# Direct mid construction verification

- `cargo test --workspace --release`: 415 passed, 7 ignored, including doctest.
- Full log: `/tmp/ipu-direct-mid-workspace.log`.
- Repeated F16 MLP (64 selected tiles, 17 tokens, dimensions 64/128, three iterations): hardware/reference PASS; max absolute error 0.000183. Modeled final cycles 32613, exchange 18087, both identical to the pre-refactor fixture.
- Small FP8 ViT (64 selected tiles, three layers, fused QKV): hardware/reference PASS; cosine 0.997459656 and max absolute error 0.189453, identical to the pre-refactor fixture. Modeled final cycles 335983, exchange 119296, also identical.
- These are randomized smoke fixtures, not full-size trained-model performance measurements.
- The initially unpriced MLP transition chose a different schedule (37518 modeled cycles). Its package, log and profile are retained with the `unpriced-transition` suffix. Restoring the existing whole-input transition estimate in baseline ranking recovered the previous selection without reintroducing deferred mid values.
- Checkpoint regression loads version-one context/recipes containing the removed deferred-view metadata, preserves choices/attempts, and still rejects changed configuration.