diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 1bf828a5b4..dd6da6de22 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -100,7 +100,7 @@ jobs:
     # its own. It runs on push only; on merge_group it is skipped at t=0, which the verdict blocks
     # below read as `skipped:::skipped` = "x86-main + determinism ran".
     if: github.event_name == 'push' && github.ref == 'refs/heads/main'
-    runs-on: [self-hosted, Linux, X64, clean-room]
+    runs-on: [self-hosted, Linux, clean-room]
     timeout-minutes: 10
     permissions:
       contents: read
@@ -314,7 +314,7 @@ jobs:
   x86-main-advisories:
     needs: [mg-reuse]
     if: ${{ !cancelled() && needs.mg-reuse.outputs.x86 == '1' }}
-    runs-on: [self-hosted, Linux, X64, clean-room]
+    runs-on: [self-hosted, Linux, clean-room]
     timeout-minutes: 30
     permissions:
       contents: read
@@ -437,11 +437,11 @@ jobs:
   # failure, never a pass.
   # ==========================================================================
   # The workspace tests, one shard per job (nextest hash partition N/12 plus the
-  # explicit runner's round-robin N/12), in parallel on the X64 clean-room pool.
+  # explicit runner's round-robin N/12), in parallel on the clean-room pool (arch-neutral, PMAT-3138).
   # The section body is ci/sections.yml `workspace-test-shard`, unchanged.
   workspace-test-shard:
-    # same lanes as x86-main: clean-room for pull_request/push, mq for a merge_group entry
-    runs-on: ${{ github.event_name == 'merge_group' && fromJSON('["self-hosted","Linux","X64","mq"]') || fromJSON('["self-hosted","Linux","X64","clean-room"]') }}
+    # clean-room for pull_request/push, mq for a merge_group entry; no X64 pin (operator 2026-10-21 14:18Z)
+    runs-on: ${{ github.event_name == 'merge_group' && fromJSON('["self-hosted","Linux","mq"]') || fromJSON('["self-hosted","Linux","clean-room"]') }}
     timeout-minutes: 180
     strategy:
       fail-fast: false
diff --git a/crates/aprender-serve/src/kernel_registry_parity.rs b/crates/aprender-serve/src/kernel_registry_parity.rs
index 0f47c86e85..bdb6e3a798 100644
--- a/crates/aprender-serve/src/kernel_registry_parity.rs
+++ b/crates/aprender-serve/src/kernel_registry_parity.rs
@@ -743,6 +743,7 @@ fn committed() -> Vec<(serde_json::Value, serde_json::Value)> {
 #[test]
 fn committed_parity_receipts_hold_on_this_host() {
     let r = registry().expect("registry");
+    let mut no_arch_receipt = Vec::new();
     for (entry, rc) in committed() {
         let id = rc["kernel_id"].as_str().expect("kernel_id");
         let path = entry["receipt"].as_str().expect("receipt path");
@@ -789,8 +790,22 @@ fn committed_parity_receipts_hold_on_this_host() {
             seed: wl["seed"].as_u64().expect("seed"),
             trials: dim("trials"),
         };
-        let bound = rc["tolerance_rel"].as_f64().expect("tolerance_rel");
         let now = measure(kernel(id), w, false);
+        // A receipt's `tolerance_rel` is the bound measured on its own `host_arch`; libm and FMA
+        // contraction move the last ulps between architectures. Any other arch must carry its own
+        // measured bound under `per_arch.<arch>.tolerance_rel` -- never a borrowed one, never a skip.
+        let bound = if rc["host_arch"] == host_arch() {
+            rc["tolerance_rel"].as_f64().expect("tolerance_rel")
+        } else if let Some(b) = rc["per_arch"][host_arch()]["tolerance_rel"].as_f64() {
+            b
+        } else {
+            no_arch_receipt.push(format!(
+                "{path}: no {} receipt (host measured max_rel_err {})",
+                host_arch(),
+                now.max_rel_err
+            ));
+            continue;
+        };
         assert_eq!(
             rc["margin"]["verdict"], "pass",
             "{path}: KTEST-02 margin verdict (max ≤ 1 under EM-DOT)"
@@ -842,6 +857,11 @@ fn committed_parity_receipts_hold_on_this_host() {
             p => panic!("{id}: no precision check for {p}"),
         }
     }
+    assert!(
+        no_arch_receipt.is_empty(),
+        "committed receipts lack a bound for this arch:\n{}",
+        no_arch_receipt.join("\n")
+    );
 }
 
 /// `KREG_INPUT_SETS_OUT`: the input-set hashes recomputed from the tree at `KREG_GIT_SHA`, one
diff --git a/evidence/kreg/parity/cpu.matvec.apr_q4.json b/evidence/kreg/parity/cpu.matvec.apr_q4.json
index e7929dfd71..da9c1fc945 100644
--- a/evidence/kreg/parity/cpu.matvec.apr_q4.json
+++ b/evidence/kreg/parity/cpu.matvec.apr_q4.json
@@ -35,6 +35,15 @@
   },
   "oracle": "apr_format_decode_f64",
   "oracle_independent": true,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 5.568608963671898e-07,
+      "source": "CI job 119289824035 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 2e-6,
+      "ts": "2026-10-21T14:51:19Z"
+    }
+  },
   "quantized_activation_oracle": null,
   "registry_precision": "f32",
   "schema": "kernel-parity-receipt/v1",
diff --git a/evidence/kreg/parity/cpu.matvec.apr_q8.json b/evidence/kreg/parity/cpu.matvec.apr_q8.json
index e95fb4c5c9..9f6f4ac2cf 100644
--- a/evidence/kreg/parity/cpu.matvec.apr_q8.json
+++ b/evidence/kreg/parity/cpu.matvec.apr_q8.json
@@ -35,6 +35,15 @@
   },
   "oracle": "apr_format_decode_f64",
   "oracle_independent": true,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 5.982009671170112e-07,
+      "source": "CI job 119289824035 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 2e-6,
+      "ts": "2026-10-21T14:51:19Z"
+    }
+  },
   "quantized_activation_oracle": null,
   "registry_precision": "f32",
   "schema": "kernel-parity-receipt/v1",
diff --git a/evidence/kreg/parity/cpu.matvec.bf16.json b/evidence/kreg/parity/cpu.matvec.bf16.json
index e6aa4240c0..57309f51c6 100644
--- a/evidence/kreg/parity/cpu.matvec.bf16.json
+++ b/evidence/kreg/parity/cpu.matvec.bf16.json
@@ -33,6 +33,15 @@
   },
   "oracle": "le_decode_half_f64",
   "oracle_independent": true,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 4.068146011742945e-07,
+      "source": "CI job 119289824035 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 9e-7,
+      "ts": "2026-10-21T14:51:19Z"
+    }
+  },
   "quantized_activation_oracle": null,
   "registry_precision": "f32",
   "schema": "kernel-parity-receipt/v1",
diff --git a/evidence/kreg/parity/cpu.matvec.f16.json b/evidence/kreg/parity/cpu.matvec.f16.json
index 17b3754d5a..af25b8ed9d 100644
--- a/evidence/kreg/parity/cpu.matvec.f16.json
+++ b/evidence/kreg/parity/cpu.matvec.f16.json
@@ -33,6 +33,15 @@
   },
   "oracle": "le_decode_half_f64",
   "oracle_independent": true,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 3.920968817889049e-07,
+      "source": "CI job 119289824035 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 8e-7,
+      "ts": "2026-10-21T14:51:19Z"
+    }
+  },
   "quantized_activation_oracle": null,
   "registry_precision": "f32",
   "schema": "kernel-parity-receipt/v1",
diff --git a/evidence/kreg/parity/cpu.matvec.f32.json b/evidence/kreg/parity/cpu.matvec.f32.json
index 9926666b94..a41e599001 100644
--- a/evidence/kreg/parity/cpu.matvec.f32.json
+++ b/evidence/kreg/parity/cpu.matvec.f32.json
@@ -33,6 +33,15 @@
   },
   "oracle": "le_decode_half_f64",
   "oracle_independent": true,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 7.356022484897627e-07,
+      "source": "CI job 119289824035 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 2e-6,
+      "ts": "2026-10-21T14:51:19Z"
+    }
+  },
   "quantized_activation_oracle": null,
   "registry_precision": "f32",
   "schema": "kernel-parity-receipt/v1",
diff --git a/evidence/kreg/parity/cpu.matvec.q4_0.json b/evidence/kreg/parity/cpu.matvec.q4_0.json
index 38ab583a76..d1a94f3c46 100644
--- a/evidence/kreg/parity/cpu.matvec.q4_0.json
+++ b/evidence/kreg/parity/cpu.matvec.q4_0.json
@@ -35,6 +35,15 @@
   },
   "oracle": "in_tree_dequant_f64",
   "oracle_independent": false,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 0.003859434297246209,
+      "source": "CI job 119289824035 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 0.008,
+      "ts": "2026-10-21T14:51:19Z"
+    }
+  },
   "quantized_activation_oracle": {
     "max_rel_err": 1.69613194525447e-7
   },
diff --git a/evidence/kreg/parity/cpu.matvec.q4_k.json b/evidence/kreg/parity/cpu.matvec.q4_k.json
index 501b8c768a..0e1ccc3ba7 100644
--- a/evidence/kreg/parity/cpu.matvec.q4_k.json
+++ b/evidence/kreg/parity/cpu.matvec.q4_k.json
@@ -36,6 +36,15 @@
   },
   "oracle": "in_tree_dequant_f64",
   "oracle_independent": false,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 0.005569836531516333,
+      "source": "CI job 119289824035 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 0.02,
+      "ts": "2026-10-21T14:51:19Z"
+    }
+  },
   "quantized_activation_oracle": {
     "max_rel_err": 1.3720446505396093e-7
   },
diff --git a/evidence/kreg/parity/cpu.matvec.q5_k.json b/evidence/kreg/parity/cpu.matvec.q5_k.json
index 01913cf231..d65c8a5ca9 100644
--- a/evidence/kreg/parity/cpu.matvec.q5_k.json
+++ b/evidence/kreg/parity/cpu.matvec.q5_k.json
@@ -36,6 +36,15 @@
   },
   "oracle": "in_tree_dequant_f64",
   "oracle_independent": false,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 1.3110594897372703e-06,
+      "source": "CI job 119289824035 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 3e-6,
+      "ts": "2026-10-21T14:51:19Z"
+    }
+  },
   "quantized_activation_oracle": null,
   "registry_precision": "f32",
   "schema": "kernel-parity-receipt/v1",
diff --git a/evidence/kreg/parity/cpu.matvec.q6_k.json b/evidence/kreg/parity/cpu.matvec.q6_k.json
index 220f53b4a7..ce60279cc6 100644
--- a/evidence/kreg/parity/cpu.matvec.q6_k.json
+++ b/evidence/kreg/parity/cpu.matvec.q6_k.json
@@ -36,6 +36,15 @@
   },
   "oracle": "in_tree_dequant_f64",
   "oracle_independent": false,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 1.6505971099429934e-6,
+      "source": "CI job 119280981403 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 4e-6,
+      "ts": "2026-10-21T14:37:43Z"
+    }
+  },
   "quantized_activation_oracle": null,
   "registry_precision": "f32",
   "schema": "kernel-parity-receipt/v1",
diff --git a/evidence/kreg/parity/cpu.matvec.q8_0.json b/evidence/kreg/parity/cpu.matvec.q8_0.json
index b69febb636..0f6306509e 100644
--- a/evidence/kreg/parity/cpu.matvec.q8_0.json
+++ b/evidence/kreg/parity/cpu.matvec.q8_0.json
@@ -36,6 +36,15 @@
   },
   "oracle": "in_tree_dequant_f64",
   "oracle_independent": false,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 0.003738839019401381,
+      "source": "CI job 119289824035 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 0.008,
+      "ts": "2026-10-21T14:51:19Z"
+    }
+  },
   "quantized_activation_oracle": {
     "max_rel_err": 2.0240994602880252e-7
   },
diff --git a/scripts/ci_mg_reuse.sh b/scripts/ci_mg_reuse.sh
index 86573f5aff..b87174df96 100755
--- a/scripts/ci_mg_reuse.sh
+++ b/scripts/ci_mg_reuse.sh
@@ -528,11 +528,12 @@ STUB
   # Each row is compared as the EXACT line (trailing comment stripped), so a third state cannot slip in.
   poolrows() { # poolrows WF -> one line per violated pool row
     local j l want
-    local plain="runs-on: [self-hosted, Linux, X64, clean-room]"
+    local plain="runs-on: [self-hosted, Linux, clean-room]" # arch-neutral job: cargo-deny, no X64 pin
     local lane="runs-on: \${{ github.event_name == 'merge_group' && fromJSON('[\"self-hosted\",\"Linux\",\"X64\",\"mq\"]') || fromJSON('[\"self-hosted\",\"Linux\",\"X64\",\"clean-room\"]') }}"
+    local nlane="runs-on: \${{ github.event_name == 'merge_group' && fromJSON('[\"self-hosted\",\"Linux\",\"mq\"]') || fromJSON('[\"self-hosted\",\"Linux\",\"clean-room\"]') }}"
     for j in x86-main x86-main-advisories workspace-test-shard; do
       l=$(awk -v j="  $j:" '$0 == j {f = 1; next} f && /^  [A-Za-z0-9_-]+:$/ {exit} f' "$1" 2> /dev/null | grep -m1 '^    runs-on:' | sed -E 's/^ +//; s/ +#.*$//')
-      want=$plain; case $j in x86-main | workspace-test-shard) want=$lane ;; esac
+      want=$plain; case $j in x86-main) want=$lane ;; workspace-test-shard) want=$nlane ;; esac
       if [ "$l" != "$want" ]; then
         echo "$j does not land on the clean-room pool (mq on merge_group for x86-main/workspace-test-shard) (runs-on: $l)"; fi
     done
@@ -552,7 +553,7 @@ STUB
   poolmut pool-advisories-event-conditional "does not land on the clean-room pool" '/^  x86-main-advisories:$/,/^    steps:/{s/^    runs-on: .*/    runs-on: ${{ github.event_name == '"'"'merge_group'"'"' \&\& fromJSON('"'"'["clean-room"]'"'"') || fromJSON('"'"'["clean-room"]'"'"') }}/}'
   poolmut mq-lane-lost-on-x86-main "does not land on the clean-room pool" '/^  x86-main:$/,/^    steps:/{s/"mq"/"clean-room"/}'
   poolmut mq-lane-dropped-on-shard "does not land on the clean-room pool" '/^  workspace-test-shard:$/,/^    steps:/{s/^    runs-on: .*/    runs-on: [self-hosted, Linux, X64, clean-room]/}'
-  poolmut pr-path-takes-mq "does not land on the clean-room pool" '/^  workspace-test-shard:$/,/^    steps:/{s/\|\| fromJSON\(.\["self-hosted","Linux","X64","clean-room"\].\)/|| fromJSON('"'"'["self-hosted","Linux","X64","mq"]'"'"')/}'
+  poolmut pr-path-takes-mq "does not land on the clean-room pool" '/^  workspace-test-shard:$/,/^    steps:/{s/\|\| fromJSON\(.\["self-hosted","Linux","clean-room"\].\)/|| fromJSON('"'"'["self-hosted","Linux","mq"]'"'"')/}'
   poolmut mq-lane-third-label "does not land on the clean-room pool" '/^  x86-main:$/,/^    steps:/{s/"X64","mq"/"X64","mq","pr-only"/}'
 
   # S1 (20X): mg-reuse runs on push only, so on merge_group it is SKIPPED and every dependent