Phonon-2: eval results at the card's numbers; reference script comments

.eval_results/open_asr_leaderboard.yaml

@@ -1,90 +1,79 @@
 - dataset:
     id: hf-audio/open-asr-leaderboard
     task_id: mean_wer
-  value: 5.15
-  date: '2026-09-26'
+  value: 5.21
+  date: '2026-09-29'
   source:
     url: https://huggingface.co/FermionResearch/Phonon-2
-    name: FermionResearch (own run of the leaderboard's nemo_asr runner at 7ce09ca6, one A100-80; pending the maintainers' run)
+    name: FermionResearch (own run of the leaderboard's scoring code on the full test sets, 2026-09-29; the leaderboard's own runner row is in open_asr_leaderboard PR #232)
     user: FermionResearch
 
 - dataset:
     id: hf-audio/open-asr-leaderboard
-    task_id: ami_wer
-  value: 9.28
-  date: '2026-09-26'
+    task_id: librispeech_clean_wer
+  value: 1.72
+  date: '2026-09-29'
   source:
     url: https://huggingface.co/FermionResearch/Phonon-2
-    name: FermionResearch (own run of the leaderboard's nemo_asr runner at 7ce09ca6, one A100-80; pending the maintainers' run)
+    name: FermionResearch (own run of the leaderboard's scoring code on the full test sets, 2026-09-29; the leaderboard's own runner row is in open_asr_leaderboard PR #232)
     user: FermionResearch
 
 - dataset:
     id: hf-audio/open-asr-leaderboard
-    task_id: earnings22_wer
-  value: 6.71
-  date: '2026-09-26'
+    task_id: librispeech_other_wer
+  value: 3.92
+  date: '2026-09-29'
   source:
     url: https://huggingface.co/FermionResearch/Phonon-2
-    name: FermionResearch (own run of the leaderboard's nemo_asr runner at 7ce09ca6, one A100-80; pending the maintainers' run)
+    name: FermionResearch (own run of the leaderboard's scoring code on the full test sets, 2026-09-29; the leaderboard's own runner row is in open_asr_leaderboard PR #232)
     user: FermionResearch
 
 - dataset:
     id: hf-audio/open-asr-leaderboard
-    task_id: gigaspeech_wer
-  value: 8.33
-  date: '2026-09-26'
+    task_id: ami_wer
+  value: 9.37
+  date: '2026-09-29'
   source:
     url: https://huggingface.co/FermionResearch/Phonon-2
-    name: FermionResearch (own run of the leaderboard's nemo_asr runner at 7ce09ca6, one A100-80; pending the maintainers' run)
+    name: FermionResearch (own run of the leaderboard's scoring code on the full test sets, 2026-09-29; the leaderboard's own runner row is in open_asr_leaderboard PR #232)
     user: FermionResearch
 
 - dataset:
     id: hf-audio/open-asr-leaderboard
-    task_id: librispeech_clean_wer
-  value: 1.72
-  date: '2026-09-26'
+    task_id: earnings22_wer
+  value: 6.96
+  date: '2026-09-29'
   source:
     url: https://huggingface.co/FermionResearch/Phonon-2
-    name: FermionResearch (own run of the leaderboard's nemo_asr runner at 7ce09ca6, one A100-80; pending the maintainers' run)
+    name: FermionResearch (own run of the leaderboard's scoring code on the full test sets, 2026-09-29; the leaderboard's own runner row is in open_asr_leaderboard PR #232)
     user: FermionResearch
 
 - dataset:
     id: hf-audio/open-asr-leaderboard
-    task_id: librispeech_other_wer
-  value: 3.9
-  date: '2026-09-26'
+    task_id: gigaspeech_wer
+  value: 8.35
+  date: '2026-09-29'
   source:
     url: https://huggingface.co/FermionResearch/Phonon-2
-    name: FermionResearch (own run of the leaderboard's nemo_asr runner at 7ce09ca6, one A100-80; pending the maintainers' run)
+    name: FermionResearch (own run of the leaderboard's scoring code on the full test sets, 2026-09-29; the leaderboard's own runner row is in open_asr_leaderboard PR #232)
     user: FermionResearch
 
 - dataset:
     id: hf-audio/open-asr-leaderboard
     task_id: spgispeech_wer
-  value: 3.68
-  date: '2026-09-26'
+  value: 3.70
+  date: '2026-09-29'
   source:
     url: https://huggingface.co/FermionResearch/Phonon-2
-    name: FermionResearch (own run of the leaderboard's nemo_asr runner at 7ce09ca6, one A100-80; pending the maintainers' run)
+    name: FermionResearch (own run of the leaderboard's scoring code on the full test sets, 2026-09-29; the leaderboard's own runner row is in open_asr_leaderboard PR #232)
     user: FermionResearch
 
 - dataset:
     id: hf-audio/open-asr-leaderboard
     task_id: voxpopuli_wer
-  value: 2.44
-  date: '2026-09-26'
+  value: 2.46
+  date: '2026-09-29'
   source:
     url: https://huggingface.co/FermionResearch/Phonon-2
-    name: FermionResearch (own run of the leaderboard's nemo_asr runner at 7ce09ca6, one A100-80; pending the maintainers' run)
+    name: FermionResearch (own run of the leaderboard's scoring code on the full test sets, 2026-09-29; the leaderboard's own runner row is in open_asr_leaderboard PR #232)
     user: FermionResearch
-
-- dataset:
-    id: hf-audio/open-asr-leaderboard
-    task_id: rtfx
-  value: 3500.59
-  date: '2026-09-26'
-  source:
-    url: https://huggingface.co/FermionResearch/Phonon-2
-    name: FermionResearch (own run of the leaderboard's nemo_asr runner at 7ce09ca6, one A100-80; pending the maintainers' run)
-    user: FermionResearch
-

SHA256SUMS.txt

@@ -2,7 +2,7 @@
 422379e411f14dde97174554deead106ec1e4816e14dd78a1bf99517c6dcc663  config.json
 690c6bc43bcbcae8df61cff0b2cace7299d023718b6f2aa9d67157407eadaa07  packed_manifest.json
 00624a5043e7ce74029317b024132ca5286116d6fbf3191d226374bc8273789f  NOTICE
-6833410ae75725ec2a9392085c2138a6d4ffb87be62ff5fe86faff45413f8586  reference_transformers.py
+cdb9019dddef44da211641022564830733b18841bdce7aa66370ff6d47526a8d  reference_transformers.py
 5cf172e8cf6b313e951c4ef588ca913c7285b0aa3b92de413c3d7ec2b939c0e4  fermion_container.py
 0dad36aa73a9b80fd5165dc33bd226f874a4ff727b06c7bd04a289c69da58452  README.md
 9ba9550ad48438d0836ddab3da480b3b69ffa0aac7b7878b5a0039e7ab429411  LICENSE-WEIGHTS-CC-BY-4.0.txt

reference_transformers.py

@@ -1,11 +1,11 @@
-"""Dense `transformers` reference for the Phonon-2 container (item 3 of the ship lane).
+"""Dense `transformers` reference for the Phonon-2 container.
 
 read container -> expand every record to fp32 (five-value: exact {0,+-lo,+-hi}; intN: q*scale;
 fp16: as stored) -> stock ParakeetForTDT(config).load_state_dict(strict=True) -> greedy TDT
-generate through ParakeetProcessor, exactly the call the official-meter read boxes make.
+generate through ParakeetProcessor, the same call the leaderboard scoring uses.
 
 Usage: python reference_transformers.py model.fermion --audio recording.wav
-       python reference_transformers.py model.fermion --utts rows.json --out receipt.json   (the gate set)
+       python reference_transformers.py model.fermion --utts rows.json --out receipt.json
 Outputs per utterance: transcript (+ token ids) and, for --dump-encoder N utts, the encoder
 output in fp32 so the MLX graph can be compared against the torch graph.  Pure reference:
 no speed claim is made or measured here.