# Limitations

- One attempt was recorded per cell.
- Provider route and inference settings are confounded with model identity.
- Qwen3-32B is an external anchor, not a jointly randomized member of the paid multimodel denominator.
- Qwen3 8B has paired coverage 18/19 and no registered interface-advantage replication.
- DeepSeek V4 Pro is infrastructure-unassessable because 38/38 cells failed at the privacy/data-policy route layer with empty final answers; do not treat it as a coding-performance denominator.
- The model-strength interaction is descriptive only, and CROSS_MODEL_GENERALIZATION_SIGNAL is false.
- The package is not a production-readiness claim and not a live controlled-evaluation runner.
- Low-affinity tasks are newly authored for calibration where supported; high-affinity tasks are adapted or reconstructed from upstream issue/task patterns with upstream bytes withheld.