diff --git a/docs/smollm2_tile128.md b/docs/smollm2_tile128.md
new file mode 100644
index 0000000..2eff8c4
--- /dev/null
+++ b/docs/smollm2_tile128.md
@@ -0,0 +1,75 @@
+# SmolLM2 GEMV tile128 schedule experiment
+
+`examples/smollm2_gemv_tile128.json` selects the existing `N_tile=128` schedule
+knob. The configuration is opt-in; the default lowerer, CUDA VM, ABI, oracle and
+kernel knobs are unchanged.
+
+On the measured SmolLM2 workload, default tiling produced 5,286 tasks per step
+and tile128 produced 1,716, with 848 buffers and 38,912 bytes of dynamic shared
+memory in both arms. A wider tile can reduce schedule/table setup work while
+reducing GPU parallelism. The result below measures complete generation; it
+is not a GPU-only GEMV claim.
+
+## Using the configuration
+
+The existing evaluator accepts the preset:
+
+```bash
+uv run amk eval HuggingFaceTB/SmolLM2-135M-Instruct --gpu rtx5090 \
+  --config examples/smollm2_gemv_tile128.json
+```
+
+That command evaluates one forward using the registered target. It does not
+reproduce the complete-generation experiment or substitute an RTX 5090 target
+for a PRO 4000. The archived generation runner reads the live SM120 device
+limits and passes the same `ScheduleConfig` to the unmodified lowerer at each
+step. `generate` currently passes `config=None`, so reproducing the full request
+requires that external adapter; this PR adds no new generation API.
+
+## Recorded complete-generation experiment
+
+Frozen upstream: `a514bbc20a03bbf698a17443f8f14a27a617fc10`.
+Model: `HuggingFaceTB/SmolLM2-135M-Instruct`, revision
+`12fd25f77366fa6b3b4b768ec3050bf629380bac`, all 30 layers and 134,515,008
+parameters, FP32 weights/activations, no quantization. Six checkpoint file hashes
+are included in the evidence. Runtime: Python 3.12, Torch 2.13.0+cu130,
+Transformers 5.16.1, CUDA 13.0, one RTX PRO 4000 Blackwell 24 GB (SM120).
+
+Four independent processes ran default/tile128/tile128/default, with independent
+first-build directories. Context lengths were 8 and 64 tokens; each request
+produced 8 new tokens. All 344 step-level logits comparisons passed the existing
+FP32 `atol=rtol=1e-4` oracle against both CPU ReferenceVM and full eager.
+All 64 correctness-stage generated tokens matched exactly. Only after both
+workloads passed did each process perform one warmup and three timed requests
+per context; timed and warmup sequences also matched.
+
+| Context | Process pair | Default median seconds | Tile128 median seconds | Latency reduction |
+|---|---|---:|---:|---:|
+| 8 | A0/P1 | 3.486624 | 1.778946 | 48.98% |
+| 8 | A3/P2 | 3.515794 | 1.755405 | 50.07% |
+| 64 | A0/P1 | 16.170751 | 7.911105 | 51.08% |
+| 64 | A3/P2 | 15.953324 | 7.895794 | 50.51% |
+
+Allocator peak: 1,320,648,704 bytes in each process. The two process-pair
+geometric means are 49.53% and 50.79% latency reduction for contexts 8 and 64.
+The baseline last/first median changes are +0.84% and -1.34% respectively.
+These are exploratory observations from two process pairs, not a confidence
+interval or independently confirmed stable speedup.
+
+Timing includes per-step lowering/validation, VM and table construction, weight
+H2D, KV transfer and sampling from a CPU-resident model. It excludes disk model
+loading, first CUDA compilation and correctness instrumentation. The existing
+path reconstructs the VM each step; this is not GPU-resident steady-state
+decode, concurrent service throughput, CUDA Graph replay, or a comparison with
+cuBLAS/vLLM.
+
+The preset passed the experiment's advancement rule: every context/process
+pair improved by at least 1%. Further independent validation is required before
+changing a default. A separate `cpa_stages=2` candidate regressed at context64
+and is not included.
+
+[All timings, per-step comparisons, checkpoint hashes and the original runner](https://gist.github.com/0z5a/7b6a65fb1899b27411706bb402201016)
+are published together. The archive documents the exact directory layout and
+commands for replaying all four processes. The compact checked-in
+results (record retained locally) retain every recorded request time and
+original receipt hashes. No GPU test was rerun for publication.
diff --git a/examples/smollm2_gemv_tile128.json b/examples/smollm2_gemv_tile128.json
new file mode 100644
index 0000000..9da9f77
--- /dev/null
+++ b/examples/smollm2_gemv_tile128.json
@@ -0,0 +1,7 @@
+{
+  "tiling": {
+    "gemv": {
+      "N_tile": 128
+    }
+  }
+}