Understanding Jev

What it is and what it answers

Jev is reached over HTTPS at https://api.typesafe.ai/v1/systemone. A call sends a compact state (a redacted description of the decision context) and one or more typed questions, and gets back a structured judgment — for example a Choice between named candidates with probabilities, a Score against a rubric, or a Noul (a tested yes/no condition). It is a second opinion on semantic questions, not a proof, not a gate, and not a runner: it never writes code, executes commands, or decides a delivery.

Typical supported questions during a delivery: is this finding about the same defect as that one, does this evidence item directly prove the stated requirement, which candidate skill is relevant here. Mechanical work — arithmetic, hashes, dates, parsing, schema checks, symbol resolution — always stays local in deterministic tools.

When it runs

• Questions are asked by the roles and the root assistant at existing decision points. No extra phase, hook or quota is introduced, and no delivery step blocks on the service.
• The client needs the TYPESAFE_API_KEY environment variable. Without a key, or when the network or service is unavailable, the caller records typesafe: unavailable and continues locally. Nothing retries blindly or invents a judgment.
• Source-blind EXAMINE does not use Jev: the examiner works only from its own observations in its public-only workspace.

Authorization

Jev is authorized by default, and the switch is yours, per project:

• Decided at install or enable. On a terminal, nwave-ai install and nwave-ai project enable ask whether Jev may be used in the project and write the answer into the project's .nwave/config.json as "jev": "authorized" or "jev": "unauthorized". The question names the machine-wide value when ~/.nwave/config.json sets one, and an empty answer keeps it. Without a terminal (CI, or --yes) the same command writes the declared default and says so: nWave wrote it and you did not choose it. The default is "jev": "unauthorized" when your machine-wide file refuses Jev, so a machine-wide refusal is never loosened by an install or an enable. A machine-wide value that is not a Jev choice fails closed to "unauthorized", with the reason on stderr. An unreadable machine-wide file stops nwave-ai install before anything is written, and nwave-ai project enable records "unauthorized" with the reason on stderr. The declaration itself adds only the jev entry and keeps every other byte. nwave-ai project enable first runs its own activation write, which sets enabled and the schema version, and the declaration then adds the jev entry to that result. A project that already holds a jev value is never asked about again, and that value is never replaced.
• Two positions per project. "jev": "authorized" allows Jev in that project, and "jev": "unauthorized" refuses it there. The same key in ~/.nwave/config.json is the machine-wide value, used by a project that declares none. The flag --jev unauthorized refuses Jev for one call and wins over both files. The repository tier wins over the global tier. Any other value is refused as malformed, and the feature and epic tiers cannot declare jev. nwdes never writes this key.
• The key is a technical prerequisite, not the authorization. What authorizes Jev is the default policy above plus your refusal switch. The TYPESAFE_API_KEY environment variable only lets the service accept requests. nwdes only checks that the variable is set; it never reads, stores or logs the key. Without the key, nothing is sent and the caller records typesafe: unavailable. Configuring Jev shows how to set it.

nWave's skill policy treats the default authorization as standing for supported task-level questions.

What is never sent

These data classes are never transmitted, regardless of authorization: credentials and secrets, personal data, private findings, source-blind material, secret-bearing logs, and confidential account, host or resource identifiers. State is kept minimal — named, relevant fields, never a repository dump. Reviewers and examiners form their own questions and never inherit producer-owned judgments.

Costs and records

Each call records its model, input/output tokens, duration and cost when the service reports them; unknown cost is not reported as zero. Request records stay local. If knowing what was asked matters to you, ask the root assistant to show the local record.

Where to read next