Add portable ATX layer-expert residency policies

README.md

@@ -29,6 +29,31 @@ Companion artifacts:
 - [`recommendations/atx_expert_residency_policy.json`](recommendations/atx_expert_residency_policy.json): local ATX residency policy artifact, when available.
 - [`upload_manifest.json`](upload_manifest.json): file hashes, source paths, and upload metadata.
 
+
+## Portable ATX Residency Policies
+
+The `policies/` directory contains four portable `--moe-residency-policy` JSON files generated from the combined coding saliency data:
+
+| File | Selected layer-expert cells | Combined importance share |
+|---|---:|---:|
+| [`policies/combined_top_05pct_layer_experts.atx.json`](policies/combined_top_05pct_layer_experts.atx.json) | 512 | 66.1% |
+| [`policies/combined_top_10pct_layer_experts.atx.json`](policies/combined_top_10pct_layer_experts.atx.json) | 1,024 | 80.0% |
+| [`policies/combined_top_15pct_layer_experts.atx.json`](policies/combined_top_15pct_layer_experts.atx.json) | 1,536 | 87.1% |
+| [`policies/combined_top_20pct_layer_experts.atx.json`](policies/combined_top_20pct_layer_experts.atx.json) | 2,048 | 91.2% |
+
+Example usage with the ATX fork:
+
+```bash
+./build-atx-metal/bin/llama-server \
+  -m /path/to/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf \
+  --moe-residency-policy policies/combined_top_10pct_layer_experts.atx.json \
+  --moe-residency-stats stats.json \
+  --ctx-size 4096 \
+  --parallel 1
+```
+
+These policies rank layer-expert cells by mean normalized importance across SciCode MVP and Terminal-Bench Hard Agent traces. They recommend which cells to duplicate/cache hot in VRAM; they are not yet throughput-validated as globally optimal.
+
 ## Important Caveats
 
 These results are observational routing/contribution-proxy measurements. They are useful for proposing which experts or layer-expert cells should remain VRAM-resident, but they are not a causal proof that those policies preserve quality or maximize throughput. Final residency recommendations should be validated with forced-residency/offload runs on the same prompts or agent tasks.

policies/README.md

@@ -0,0 +1,16 @@
+# ATX Combined Coding Layer-Expert Residency Policies
+
+These JSON files are portable policies for the ATX llama.cpp fork. Use them with `--moe-residency-policy FILE` and record `--moe-residency-stats stats.json`.
+
+The percentages are fractions of all Qwen3.6 MoE layer-expert cells: `40 layers * 256 experts = 10,240` cells.
+
+| Policy | Selected cells | Intended use |
+|---|---:|---|
+| `combined_top_05pct_layer_experts.atx.json` | 512 | Tight VRAM budget / strongest hot-cell hypothesis |
+| `combined_top_10pct_layer_experts.atx.json` | 1,024 | Balanced first test |
+| `combined_top_15pct_layer_experts.atx.json` | 1,536 | Broader hot-cell coverage |
+| `combined_top_20pct_layer_experts.atx.json` | 2,048 | Approximately the top fifth of observed coding-important cells |
+
+Each file includes a `keep_layer_experts` array accepted by the ATX policy parser. Extra rank and score fields are ignored by the runtime but kept for inspection.
+
+These are observational saliency recommendations, not proven optimal throughput policies. Compare them against whole-layer policies and random controls on the same prompts.

policies/combined_layer_expert_policy_index.json

@@ -0,0 +1,49 @@
+{
+  "schema_version": "atx-moe-residency-policy-index-v1",
+  "description": "Portable ATX layer-expert residency policies generated from combined coding saliency traces.",
+  "usage": "./build-atx-metal/bin/llama-server -m /path/to/model.gguf --moe-residency-policy policies/combined_top_10pct_layer_experts.atx.json --moe-residency-stats stats.json",
+  "basis": "Mean of per-dataset normalized importance_score shares from SciCode MVP and Terminal-Bench Hard Agent traces.",
+  "total_layer_expert_pairs": 10240,
+  "policies": [
+    {
+      "percent": 5,
+      "path": "policies/combined_top_05pct_layer_experts.atx.json",
+      "selected_pairs": 512,
+      "cumulative_mean_importance_share": 0.660590412743,
+      "cumulative_scicode_importance_share": 0.742969757523,
+      "cumulative_terminal_bench_hard_agent_importance_share": 0.578211067963,
+      "sha256": "a7dde8b4fc3427913fecc4bb98d7fe20acdee13d7c38f5ce6456b796c3803d66",
+      "bytes": 259724
+    },
+    {
+      "percent": 10,
+      "path": "policies/combined_top_10pct_layer_experts.atx.json",
+      "selected_pairs": 1024,
+      "cumulative_mean_importance_share": 0.800313006063,
+      "cumulative_scicode_importance_share": 0.858042322518,
+      "cumulative_terminal_bench_hard_agent_importance_share": 0.742583689608,
+      "sha256": "c1561d95890d6da837100317fd4aff97de0b51c5a185af9880634e90eb2f8f0a",
+      "bytes": 516940
+    },
+    {
+      "percent": 15,
+      "path": "policies/combined_top_15pct_layer_experts.atx.json",
+      "selected_pairs": 1536,
+      "cumulative_mean_importance_share": 0.871043543523,
+      "cumulative_scicode_importance_share": 0.9104168024,
+      "cumulative_terminal_bench_hard_agent_importance_share": 0.831670284645,
+      "sha256": "96d990967c23ef65f207fec3e71fc6fbcd32dbfaeb46b8d699823722c255ad15",
+      "bytes": 774273
+    },
+    {
+      "percent": 20,
+      "path": "policies/combined_top_20pct_layer_experts.atx.json",
+      "selected_pairs": 2048,
+      "cumulative_mean_importance_share": 0.912287774427,
+      "cumulative_scicode_importance_share": 0.940337427439,
+      "cumulative_terminal_bench_hard_agent_importance_share": 0.884238121415,
+      "sha256": "33b64099ce42008d2e30046253210a481051c86eb945354b0f9f372df3d74086",
+      "bytes": 1030436
+    }
+  ]
+}