Study the diff-surface redesign for large changes, especially rendering, navigation and inline-comment handling in review-heavy workflows.

Assess split inference for robotics workloads by measuring task success, latency, network dependence and onboard resource use.

Benchmark prefill-only decisions on text and images, measuring latency, calibration and option-order stability before safe rollout.

Treat reported scientific assistance as a workflow lead, not validation: separate model contribution from full laboratory replication.

Expose catalog search and read-only diagnostics first; require dry-run confirmation and per-user authorisation for every write action.

Use one compatible endpoint to compare local models, but record routing rules, warm-up time and failure behaviour in logs per request.