Upload docs/RESULTS_l3long_2026-09-09.md with huggingface_hub

docs/RESULTS_l3long_2026-09-09.md

@@ -20,7 +20,10 @@ window. Step count was the remaining untested lever.
     rate            11.56 s/step measured
 
 IMPORTANT: this is a 16k run, NOT a 32k run. The "32768" in the data path is the
-document-length bucket, not the context window. No 32k extension has been run.
+document-length bucket, not the context window. A SEPARATE 32k extension
+(ext32k_b4, parent l2, YaRN scale 2 x orig 16384) was run and evaluated earlier
+on 2026-09-09 -- see docs/EXT32K_RESULTS_2026-09-09.md. The local ext_32k.sh
+(parent l2, b3 data) is a DIFFERENT, still-unrun variant.
 
 ## NIAH single-needle accuracy (%)
 
@@ -30,6 +33,20 @@ document-length bucket, not the context window. No 32k extension has been run.
 | 8192  | 53.0 | 48.0 | 62.0 | 64.0 | 55.0 | 54.0 |
 | 16360 | 33.0 | 30.0 | 35.0 | 35.0 | 36.0 | 43.0 |
 
+## Cross-comparison with the 32k extension at 16,360
+
+| arm | what it changed | 4,096 | 8,192 | 16,360 | 32,744 |
+|---|---|---|---|---|---|
+| l2 (baseline) | --                    | 80.0 | 62.0 | 35.0 | cannot run |
+| l3long@2000   | more steps at 16k     | 68.0 | 54.0 | 43.0 | cannot run |
+| ext32k_b4     | widened window to 32k | 75.0 | 60.0 | 45.0 | 27.0 |
+
+Widening the window and adding step count buy a SIMILAR gain at 16,360 (+10.0
+and +8.0 over l2 respectively, neither resolvable at N=100), but only the
+widening unlocks 32,744 at all, and it costs less short-context accuracy
+(4,096: 75.0 vs 68.0). If only one lever survives, the window widening looks
+like the better buy -- but note ext32k is decayed and l3long@2000 is not.
+
 ## Reading
 
 Step count IS a live lever at 16k, unlike document length. At 16360 l3long went