StrataFold

Fold experts, not precision.

StrataFold is an independent clean-room laboratory for structural Mixture-of-Experts compression without additional quantization. It investigates whether routing-aware shared structure can reduce stored expert weights while preserving each source tensor's native dtype. The primary adapter targets deepseek-ai/DeepSeek-V4-Flash-0731 at the exact pinned revision 7872f01b1d1fe23eabc4c98b48bffcef5a386062.

The target baseline is described from bounded official metadata. Its config declares expert_dtype=fp4, torch_dtype=bfloat16, and an FP8 e4m3 quantization configuration. These are configuration declarations—not shard-header or payload observations. StrataFold does not call the target unquantized and does not count a dtype or bit-width change as structural compression.

The adopted M1 architecture view is source-backed by committed metadata and exact evidence surfaces. It does not represent a checkpoint run or a compression result.

Validated M1 target genome

M1 is now a validated, pinned-official-metadata snapshot:

• Topology declarations: 43 hidden layers, hidden size 4096, 256 routed experts plus one shared expert per layer, top-6 routing, three hash layers, one next-token-prediction layer, and DSpark target layers 40–42.
• Index declarations: 72,317 tensor names, 48 listed shard filenames, and 166,878,536,440 declared tensor payload bytes.
• Verified API projection: 166,886,535,336 weight-shard bytes, 12,125,738 non-weight-file bytes, and 166,898,661,074 total repository-file bytes.
• API-reported parameter classes: 304,180,418,494 total, reproduced from the committed receipt and not recomputed from shard headers.
• Derived difference: 7,998,896 bytes between API-reported weight-shard bytes and index-declared tensor payload bytes. It is unattributed metadata—not measured compression or proven container overhead.

Declared topology	Separate byte ledgers
[source-reproduced] configuration declarations only	[source-reproduced] API/index totals; the [derived] inset is explicitly not to scale

Evidence surfaces are committed directly:

• raw M1 JSON
• raw M1 CLI transcript
• raw M1 capture provenance
• deliberate rejection record
• project provenance ledger
• snapshot manifest
• repository projection receipt
• claim registry

Every quantitative claim is registered in CLAIMS.json with an allowed evidence tag, exact committed evidence paths, and a reproduce command.

Rasterized exact verified CLI transcript — this is not an OS-terminal screenshot. The image links to the exact committed text bytes.

The full visual atlas documents every caption and evidence boundary. Remaining focused views: expert census, 48-shard inventory, API-reported parameter classes, generator manifest, and adoption provenance.

Setup and verification

The control plane is dependency-free and requires Python 3.10+ and make:

git clone https://github.com/omar07ibrahim/stratafold.git
cd stratafold
make doctor
make test
make evidence-m1-check
make evidence-m1-verify
PYTHONPATH=src python3 -m stratafold inspect-target

make doctor enforces the workspace, cache, object-size, revision, and metadata-only request boundaries before any evidence command runs. make evidence-m1-check reconstructs the expected target-genome surfaces in memory and compares them byte-for-byte without rewriting; make evidence-m1-verify validates the adopted evidence identities and semantics. The separate hosted visual job renders twice and runs make visuals-m1-verify-adopted to compare all eleven generated files with the committed atlas.

Evidence status and safety

Evidence labels remain deliberately separate:

• [source-reproduced] config, topology, index, repository byte ledgers, and API-reported parameter classes were reproduced from the committed metadata and post-capture receipt;
• [derived] the 7,998,896-byte difference is arithmetic with no causal attribution;
• [measured] the validated decoder invocation found zero shard files in the snapshot, opened zero shard files, executed no target code, and performed no full-checkpoint operation;
• [unverified] the capture-time attestation is not an independent audit, and the hostwide download state is unaudited;
• [not-run] the full checkpoint is NOT DOWNLOADED / NOT RUN for the M1 project record.

Deliberate validation experiment — a temporary copied snapshot changes the same-length token config.expert_dtype: fp4 → fp3, refreshes the manifest config entry, and the real CLI rejects it. This is not an upstream incident.

The semantic rejection occurs before reviewed identity gates; its exact canonical stderr and hashes are committed in the rejection record.

This host supports generated micro-checkpoints and metadata-only validation. No M1 result is a compression-ratio, quality, throughput, active-compute, performance, or state-of-the-art claim.

See the target ledger, benchmark contract, clean-room decision, threat model, and prior-art boundary before interpreting future results.

StrataFold is not affiliated with or endorsed by DeepSeek, Hugging Face, Moonshot AI, or any cited researcher or institution.

License

Code is licensed under Apache-2.0. Documentation and original project artwork are licensed under CC BY 4.0; see LICENSE-DOCS and THIRD_PARTY_NOTICES.md.