Read your first report
Read status, context, evidence and limitations before drawing a conclusion.
v0.2.1Local CLI · Offline reports
On this page
Check status and configuration
| State | Meaning |
|---|---|
| ingestion_only | No questions; retrieval was not executed. |
| retrieval_inspection | Actual retrieval; relevance and quality are unscored. |
| reviewed_evaluation | Reviewed original-span evidence was validated and scored. |
Confirm loaded/skipped files, failed extraction, question counts, k, chunking, models and context budgets. Low lexical scores are observations, not probabilities or command failures.
Prioritize changed questions
| Class | Interpretation |
|---|---|
| fail → pass | All required evidence became available. |
| pass → fail | Required evidence was lost: investigate the regression. |
| pass → pass | Both contexts are sufficient; ranks/resources may differ. |
| fail → fail | The gap remains; inspect partial gains and missing spans. |
Inspect the actual context
Read original source locations, exact returned text and stage traces. A parent can restore missing qualifications while increasing context size. Compression can drop a necessary condition. A valid model output can still be unhelpful.
Treat diagnosis as a hypothesis
An observed change does not prove a universal cause. Use a relevant next experiment and reserve separate heldout data before making a decision.
Was something unclear? Send documentation feedback.