Is the stand-in emulator calibrated like real Jev?

No. The emulator reuses the agent LLM in the same format and is explicitly marked uncalibrated. Real probabilities, confidence, latency, and calibration flags land in the decisions table for later measurement.

Will a confident score let a risky write skip humans?

Never. Reads execute directly and reversible writes pass a Jev adequacy check, but high-risk irreversible actions always require Telegram approval regardless of probability.

Do long threads blow past the API token limit?

No. A token budget shrinks state from history down through attachments and tool results to fit before any Jev call, and whatever was cut is recorded in the item notes.

Does passing replay evaluation prove the model is good?

No. The versioned cassette is rebuilt from labels, so replay fails on prompt, question, taxonomy, or plumbing changes rather than measuring model quality or catching accuracy loss.

What is the default confidence routing?

Category confidence decides the path: 0.85 or above follows automatically, 0.55 or above sends the draft to a human for approval, and below that escalates, with per-category overrides in the environment.

How does it verify its own draft?

Jev answers a verification set about whether the draft resolves the request, its quality, and any unsupported claims; a failure regenerates once with feedback naming the failing fact, and a second failure escalates.

How are facts checked one by one?

The agent ends its reply with a ---FATOS--- line listing the claims it makes, each becomes a yes/no Jev question in the same call, and one unsupported fact fails the draft; backoffice show lists the facts with the verdict.

CommunityJev Agents & MCP

Backoffice Agents

Backoffice email pilot where the LLM generates and uses CRM and ERP tools while Jev triages, gates tool calls, and verifies drafts before sending, with Telegram human approval.

See highlights

Open github.com

Stars: 1
Language: Python
Published by: Community
Author: lucianoon
Last verified: Oct 2026

lucianoon/backoffice-agents

User repository on GitHub

View on GitHub

Piloto de agentes de backoffice: LLM gera e raciocina, Jev (TypeSafe AI) decide, roteia e verifica

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Piloto de agentes de backoffice em que o LLM gera e raciocina e o Jev decide, roteia e verifica. Entrada por e-mail, ferramentas de CRM e ERP, operador humano pelo Telegram. Jev é o modelo System One da TypeSafe AI: não gera texto, responde perguntas tipadas (sim/não, escolha, nota) com probabilidades calibradas, em 70 a 500 ms, a US$ 0,042 por milhão de tokens de entrada. Enquanto a chave não chega (early access), o projeto roda com um emulador que usa o próprio LLM no mesmo formato, marcado como não calibrado.

Quoted from the project's README. Read the full README.

Tagscommunity github

Works forJev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

One-call triage asks category, urgency, human need, sensitive data, and injection; confident spam is dropped and sensitive cases escalate.
Category confidence bands route automatically, to human review, or to escalation, with tenant versions recorded per decision.
Per-tool gate lets reads run, checks reversible writes with Jev, and forces human approval for irreversible actions.
Draft facts become per-fact noul checks naming the failing claim; a second failure escalates instead of retrying forever.

Watch out

While the API key is unavailable an uncalibrated LLM emulator stands in for Jev; eval sets are synthetic and replay checks plumbing, not quality. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

agent-guard

Claude Code PreToolUse hook that screens Bash, file edits, and tool calls with Jev risk, policy, and allow-ask-deny judgments before execution.