Tinnci · Tinnci · Jul 12, 2026 · Jul 12, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -24,6 +24,8 @@ This project follows semantic versioning once published to PyPI.
 - Added a pinned external Whisper evaluator specification and Colab-only transcript workflow for
   precomputed downstream WER/CER evidence without adding an ASR package dependency.
 - Added a Colab-only verified MUSHRA bundle export for the licensed LibriSpeech sinc/LavaSR matrix.
+- Added a bounded Colab-only MossFormer2 feasibility workflow covering minimal checkpoints, safe
+  conversion, offline CPU inference, channel behavior, alignment, and parity-fixture constraints.
 
 ## 0.6.0 - 2026-07-12
 

diff --git a/ROADMAP.md b/ROADMAP.md
@@ -79,7 +79,7 @@ human ratings is a follow-up study, not a prerequisite for completing the export
 Do not collapse quality, downstream usefulness, speed, stability, and governance into one ranking.
 The evaluation policy lives in [docs/EVALUATION.md](docs/EVALUATION.md).
 
-## Priority 3: One Candidate Spike, Not Three Implementations
+## Completed: One Candidate Spike, Not Three Implementations
 
 ClearerVoice `MossFormer2_SR_48K` is the first candidate to revisit because it has a clear 48 kHz
 speech SR contract and identifiable checkpoint files. The spike must stop before backend
@@ -96,6 +96,12 @@ Resemble Enhance remains a possible external speech-enhancement backend, not an
 package-owned 48 kHz SR path. Detailed evidence lives in
 [docs/BACKEND_CANDIDATES.md](docs/BACKEND_CANDIDATES.md).
 
+The Colab CPU spike passed with only the pointer plus `m` and `g` checkpoints. Both state dicts can
+be loaded with `weights_only=True` and converted exactly to safetensors. Offline 16 kHz mono/stereo
+and native 48 kHz inputs ran successfully; stereo channels were preserved. The sub-second fixture
+exposed a consistent 224-sample output shortening and cross-process floating-point drift, so any
+future backend needs explicit alignment plus tolerance-based parity rather than hash equality.
+
 ## Priority 4: Optimize Only After Measurement
 
 Keep `lavasr-compat` on `torch-eager` until another provider has real-weight benchmarks and parity

diff --git a/docs/ACCELERATORS.md b/docs/ACCELERATORS.md
@@ -269,6 +269,23 @@ The bundle uses MUSHRA metadata with seed `0` and four stimuli per trial: refere
 anchor, sinc output, and LavaSR output. Public stimuli contain only blind IDs and generic paths;
 backend, role, and source mapping remain in the separate answer key.
 
+Run the bounded MossFormer2 feasibility spike on a fresh CPU session. Checkpoint download,
+inspection, conversion, and real inference all remain remote:
+
+```sh
+colab new -s asr-mossformer2
+colab exec -s asr-mossformer2 -f examples/colab_mossformer2_spike.py --timeout 3600
+colab download -s asr-mossformer2 \
+  /content/asr-mossformer2-spike.tar.gz \
+  runs/asr-mossformer2-spike.tar.gz
+colab stop -s asr-mossformer2
+```
+
+The script pins both upstream revisions, downloads only the pointer and two inference checkpoints,
+forces offline mode before model construction, uses an isolated Python worker after dependency
+installation, and records safetensors round-trip plus mono/stereo/alignment evidence. Do not run it
+locally.
+
 For upstream LavaSR parity, install upstream dependencies deliberately and use the environment gates
 in [../tests/README.md](../tests/README.md). Validate AudioSR separately because its external package
 owns checkpoint and accelerator behavior.

diff --git a/docs/BACKEND_CANDIDATES.md b/docs/BACKEND_CANDIDATES.md
@@ -5,14 +5,15 @@ This document owns implementation-grade evidence for model candidates. Admission
 
 ## Current Decision
 
-Do not start another package-owned backend until v0.6.0 is released and broader LavaSR evidence is
-recorded. When capacity returns, run one bounded ClearerVoice `MossFormer2_SR_48K` feasibility spike.
-Do not implement MossFormer2, FlowHigh, and Resemble Enhance in parallel.
+The bounded ClearerVoice `MossFormer2_SR_48K` feasibility spike is complete. It clears the checkpoint,
+license, offline-loading, CPU-fallback, and basic I/O gates, but backend implementation remains a
+separate decision because the model is heavyweight and its output has a fixed alignment offset. Do
+not implement MossFormer2, FlowHigh, and Resemble Enhance in parallel.
 
 | Candidate | Domain | Decision |
 | --- | --- | --- |
 | `lavasr-v2-bwe` | Speech BWE | Implemented experimental baseline; broaden evidence before stable promotion. |
-| ClearerVoice `MossFormer2_SR_48K` | Speech SR to 48 kHz | First future spike; defer implementation pending safe checkpoint and parity evidence. |
+| ClearerVoice `MossFormer2_SR_48K` | Speech SR to 48 kHz | Feasibility passed; eligible for one bounded backend plan with tolerance-based parity. |
 | Resemble Enhance | Speech enhancement/BWE to 44.1 kHz | Deferred; consider only as an optional external backend. |
 | AudioSR `basic` / `speech` | General audio / speech | Keep the existing external-package backend; do not rewrite now. |
 | FlowHigh | General audio SR to 48 kHz | Deferred; CUDA-first runtime and checkpoint-license/provider blockers remain. |
@@ -24,6 +25,7 @@ Source:
 
 - repository: <https://github.com/modelscope/ClearerVoice-Studio>
 - code license: Apache-2.0
+- checkpoint license: Apache-2.0 in the pinned Hugging Face model card and repository tag
 - model repository: <https://huggingface.co/alibabasglab/MossFormer2_SR_48K>
 - inspected source revision: `6b3774dc79c46ae8bed2a4fa5f706f0ac8c75c61`
 - inspected model revision: `39eb1f25ea84f5e0315ade9ac0070fff216fc690`
@@ -38,24 +40,38 @@ Observed weight files:
 | `last_best_checkpoint_m.pt` | 218,471,889 | `6cbadb2b6b839e444bb65223c69eea162c8ad08f36e9d0a64144672c4095ab36` |
 | `do_03925000` | 1,744,458,369 | `549035d29a03928a854815b9c1b21c02d825845aa2e5f093af45694a236c1f19` |
 
-The public SR loader appears to use the pointer plus `m` and `g` checkpoints; the larger file should
-not enter a managed manifest unless the spike proves it is required.
-
-Blockers:
-
-- upstream loading can download implicitly when files are absent;
-- checkpoints use PyTorch `.pt` loading and need a safe conversion/loading plan;
-- package-owned module boundaries and key mapping are not reproduced;
-- a short upstream CPU smoke and parity fixture have not been recorded;
-- the clean-speech limitation must remain visible in catalog metadata.
-
-Spike exit criteria:
-
-1. Inspect tensor keys for the two required checkpoints from a temporary cache.
-2. Run upstream CPU inference on a sub-second fixture without provider calls during inference.
-3. Record input/output sample-rate, channel, alignment, and preprocessing behavior.
-4. Decide whether conversion to a safer package-owned format is practical.
-5. Define a parity fixture before opening backend implementation.
+The Colab spike proved that the pointer plus `m` and `g` checkpoints are sufficient. The 1.74 GB
+`do_03925000` file was not downloaded and must not enter a managed manifest.
+
+Recorded Colab CPU evidence:
+
+- `last_best_checkpoint_m.pt` contains a `mossformer` state with 929 tensors and 54,497,666
+  parameters; `last_best_checkpoint_g.pt` contains a `generator` state with 311 tensors and
+  55,150,402 parameters;
+- both files load with `torch.load(..., weights_only=True)` and round-trip exactly through
+  safetensors after cloning shared rotary-frequency buffers;
+- converted sizes are 218,118,376 and 220,633,048 bytes, with recorded deterministic conversion
+  hashes in the evidence bundle;
+- a generated 0.25-second 16 kHz mono fixture ran on CPU with Hugging Face offline mode enabled;
+  model construction took 14-21 seconds and first inference took 4.1-5.2 seconds on Colab CPU;
+- repeated inference in one process was bit-exact, but separate clean sessions produced small
+  floating-point differences, so parity must use numeric/audio tolerances rather than a WAV hash;
+- 16 kHz mono input produced 48 kHz-domain shape `[1, 11776]`; 16 kHz stereo preserved two
+  independently processed channels with shape `[2, 11776]`;
+- both a 4000-sample 16 kHz input and a 12000-sample 48 kHz input produced 11776 samples, a fixed
+  `-224` sample (`-4.67 ms`) alignment delta for this fixture;
+- upstream preprocessing resamples inputs to 48 kHz, processes channels independently, and marks SR
+  output as 48 kHz rather than resampling it back to the original input rate.
+
+The package-owned parity fixture should generate the same seeded 440 Hz mono input, compare the
+pre-alignment upstream float output, explicitly account for the 224-sample generator boundary loss,
+and use bounded waveform/spectral tolerances. Do not use a cross-process file hash as the parity
+oracle. The clean-speech limitation must remain visible in catalog metadata.
+
+Implementation is now technically admissible, but it should be accepted only as a bounded,
+self-contained backend task with two managed safetensors files, explicit alignment policy, no
+implicit downloads, and gated real-weight tests. Its approximately 110 million parameters and slow
+CPU startup/inference make it unsuitable as the default lightweight backend.
 
 ## Resemble Enhance