explainer/task-triage.json
No model answers or timings are shown because no live Jev request succeeded. A successful smoke test would establish access—not production accuracy.