## Evaluation metadata
- Project/status/date:
- Model, provider, version, configuration, and code commit:
- My contribution and AI assistance:

## Problem and scope
Intended user, workflow, and consequence of failure.

## Baseline
Manual, rule-based, or keyword baseline. Why might AI add value?

## System and data flow
Corpus, retrieval, access rules, context assembly, and outputs.

## Dataset and provenance
Source, permissions, sample size, case mix, and exclusions. Separate development and held-out data.

## Evaluation method
Rubric, metric definitions, denominators, evaluator, human review, and repeated-run conditions.

## Results
Baseline comparison, error counts, latency, and cost under stated conditions. Write "not measured" where appropriate.

## Failure analysis
Representative failures, citation errors, abstention, denied requests, and boundary cases.

## Reproduction
Dependencies, safe fixtures, commands, expected output, and version information.

## Operational and security considerations
Secrets, permissions, logging, fallback behavior, monitoring, and cleanup.

## Limitations and next decision
What does this evaluation not establish? What should be tested next?

## References and correction history
Link permitted artifacts and record substantive changes.