wuyoscar Find and verify source evidence
Use safely
Choose the service once and keep that choice. If unset, ask A: real Jev via
OpenRouter (OPENROUTER_API_KEY) or TypeSafe (TYPESAFE_API_KEY), or B: simulation
with this agent or an explicitly chosen available model such as DeepSeek. Wait for
consent; errors do not authorize switching. Check key presence only, never values.
Real calls send evidence and cost money; get approval before sending private data.
For B, skip CLI/API calls. Mark agent_simulation or model_simulation, identify
the actual model when available, set jev_called: false, probability: null and
confidence: null. Return a value, evidence-based reason and needs_review; use
null/review when evidence is missing. Do not invent Jev output or probabilities.
Choice uses supplied labels, Noul uses booleans, Score uses integer rubric indices.
For A, use the existing jev-decide CLI with the chosen --provider openrouter
or --provider typesafe. If absent, explain the dependency; do not silently install.
--dry-run is offline validation, not a judgment. Exit 0 means selected/scored,
2 means review, 1 means error. Read each value: false Noul remains false. Selection
is not permission, and confidence is not accuracy. Keep unknown/review paths.
First request
Adapt the example. The shared CLI needs Python 3.10+;
no sibling skill is needed. Host tools still own collection and actions.
Resolve <skill-dir> to this installed folder:
jev-decide decide <skill-dir>/assets/example.json --dry-run
# After approval, send the edited request with the selected provider:
jev-decide decide /path/to/request.json --provider openrouter
Choose the evidence workflow
- Documents: follow the workflow below for original spans, passage relevance and claim checks; adapt the document template.
- Code locations: read code-location selection and adapt the code-location template. Use the project's required graph/index tools first. Inspect selected code before making claims.
- Judging whether a change is correct: use
jev-evalif installed, rather than treating a relevance score as a code-review result.
Workflow
- Read the authorized source and retain document/page/line identifiers. Have parsers or regex produce exact candidate spans when possible.
- Define the requested role precisely: invoice destination is not any email address. Include none when no candidate fits.
- Use independent relevance questions when ranking all passages; winning a relative Choice does not establish an answer exists.
- Copy the original span selected by ID. Do not ask Jev to synthesize the extracted field or fabricate a quotation.
- Check each claim against its cited evidence separately. Report unsupported/contradicted statements and preserve source links for human checking.
Context and parallelism
Jev does not inherit the agent's history. Give every request sufficient context: the user's information need, exact claim, source IDs, surrounding passages, definitions and relevant exceptions. Supply the text, not just a URL or your own summary verdict. Keep needed cross-references; omit unrelated material and secrets.
Batch independent claim checks or per-passage relevance scores over shared state instead of serial LLM calls. For separate document groups, use bounded concurrency with stable document/question IDs, rate limits and a cost/time budget. The host schedules calls; the CLI has no parallel scheduler. Questions cannot read other answers in the same request: fetch a selected source before asking about unseen contents. Use Jev's low latency for repeated judgments, not document generation.
Make it yours
Replace the example's evidence, candidate IDs and criteria together. Preserve a no-match route when the real task can fall outside the labels. Agree on how the host or person consumes each answer before enabling any automatic effect.
Precedent
Related project or author example. Our workflow is an adaptation, not that project's code, an automatic installer, or a reproduced benchmark. OpenRouter request contract.
Examples
Rerank search and retrieval results · Repository navigation · Find meaning on a page, not just matching words
More tasks and local templates. Open only the matching row; there is no need to read the full README before a judgment.