Phonon-1

.gitattributes

@@ -0,0 +1,7 @@
+# LFS rules — commit this file FIRST, before any weight file (Neutrino 0.6B
+# pattern; a file committed before its rule exists stays non-LFS forever).
+# Never add a glob that a sibling small file would match: the Hub resolves LFS
+# by "does any filter=lfs pattern match" and does NOT honour a later unset.
+*.tar.zst  filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+*.bin *.pt *.pth *.ckpt *.npz *.onnx *.msgpack *.tar *.tar.gz *.zip filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

NOTICE

@@ -0,0 +1,60 @@
+Phonon
+Copyright 2026 Fermion Research
+
+This product includes software and models developed by Fermion Research.
+
+================================================================================
+BASE MODEL
+================================================================================
+
+This model is derived from Qwen/Qwen3-ASR-0.6B, licensed under the
+Apache License, Version 2.0.
+
+    https://huggingface.co/Qwen/Qwen3-ASR-0.6B
+    https://www.apache.org/licenses/LICENSE-2.0
+
+The audio encoder architecture, tokenizer (vocab.json, merges.txt,
+tokenizer_config.json), chat template, generation config and preprocessor
+config derive from that model and remain under Apache-2.0.
+
+
+================================================================================
+BUNDLED SAMPLE AUDIO
+================================================================================
+
+samples/1089-134691-0000.flac
+samples/1089-134691.trans.txt
+
+    From the LibriSpeech ASR corpus (Panayotov, Chen, Povey, Khudanpur;
+    ICASSP 2015; https://www.openslr.org/12/), licensed CC-BY 4.0.
+    Used as the runtime warm-up clip and as the quickstart example.
+
+================================================================================
+RUNTIME DEPENDENCIES
+================================================================================
+
+MLX ......................... MIT License ......... https://github.com/ml-explore/mlx
+mlx-audio ................... MIT License ......... https://github.com/Blaizzy/mlx-audio
+NumPy ....................... BSD-3-Clause ........ https://numpy.org
+SciPy ....................... BSD-3-Clause ........ https://scipy.org
+sounddevice ................. MIT License ......... https://github.com/spatialaudio/python-sounddevice
+soundfile ................... BSD-3-Clause ........ https://github.com/bastibe/python-soundfile
+    soundfile links libsndfile (LGPL-2.1), dynamically and unmodified, as
+    distributed in the standard soundfile wheel.
+
+================================================================================
+OPTIONAL --punctuate DEPENDENCIES
+================================================================================
+
+Not installed by default. Required only for the display-only punctuation and
+truecasing pass.
+
+punctuators ................. Apache-2.0 .......... https://github.com/1-800-BAD-CODE/punctuators
+onnxruntime ................. MIT License ......... https://onnxruntime.ai
+
+Punctuation / truecasing / sentence-boundary models, all Apache-2.0, downloaded
+on first use and run locally on CPU:
+
+    1-800-BAD-CODE/punctuation_fullstop_truecase_english
+    1-800-BAD-CODE/punct_cap_seg_47_language
+    1-800-BAD-CODE/xlm-roberta_punctuation_fullstop_truecase

README.md

@@ -0,0 +1,69 @@
+---
+license: apache-2.0
+base_model: Qwen/Qwen3-ASR-0.6B
+base_model_relation: quantized
+language:
+- en
+library_name: mlx
+pipeline_tag: automatic-speech-recognition
+tags:
+- mlx
+- apple-silicon
+- speech-to-text
+- asr
+- stt
+- low-bit
+- ternary
+- quantization-aware-training
+- on-device
+- streaming
+metrics:
+- wer
+---
+
+# Phonon-1
+
+Phonon-1 is an open speech recognition model for English. It downloads in
+415 MB, runs on a laptop or a datacenter GPU, and transcribes an hour of audio
+in about two and a half minutes. It was trained at 2.4 bits per weight from
+the start, and it is the second model in the lab's low-bit lane after
+Neutrino-1.
+
+## Benchmarks
+
+| Dataset | Phonon-1 (415 MB) | Phonon-1 Micro (285 MB) | Parakeet-0.6B 4-bit (637 MB) | Moonshine base (248 MB) | Whisper large-v3-turbo (1,619 MB) | Whisper small (967 MB) | wav2vec2-large (1,262 MB) | Qwen3-ASR teacher (1,569 MB) |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|
+| LibriSpeech test-clean | 2.640 | 3.002 | 2.186 | 3.417 | 2.10 | 3.4† | 2.8† | 2.235 |
+| LibriSpeech test-other | 5.699 | 6.511 | 3.937 | 8.262 | 4.07 | 7.6† | 6.3† | 4.618 |
+| TED-LIUM | 3.421 | 3.878 | 2.829 | 5.272 | — | — | — | 2.889 |
+| SPGISpeech | 4.163 | 4.858 | 4.104 | 5.731 | 2.79† | — | 13.31† | 3.074 |
+| VoxPopuli | 8.394 | 9.177 | 6.345 | 10.470 | 11.22† | — | — | 7.151 |
+| GigaSpeech | 11.396 | 11.882 | 9.614 | 12.114 | 8.52† | — | — | 9.321 |
+| Earnings-22 | 12.571 | 14.771 | 11.190 | 17.872 | 11.07† | — | 36.28† | 11.188 |
+| AMI | 13.084 | 14.094 | 12.723 | 17.790 | 15.16† | — | — | 12.560 |
+| Macro (eight benchmarks) | 7.67 | 8.52 | 6.62 | 10.1 | — | — | — | 6.63 |
+
+Word error rate, lower is better. Unmarked cells: measured by us — full test sets, Whisper English text normalizer, greedy decoding. † = published figure (model card, paper, or the Open ASR Leaderboard). Dash = no comparable measurement.
+
+Median 23.9× realtime across nine corpora on a base M5 MacBook Air.
+
+## Run it
+
+```bash
+pip install fermion-research
+fermion transcribe recording.wav
+```
+
+```bash
+fermion serve
+curl -s http://127.0.0.1:8000/v1/audio/transcriptions \
+  -F "file=@recording.wav" \
+  -F "model=FermionResearch/Phonon-1"
+```
+
+The same weights run on a Mac (via MLX) or an NVIDIA GPU; the CUDA runtime and
+Docker image are in the [GitHub repo](https://github.com/fermionresearch/phonon).
+
+## License
+
+**Apache License 2.0** for the weights and the [command line](https://pypi.org/project/fermion-research/). Base model: [`Qwen/Qwen3-ASR-0.6B`](https://huggingface.co/Qwen/Qwen3-ASR-0.6B), Apache-2.0.

config.json

@@ -0,0 +1,74 @@
+{
+  "config_schema": "fermion.phonon/1",
+  "model_id": "FermionResearch/Phonon-1",
+  "name": "Phonon-1",
+  "summary": "The default Phonon model. Statistically tied with Phonon-1 Big on accuracy at 71 % of the download.",
+  "license": "apache-2.0",
+  "library_name": "mlx",
+  "pipeline_tag": "automatic-speech-recognition",
+  "language": [
+    "en"
+  ],
+  "base_model": "Qwen/Qwen3-ASR-0.6B",
+  "base_model_relation": "quantized",
+  "audio": {
+    "sample_rate": 16000,
+    "channels": 1,
+    "designed_utterance_seconds": [
+      0.5,
+      30.0
+    ],
+    "punctuation": "native"
+  },
+  "runtime": {
+    "platform": "macOS on Apple silicon (arm64)",
+    "framework": "mlx"
+  },
+  "artifact": {
+    "profile": "audio6",
+    "backend": "audio6",
+    "filename": "phonon-audio6.bps.tar.zst",
+    "download_bytes": 415077202,
+    "sha256": "214c3b45aa57257013811a53f99a905848466ad4ab21f2b2f8368f7ac79427b2",
+    "unpacked_bytes": 454527730
+  },
+  "profiles": {
+    "default": "audio6",
+    "listed": [
+      "audio6"
+    ],
+    "entries": [
+      {
+        "key": "audio6",
+        "name": "Audio6",
+        "backend": "audio6",
+        "filename": "phonon-audio6.bps.tar.zst",
+        "download_bytes": 415077202,
+        "sha256": "214c3b45aa57257013811a53f99a905848466ad4ab21f2b2f8368f7ac79427b2",
+        "unpacked_bytes": 454527730,
+        "listed": true
+      }
+    ]
+  },
+  "family": [
+    {
+      "repo_id": "FermionResearch/Phonon-1-Big",
+      "name": "Phonon-1 Big",
+      "summary": "The largest Phonon build, and the quickest response of the three. Not the default: it costs 166 MB more than Phonon-1 for no measured accuracy gain."
+    },
+    {
+      "repo_id": "FermionResearch/Phonon-1",
+      "name": "Phonon-1",
+      "summary": "The default Phonon model. Statistically tied with Phonon-1 Big on accuracy at 71 % of the download."
+    },
+    {
+      "repo_id": "FermionResearch/Phonon-1-Micro",
+      "name": "Phonon-1 Micro",
+      "summary": "The smallest Phonon model, for the smallest install. It trades some accuracy for size."
+    }
+  ],
+  "notes": [
+    "Accuracy and latency figures are published in README.md, not here, so that a card revision can never leave a stale number in machine-readable metadata.",
+    "The byte size and SHA-256 above describe this release and do not change for it. Check the archive against them before unpacking."
+  ]
+}

package_release_bps.py

@@ -0,0 +1,281 @@
+#!/usr/bin/env python3
+"""Byte-plane-split release packer/unpacker.
+
+A BF16 array interleaves a highly predictable exponent byte with a near-random
+mantissa byte.  Compressed interleaved, zstd models neither population well.
+Split into per-byte-position planes, the exponent plane compresses hard.  U8
+code planes are left untouched.
+
+This is a *transport* transform only:
+
+* the transform is applied per tensor, using the dtype in the safetensors
+  header, so it never has to guess;
+* bytes not covered by any tensor (the header, alignment padding) are copied
+  verbatim;
+* ``unpack`` reconstructs the original file and the manifest carries the
+  original SHA-256 and length of every member, so an install proves
+  byte-identity before anything is used.
+
+Runtime tensors are therefore provably unchanged and accuracy cannot move.
+
+    python package_release_bps.py pack   parity --level 19
+    python package_release_bps.py unpack local_stt/releases_bps/phonon-parity.bps.tar.zst DEST
+    python package_release_bps.py verify parity          # full roundtrip proof
+"""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import io
+import json
+import subprocess
+import tarfile
+import tempfile
+import time
+from pathlib import Path
+
+import numpy as np
+
+ROOT = Path(__file__).resolve().parent
+PROFILES = {
+    "parity": ROOT / "model_v18_mlx_quint5",
+    "micro": ROOT / "model_v18_mlx_hybrid4_quint5",
+    "audio6": ROOT / "model_v18_mlx_head8audio6_quint5",
+}
+ITEMSIZE = {"BOOL": 1, "U8": 1, "I8": 1, "U16": 2, "I16": 2, "F16": 2, "BF16": 2,
+            "U32": 4, "I32": 4, "F32": 4, "U64": 8, "I64": 8, "F64": 8, "F8_E4M3": 1,
+            "F8_E5M2": 1}
+BLOCK = 16 << 20
+FORMAT = "phonon-byteplane-tar-zstd-v1"
+
+
+def sha256_bytes(data: bytes) -> str:
+    return hashlib.sha256(data).hexdigest()
+
+
+def sha256_file(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as handle:
+        for chunk in iter(lambda: handle.read(BLOCK), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def tensor_spans(raw: np.ndarray):
+    """Return (base, [(start, end, itemsize), ...]) sorted, non-overlapping."""
+    header_len = int.from_bytes(raw[:8].tobytes(), "little")
+    header = json.loads(raw[8:8 + header_len].tobytes())
+    base = 8 + header_len
+    spans = []
+    for name, meta in header.items():
+        if name == "__metadata__":
+            continue
+        start, end = meta["data_offsets"]
+        itemsize = ITEMSIZE[meta["dtype"]]
+        if itemsize > 1 and (end - start) % itemsize == 0:
+            spans.append((int(start), int(end), itemsize))
+    spans.sort()
+    merged = []
+    last_end = 0
+    for start, end, itemsize in spans:
+        if start < last_end:                       # overlapping/aliased tensors
+            continue
+        merged.append((start, end, itemsize))
+        last_end = end
+    return base, merged
+
+
+def split_file(path: Path) -> tuple[bytes, dict]:
+    raw = np.fromfile(path, dtype=np.uint8)
+    base, spans = tensor_spans(raw)
+    out = io.BytesIO()
+    out.write(raw[:base].tobytes())               # header verbatim
+    cursor = 0
+    plan = []
+    for start, end, itemsize in spans:
+        if start > cursor:                         # padding / uncovered bytes
+            out.write(raw[base + cursor: base + start].tobytes())
+        chunk = raw[base + start: base + end]
+        for i in range(itemsize):
+            out.write(chunk[i::itemsize].tobytes())
+        plan.append([start, end, itemsize])
+        cursor = end
+    tail = raw[base + cursor:]
+    if tail.size:
+        out.write(tail.tobytes())
+    meta = {
+        "base": base,
+        "plan": plan,
+        "payload_bytes": int(raw.size - base),
+        "original_bytes": int(raw.size),
+        "original_sha256": sha256_bytes(raw.tobytes()),
+    }
+    return out.getvalue(), meta
+
+
+def join_file(data: bytes, meta: dict) -> bytes:
+    raw = np.frombuffer(data, dtype=np.uint8)
+    base = meta["base"]
+    out = np.empty(meta["original_bytes"], dtype=np.uint8)
+    out[:base] = raw[:base]
+    src = base
+    cursor = 0
+    for start, end, itemsize in meta["plan"]:
+        if start > cursor:
+            width = start - cursor
+            out[base + cursor: base + start] = raw[src: src + width]
+            src += width
+        n = end - start
+        per = n // itemsize
+        block = raw[src: src + n].reshape(itemsize, per)
+        out[base + start: base + end] = block.T.reshape(-1)
+        src += n
+        cursor = end
+    remaining = meta["payload_bytes"] - cursor
+    if remaining:
+        out[base + cursor:] = raw[src: src + remaining]
+    return out.tobytes()
+
+
+def tar_info(name: str, size: int) -> tarfile.TarInfo:
+    info = tarfile.TarInfo(name)
+    info.size = size
+    info.mtime = 0
+    info.mode = 0o644
+    info.uid = info.gid = 0
+    info.uname = info.gname = ""
+    return info
+
+
+def pack(profile: str, level: int, out_dir: Path) -> dict:
+    source = PROFILES[profile]
+    out_dir.mkdir(parents=True, exist_ok=True)
+    archive = out_dir / f"phonon-{profile}.bps.tar.zst"
+
+    members = sorted(p for p in source.rglob("*") if p.is_file())
+    manifest = {
+        "release_format": FORMAT,
+        "profile": profile,
+        "compression": {"codec": "zstd", "level": level},
+        "transform": "byte-plane-split-per-tensor-v1",
+        "files": [],
+    }
+    payloads: list[tuple[str, bytes]] = []
+    for path in members:
+        rel = str(path.relative_to(source))
+        blob = path.read_bytes()
+        entry = {"path": rel, "original_bytes": len(blob),
+                 "original_sha256": sha256_bytes(blob)}
+        if path.name.startswith("model-") and path.suffix == ".safetensors":
+            transformed, meta = split_file(path)
+            entry["transform"] = meta
+            entry["stored_bytes"] = len(transformed)
+            payloads.append((rel + ".bps", transformed))
+        else:
+            entry["stored_bytes"] = len(blob)
+            payloads.append((rel, blob))
+        manifest["files"].append(entry)
+
+    manifest_bytes = (json.dumps(manifest, indent=2, sort_keys=True) + "\n").encode()
+    started = time.perf_counter()
+    args = ["zstd", "-q", f"-{level}", "-T0", "-f", "-o", str(archive), "-"]
+    if level >= 20:
+        args.insert(1, "--ultra")
+    process = subprocess.Popen(args, stdin=subprocess.PIPE)
+    assert process.stdin is not None
+    with tarfile.open(fileobj=process.stdin, mode="w|") as tar:
+        tar.addfile(tar_info("bps_manifest.json", len(manifest_bytes)),
+                    io.BytesIO(manifest_bytes))
+        for name, blob in payloads:
+            tar.addfile(tar_info(name, len(blob)), io.BytesIO(blob))
+    process.stdin.close()
+    if process.wait() != 0:
+        raise RuntimeError("zstd failed")
+    pack_s = time.perf_counter() - started
+    return {"profile": profile, "archive": str(archive),
+            "archive_bytes": archive.stat().st_size,
+            "source_bytes": sum(p.stat().st_size for p in members),
+            "level": level, "pack_seconds": pack_s,
+            "archive_sha256": sha256_file(archive)}
+
+
+def unpack(archive: Path, dest: Path) -> dict:
+    dest.mkdir(parents=True, exist_ok=True)
+    started = time.perf_counter()
+    process = subprocess.Popen(["zstd", "-q", "-d", "-c", str(archive)],
+                               stdout=subprocess.PIPE)
+    assert process.stdout is not None
+    manifest = None
+    written = []
+    with tarfile.open(fileobj=process.stdout, mode="r|") as tar:
+        for member in tar:
+            handle = tar.extractfile(member)
+            if handle is None:
+                continue
+            blob = handle.read()
+            if member.name == "bps_manifest.json":
+                manifest = json.loads(blob)
+                index = {row["path"]: row for row in manifest["files"]}
+                continue
+            if manifest is None:
+                raise RuntimeError("bps_manifest.json must be the first member")
+            rel = member.name[:-4] if member.name.endswith(".bps") else member.name
+            row = index[rel]
+            data = join_file(blob, row["transform"]) if "transform" in row else blob
+            got = sha256_bytes(data)
+            if got != row["original_sha256"] or len(data) != row["original_bytes"]:
+                raise RuntimeError(f"checksum mismatch on {rel}")
+            target = dest / rel
+            target.parent.mkdir(parents=True, exist_ok=True)
+            target.write_bytes(data)
+            written.append(rel)
+    if process.wait() != 0:
+        raise RuntimeError("zstd decompression failed")
+    missing = {row["path"] for row in manifest["files"]} - set(written)
+    if missing:
+        raise RuntimeError(f"archive is missing members: {sorted(missing)}")
+    return {"files": len(written), "unpack_seconds": time.perf_counter() - started}
+
+
+def verify(profile: str, level: int, out_dir: Path) -> dict:
+    """Pack, unpack to a temporary directory, and prove byte-identity."""
+    packed = pack(profile, level, out_dir)
+    source = PROFILES[profile]
+    with tempfile.TemporaryDirectory() as tmp:
+        stats = unpack(Path(packed["archive"]), Path(tmp))
+        mismatched = []
+        for path in sorted(p for p in source.rglob("*") if p.is_file()):
+            rel = path.relative_to(source)
+            other = Path(tmp) / rel
+            if not other.exists() or sha256_file(other) != sha256_file(path):
+                mismatched.append(str(rel))
+    packed.update(stats)
+    packed["roundtrip_byte_identical"] = not mismatched
+    packed["mismatched"] = mismatched
+    return packed
+
+
+def main() -> None:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("command", choices=("pack", "unpack", "verify"))
+    ap.add_argument("target")
+    ap.add_argument("dest", nargs="?")
+    ap.add_argument("--level", type=int, default=19)
+    ap.add_argument("--out-dir", type=Path, default=ROOT / "releases_bps")
+    args = ap.parse_args()
+
+    if args.command == "unpack":
+        print(json.dumps(unpack(Path(args.target), Path(args.dest)), indent=2))
+        return
+    fn = pack if args.command == "pack" else verify
+    result = fn(args.target, args.level, args.out_dir)
+    ratio = 100 * result["archive_bytes"] / result["source_bytes"]
+    result["percent_of_source"] = ratio
+    print(json.dumps(result, indent=2, sort_keys=True))
+    print(f"{args.target}: {result['source_bytes']/1e6:.1f} MB -> "
+          f"{result['archive_bytes']/1e6:.1f} MB ({ratio:.2f}%)")
+
+
+if __name__ == "__main__":
+    main()

phonon-audio6.bps.tar.zst

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:214c3b45aa57257013811a53f99a905848466ad4ab21f2b2f8368f7ac79427b2
+size 415077202

verify_install.py

@@ -0,0 +1,76 @@
+#!/usr/bin/env python3
+"""Verify a Phonon deployment artifact without loading or mutating it."""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import platform
+import sys
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parent
+MODELS = {
+    "parity": ROOT / "model_v18_mlx_quint5",
+    "audio6": ROOT / "model_v18_mlx_head8audio6_quint5",
+    "micro": ROOT / "model_v18_mlx_hybrid4_quint5",
+}
+# The three published models, and the only values shown in `--help`.
+PUBLIC_PROFILES = ("parity", "audio6", "micro")
+
+
+def sha256(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as handle:
+        while chunk := handle.read(8 << 20):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    # Every key in MODELS stays valid; only the three published models are
+    # advertised.  `metavar` controls the help text, `choices` controls what is
+    # accepted, so the unpublished ones remain checkable by name.
+    parser.add_argument(
+        "--profile",
+        choices=MODELS,
+        metavar="{" + ",".join(PUBLIC_PROFILES) + "}",
+        default="audio6",
+    )
+    args = parser.parse_args()
+    model = MODELS[args.profile]
+    if platform.machine() != "arm64":
+        raise RuntimeError("the optimized local runtime requires Apple Silicon")
+    manifest = json.loads((model / "packed_manifest.json").read_text())
+    if manifest.get("status") != "PASS":
+        raise RuntimeError("packed manifest is not PASS")
+    if manifest.get("source_checkpoint_sha256") != (
+        "27f01f214a0c0916944118458d0f43791b5377431fb6a230d7a2f4248368a49e"
+    ):
+        raise RuntimeError("this artifact does not match the published Phonon-1 release checkpoint")
+    if len(manifest.get("modules", [])) != 196:
+        raise RuntimeError("expected 196 packed decoder layers")
+    total = 0
+    for row in manifest["shards"]:
+        path = model / row["name"]
+        if path.stat().st_size != row["bytes"]:
+            raise RuntimeError(f"size mismatch: {path.name}")
+        actual = sha256(path)
+        if actual != row["sha256"]:
+            raise RuntimeError(f"SHA-256 mismatch: {path.name}")
+        total += path.stat().st_size
+        print(f"PASS {path.name} {actual[:16]}…")
+    if total != manifest["total_bytes"]:
+        raise RuntimeError("packed byte total mismatch")
+    print(
+        f"PASS Phonon {args.profile} model: {len(manifest['modules'])} linears, "
+        f"{total / 1e9:.3f} GB"
+    )
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())