[shipitdev]

Changes

• Bound standalone generation to one writer and one verifier.
• Format existing valid writer-selected citations in code, preserving claims and verifying the resulting answer.
• Strengthen scope, completeness, source-support and version instructions. Withhold factual rejections without repair calls.
• Publish the fixed-cohort comparison and expose provider/model labels in saved dashboard outputs.

Evaluation

• Same-model 28-case comparison released 14 to 18 answers and reduced citation-format failures from six to zero.
• Mean model stages fell from 2.91 to 2.00 across 11 shared released answers.
• Two separately labeled TokenHarbor fallback cases complete the 30-question cohort. One Groq quota failure remains.
• Paired reference correctness declined on four cases. Reference alignment and changed withholding require review. No broad quality-gain or held-out release claim.
• Further judge calls stopped on user direction. Cache-only assembly used zero model requests. Relevancy was not rescored.

Validation

• 160 Python tests and three subtests passed, with two optional skips.
• Separate RAGAS environment checks passed.
• Asserted unchanged evidence/references, provider separation, budgets and redaction.
• Saved results published to the existing TechQA dataset.