All 497 real tensors, all 7 real plans side by side: Production, Plan A/B/C (original, real default --pareto meaningful + --hessian-weight-strength 2.0), and A2/B2/C2 (same real settings, corrected: --pareto none + --hessian-weight-strength 0). Spread is now the max-min bits across all 7. Raw Hessian/KL signal data kept at the end as reference only.
| name | layer | component | Prod bits | A bits | B bits | C bits | A2(nw) bits | B2(nw) bits | C2(nw) bits | spread | Prod hessian floor | A hessian floor | B hessian floor | C hessian floor | A2(nw) hessian floor | B2(nw) hessian floor | C2(nw) hessian floor | Prod early qkv floor | A early qkv floor | B early qkv floor | C early qkv floor | A2(nw) early qkv floor | B2(nw) early qkv floor | C2(nw) early qkv floor | Prod late attn floor | A late attn floor | B late attn floor | C late attn floor | A2(nw) late attn floor | B2(nw) late attn floor | C2(nw) late attn floor | param count | hess score | hess rank | kl4 | kl rank |
|---|
Plan A/B/C used the real default --pareto meaningful and --hessian-weight-strength 2.0. Plan A2/B2/C2 are the identical real commands corrected with --pareto none and --hessian-weight-strength 0 (2026-09-18 finding).
| Parameter | Prod | A | B | C | A2(nw) | B2(nw) | C2(nw) |
|---|---|---|---|---|---|---|---|
| Achieved BPW | 5.028552 | 5.028571 | 5.028574 | 5.028574 | 5.028443 | 5.028568 | 5.028568 |
| --pareto | meaningful | meaningful | meaningful | meaningful | none | none | none |
| --hessian-weight-strength | (n/a) | 2.0 | 2.0 | 2.0 | 0.0 | 0.0 | 0.0 |
| --hessian-primary | no | yes | yes | yes | yes | yes | yes |
| Hessian floor tensors | (none) | 61 | 60 | 61 | 61 | 61 | 61 |
| Late-attn floor tensors | 76 | 76 | 76 | 0 | 76 | 76 | 0 |
| Early-QKV floor tensors | 48 | 0 | 48 | 0 | 0 | 48 | 0 |
| Raw Sigma isolated KL | 37.45 | 50.517 | 53.114 | 52.25 | 38.889 | 50.361 | 50.093 |
| OptiQ-weighted Sigma KL | 57.136 | 388.189 | 412.61 | 411.397 | 58.575 | 221.598 | 221.33 |
| Parameter | Prod | A | B | C | A2(nw) | B2(nw) | C2(nw) |
|---|---|---|---|---|---|---|---|
| Q4 | 236 | 291 | 268 | 309 | 221 | 262 | 298 |
| Q5 | 92 | 114 | 121 | 83 | 80 | 114 | 79 |
| Q6 | 13 | 29 | 43 | 33 | 35 | 37 | 33 |
| Q8 | 3 | 20 | 23 | 21 | 19 | 20 | 19 |
| Q16 | 153 | 43 | 42 | 51 | 142 | 64 | 68 |
Note: Plan C2 was re-run (2026-09-18) with the correct --hessian-weight-strength 0 -- the file now correctly matches A2/B2 (weighted ΣKL 221.3, consistent with the other clean strength=0 runs). Raw KL remains the one to trust for any real comparison regardless -- weighted ΣKL is dominated by a handful of tensors where --protection-multiplier (100×, boundary/lm_head only) and Hessian strength compound multiplicatively; it barely reflects the other ~492 tensors even when clean. Full mechanism: MILP_AWARE_HESSIAN_GUIDE.html.