DeepResearch Jev Report Evaluation
A research agent evaluates report relevance, evidence and citations with a strict Jev decision adapter.
Added to Jevfast
How it uses Jev
After a research report is produced, the evaluator sends a bounded projection of the research question, report, cited sources, relevant observations, and plan summary to Jev. One request asks five Noul questions about relevance, evidence, citations, sufficiency, and whether more research is needed, plus a Score question about source quality. The evaluator combines these answers into a report score and recommended action.
What Jev decides
Research report evidence review (illustrative request shape)Example answers · not a recorded Jev response · Source ↗
Question 1 · answer relevance
YOUR APP
INSTRUCTION
Does the report directly answer the user's research question?
STATE
A bounded report-evaluation projection: research_question, report, cited_sources, relevant evidence observations, and a summary of plan steps. The example describes source-defined fields, not captured runtime data.
JEV · NOUL
YesNo
Question 2 · evidence support
YOUR APP
INSTRUCTION
Does the evidence directly support the report's main factual conclusions without unsupported inference?
STATE
A bounded report-evaluation projection: research_question, report, cited_sources, relevant evidence observations, and a summary of plan steps. The example describes source-defined fields, not captured runtime data.
JEV · NOUL
YesNo
Question 3 · citation coverage
YOUR APP
INSTRUCTION
Do report citations adequately cover important factual claims using the cited source URLs?
STATE
A bounded report-evaluation projection: research_question, report, cited_sources, relevant evidence observations, and a summary of plan steps. The example describes source-defined fields, not captured runtime data.
JEV · NOUL
YesNo
Question 4 · evidence sufficient
YOUR APP
INSTRUCTION
Is the supplied evidence sufficient to answer the research question responsibly?
STATE
A bounded report-evaluation projection: research_question, report, cited_sources, relevant evidence observations, and a summary of plan steps. The example describes source-defined fields, not captured runtime data.
JEV · NOUL
YesNo
Question 5 · continue research
YOUR APP
INSTRUCTION
Would additional web research likely be necessary to answer responsibly?
STATE
A bounded report-evaluation projection: research_question, report, cited_sources, relevant evidence observations, and a summary of plan steps. The example describes source-defined fields, not captured runtime data.
JEV · NOUL
YesNo
Question 6 · source quality
YOUR APP
INSTRUCTION
Rate the overall quality and directness of the report's cited sources.
STATE
A bounded report-evaluation projection: research_question, report, cited_sources, relevant evidence observations, and a summary of plan steps. The example describes source-defined fields, not captured runtime data.
JEV · SCORE
0
App workflow
- Project report provenance
Require a non-empty final report, identify cited URLs from explicit provenance or collected sources appearing in the report, and select observations tied to those sources.
- Bound and evaluate
Clip the report and evidence to configured payload limits, summarize the plan, then ask all six report questions in one provider call.
- Compose an evaluation
Validate typed answers, combine the Noul values and normalized source-quality Score, and return an evaluation with report/evidence hashes and a recommended action.
Why it is interesting
It evaluates a report against its linked evidence and provenance, hashes the report and evidence for an evaluation signature, and makes the evaluation result part of an auditable research workflow rather than relying on a free-form critique alone.