docs: Qwen3.5 graft Report 13 cross-attention results

docs/benchmarking/qwen35-grafting-report-13.md

@@ -0,0 +1,83 @@
+# Qwen3.5 graft Report 13: causal cross-attention fusion
+
+## Summary
+
+Report 13 evaluated a causal cross-attention graft from Qwen3.5-27B into the
+complete Qwen3.5-9B host. The graft improved the combined MMLU-Pro and BBH score
+at 100M continuation tokens, but the improvement did not persist through the
+250M checkpoint.
+
+The best checkpoint was 100M:
+
+- MMLU-Pro: `0.514628` (`+0.000582` versus the 9B baseline)
+- BBH: `0.623503` (`+0.005207` versus the 9B baseline)
+- Combined: `1.138131` (`+0.005789` versus the 9B baseline)
+
+The 250M checkpoint fell to a combined score of `1.122947`, which is
+`-0.009395` below baseline. The result supports checkpoint selection and early
+stopping; it does not support treating longer continuation as monotonically
+beneficial.
+
+## Configuration
+
+The experiment kept the complete Qwen3.5-9B host and inserted frozen donor
+layers 31--33 from Qwen3.5-27B after host layer 15. The donor layers executed on
+every forward pass. A trainable 512-wide, eight-head causal cross-attention
+module fused donor information into the host stream, with a maximum residual
+scale of `0.05`.
+
+The host and original donor parameters remained frozen. Training was restricted
+to the input/output projections and report-specific graft parameters. Each
+checkpoint used the same protocol identity:
+`b82fc1daf569d506e0a1ff24d1e96674879315248311badd10966164a685e111`.
+
+Evaluation covered the complete deterministic task sets:
+
+- MMLU-Pro: 12,032 samples
+- BBH: 5,761 samples
+
+The baseline and baseline-repeat evaluations were identical.
+
+## Results
+
+| Checkpoint | MMLU-Pro | BBH | Combined | Combined delta vs. baseline |
+|---:|---:|---:|---:|---:|
+| 9B baseline | 0.514046 | 0.618295 | 1.132341 | -- |
+| 1M | 0.512799 | 0.617428 | 1.130227 | -0.002115 |
+| 50M | 0.512716 | 0.616039 | 1.128755 | -0.003586 |
+| **100M** | **0.514628** | **0.623503** | **1.138131** | **+0.005789** |
+| 150M | 0.509475 | 0.616560 | 1.126034 | -0.006307 |
+| 200M | 0.514295 | 0.618990 | 1.133285 | +0.000944 |
+| 250M | 0.508644 | 0.614303 | 1.122947 | -0.009395 |
+
+## Interpretation
+
+The cross-attention graft can outperform the frozen host at a selected
+checkpoint, driven mostly by BBH. The trajectory is unstable: the 100M gain
+disappears at 150M, partially returns at 200M, and reverses at 250M. This makes
+the 100M result evidence for a useful intermediate state, not evidence that the
+architecture reliably improves with additional continuation.
+
+Follow-up work should evaluate the sealed 100M checkpoint on the campaign's
+held-out tasks and compare multiple seeds before making a broader capability
+claim. Promotion rules should select on a fixed validation protocol and retain
+the selected checkpoint rather than automatically using the final checkpoint.
+
+## Artifact evidence
+
+The local campaign artifacts are not committed to this repository. The hashes
+below bind this report to their immutable completion receipts.
+
+| Checkpoint | Training completion SHA-256 | Evaluation completion SHA-256 |
+|---:|---|---|
+| 1M | `880dbbef7173c3356a9b42ed50946f7856ea7dcfa8b752726a98d97b13be002b` | `d6c1b16825dac97fcd42e62cf9bc61ec84b13b05f850fe6488a29cc79e11dcf4` |
+| 50M | `dc23d8d18c4179dc96e68729ff66538c8599fac47344875a15cd157d922800dd` | `bacd0deeb4aaf986b6cbb5338af0bac3d0709b0f00eab469da7af542a5ce21ba` |
+| 100M | `ea07790362318cb0e59b7c6804d1d5ccf9bd33062d3516872fc299318f7a7c32` | `414b5821cddcefe0ae9766ce811b41040dfda4bbe27648c6443f834227edee60` |
+| 150M | `f5c4f3f64cc971524064e9ac55a4e8ba10947f7a97aba0a569c85e0d19f6a6fb` | `f69f219eef484681e92949f59b9c4a24e52c765efb0f36492f8cc4c408c7d1b5` |
+| 200M | `07da1faeeeacd8304d3292abc3c1bc85fdd4d22078958ada53ffd383b4b5d73b` | `9a4a96ef9f6fd7f8f9ffb6d8b9e79234b17208f3f67a0a5be5f279058a4cc2f7` |
+| 250M | `1f408110f1680e234a426a0ade31c113cd311b3384096ecaeff3433de5eb933b` | `a4c0f9cda514bafe31dcf9128e80e27914875b9c9ddec2e112c0af765a48851c` |
+
+The runtime receipt binds runner SHA-256
+`e727e864c0703be6003a0e707983660e05d24cb1e2d64e56a7dcc2cbd9eaa950`
+and core SHA-256
+`a3cdcd11273b87ed6edbaf48990e304a915767a9de1c9cce7ec8d11fd6364b8a`.