cards: ling duel + v7.1, ornith 1.5 PPLs

docs/README-Ling.md

@@ -39,6 +39,7 @@ MoE-specific handling: routers (`ffn_gate_inp`, `exp_probs_b`) pinned at F16 out
 | `Ling-3.0-tiny-BF16-ASHQ1-5000.gguf` | 4861 MiB | **13.0681 ± 0.1051** | Bottom-up, best of the duel — smaller and better |
 | `Ling-3.0-tiny-BF16-TOPDOWN-5000.gguf` | 4999 MiB | 13.1094 ± 0.1054 | Top-down challenger (local only) |
 | `Ling-3.0-tiny-BF16-TOPDOWN-6000.gguf` | 6013 MiB | **13.0001 ± 0.1047** | Top-down, first MoE quant |
+| `Ling-3.0-tiny-BF16-TOPDOWN-6000-v7.1.gguf` | 6016 MiB | 12.9802 ± 0.1045 | Top-down + phase-2 slack filling (local only) |
 | `Ling-3.0-tiny-i1-Q6_K.gguf` | 6198 MiB | 13.0123 ± 0.1049 | Uniform Q6_K with imatrix (reference, local only) |
 
 ## Bottom-up vs top-down (@5000, same budget)

docs/README-Ornith-MTP-1.5.md

@@ -41,7 +41,7 @@ ASHQ1 quantization of Ornith with **explicit MTP (Multi-Token Prediction) handli
 | `Ornith-1.5-9B-MTP-BF16-TOPDOWN-6500.gguf` | 6509 MiB | **8.6337 ± 0.0611** | Experimental top-down mode, norms at Q4_K — same PPL |
 | `Ornith-1.5-9B-MTP-BF16-ASHQ1-4500.gguf` | 4518 MiB | **10.1077 ± 0.0719** | Compact, IQ tiers via `--allow-q3-or-lower` |
 | `Ornith-1.5-9B-MTP-BF16-ASHQ1-4500-TOX.gguf` | 4518 MiB | **9.9629 ± 0.0709** | Same budget, sub-4-bit toxicity (`TOXICITY_SUB4`) — 99 → 79 sub-4 tensors, −0.145 PPL |
-| `Ornith-1.5-9B-i1-Q6_K.gguf` | TBD | TBD | Uniform Q6_K with imatrix (reference) |
+| `Ornith-1.5-9B-i1-Q6_K.gguf` | 7209 MiB | 9.2025 ± 0.0677 | Uniform Q6_K with imatrix (reference) |
 
 ## Benchmark vs ASHQ1-Remix (same model, same wiki.test.raw band)