Use qualified default specialization for Watch Core ML

README.md

@@ -141,7 +141,7 @@ signatures and operation counts. Hugging Face provides file checksums for each r
 
 ## Experimental Core ML bundle for Apple Watch
 
-`coreml-watch/bundle.tar` contains compiled watchOS 27 Core ML graphs, runtime configuration, frontend/decoder sidecars, license files and `loading.json` with measured per-file loading baselines. Extract the uncompressed POSIX ustar archive into one model directory. The existing Core AI bundles remain available separately. These Watch graphs request CPU and Neural Engine execution using public Core ML APIs.
+`coreml-watch/bundle.tar` contains compiled watchOS 27 Core ML graphs, runtime configuration, frontend/decoder sidecars, license files and `loading.json` with measured per-file loading baselines. Extract the uncompressed POSIX ustar archive into one model directory. The existing Core AI bundles remain available separately. These Watch graphs request CPU and Neural Engine execution using public Core ML APIs, with the default specialization strategy.
 
 This English TDT configuration uses up-to-15-second chunks, a CPU decoder and six resident encoder stages. Native word timings are available. A 90-second numerical qualification matched source token IDs in three runs.
 
@@ -149,8 +149,8 @@ Measured on Apple Watch Ultra 4, watchOS 27.0.1, Release runtime, resident graph
 
 | Measurement | Apple Watch Ultra 4 |
 | --- | ---: |
-| Processing throughput, 20-second input | 25.3× real time |
-| Observed first preparation | 189.7 s |
+| Processing throughput, 20-second input | 27.2× real time |
+| Observed first preparation | 70.2 s |
 | Download size | 668.6 MB |
 
-Preparation is an observed first load, not a controlled cold-cache benchmark. Processing includes the frontend, model work, cache updates and host decoding where applicable; it excludes preparation, warmup and result writing. A separate hardware trace confirmed ANE execution of encoder stage 0; this does not establish whole-pipeline placement or ALU utilization. First preparation on other devices may differ.
+Preparation is an observed first load of the default strategy with existing caches preserved, not a controlled cold-cache benchmark. Processing includes the frontend, model work, cache updates and host decoding where applicable; it excludes preparation, warmup and result writing. A separate hardware trace confirmed ANE execution of encoder stage 0; this does not establish whole-pipeline placement or ALU utilization. First preparation on other devices may differ.

coreml-watch/bundle.tar

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:144a3edee53a3d0adcc5fe300e4f367fa177cec921bb93229e80d87de53e3855
-size 668620800
+oid sha256:26dbd456056ca3d10e07e087857318733dd0dcfd5fbca06edb57bf38f5b483f1
+size 668641280

coreml-watch/loading.json

@@ -1,14 +1,15 @@
 {
+  "specialization_strategy": "default",
   "device": "Apple Watch Ultra 4",
-  "os": "watchOS 27.0.1",
-  "measurement_scope": "Observed first model preparation; not a controlled cache-purge measurement. Includes any on-device specialization.",
+  "os": "Version 27.0.1 (Build 24R365)",
   "cold_load_seconds": {
-    "subsampling": 4.074,
-    "encoder_0": 31.299,
-    "encoder_1": 30.97,
-    "encoder_2": 30.778,
-    "encoder_3": 30.999,
-    "encoder_4": 30.549,
-    "encoder_5": 30.814
-  }
+    "subsampling": 0.6473711666767485,
+    "encoder_0": 11.457231166670681,
+    "encoder_1": 11.467421708337497,
+    "encoder_2": 11.44494304167165,
+    "encoder_3": 11.574859916669084,
+    "encoder_4": 12.016754333337303,
+    "encoder_5": 11.474410083334078
+  },
+  "measurement": "Observed first strategy-specific load; existing caches preserved. Actual per-file preparation calibrates the ETA."
 }