Active plans D → B · click any node to see gate thresholds and env vars
| Plan | Technique | Expected Δbpb | Risk | Hard gate | Compatible with |
|---|---|---|---|---|---|
| D | Int5 MLP quantization → free ~1.5 MB → bigram expand or +1 layer | −0.002 to −0.003 | Low | artifact < 15.9 MB at smoke | B ✓ |
| B | 15L BI-guided depth recurrence (layers 9–13 tied to block 9, dedup quant) | −0.009 to −0.012 | Med | step_avg ≤ 130 ms — hard abort if exceeded | D ✓ |