brandonmusic/shapleymcg-qwen3-30b-a3b-reproducibility

ShapleyMCG Qwen3-30B-A3B reproducibility artifacts

This dataset preserves the calibration statistics, exact corrected-R10 EXL3/MCG candidates, source and corpus identities, BF16 teacher/student logits, tokenwise KLD, allocations, attribution ledgers, hashes, and publication receipts for the Qwen3-30B-A3B experiments in `brandonmmusic-max/shapleymcg`.

The complete cross-checkpoint `results ledger` and `method specification` distinguish the predecessor routed-p2 allocator from the full causal method. The fresh machine-readable `Base completion audit` records the immutable tree identities, layer receipt closure, model-manifest audit, corpus lineage hashes, and independent KLD replay used for final verification.

Causal allocation result

The complete two-level Aumann–Shapley/Fisher allocation improves on the historical Hessian/router allocation at an identical 3.5 logical bits per routed-expert weight. The primary comparison holds the Base revision, exact MCG candidates, BF16 teacher, 20,480 token positions, SDPA attention, every non-expert weight, and 9,216 K3 plus 9,216 K4 choices fixed.

This is a 7.731248% KLD reduction and a 0.268555 percentage-point top-1 gain. Independent NumPy float64 replay reproduced both stored per-token vectors with zero maximum difference. A separate 2,047-position SDPA comparison also favors the causal allocation by 11.896262%.

• —Parent: Qwen/Qwen3-30B-A3B-Base@1b75feb79f60b8dc6c5bc769a898c206a1c6a4f9
• —Candidate inventory: 205cdd1e0b36dcaa545e551df9ca3eb88a6ebb385276236fc45a3f105c25b4e9
• —Causal allocation: 10bb3b71a2258500b281e3b37d57657e5e0ecf9318bb9c1a3fc6dbd2326e63c4
• —20k panel: 6c509f3496d5a1fe6739c58ffa152c3d4202932f9d8a6514e55d17f8a4cfa741
• —20k comparison: 8d8438df03d0f9a34a4a5a79912dd232baa401f24d8e6e14bb942bf00062a608
• —Causal/historical verifier seals: 6e4215f276d69ca53ce1ed29c2a900f5e040d41594a9e2ab24cb247ddec90265 / 752e83112522f1f8bfe2cbf7797196cc7961164b7617ccb22d92f7e60c8a1911

The raw five-node layer attribution summed to 0.0306227734756427 against a measured uniform-K4 endpoint of 0.04401261771637709. The unresolved 0.013389844240734392 remainder (30.42%) is published explicitly. Reconciled expert scores close to measured layer attribution and reconciled layer scores close to endpoint KLD, but exact bookkeeping closure is not represented as raw proxy additivity.

The full causal evidence is under causal-arm-v3/. Teacher and both allocation arms' panel logits, reports, per-token vectors, independent verifiers, attribution/reconciliation artifacts, and retention receipts are preserved.

Verified validation model publication

The exact expanded validation reconstruction is published at `brandonmusic/Qwen3-30B-A3B-ShapleyMCG-K34-Validation-Reconstruction`.

• —Manifest-verified model revision: c3447f3f8e231ae83afa10ba263346d1ceb98c11
• —Publication-receipt revision: fbdd05904fd4018dea9fd0f25ed4800c15df1493
• —Model manifest seal: 32fec13a96c6c85eb394939a6eb511c821bd05e8ac0960268852d33150d0cde1
• —Publication receipt seal: e7b3b344321c11e872fc006cbd009c935b9fdbf0fdb8b2b0a99197cce14bffe8
• —Dataset receipt mirror: causal-arm-v3/validation-model-hf-publication-receipt.json
• —Verified content inventory: 31 files and 61,079,798,645 bytes; every file matched its remote size and either Hub LFS SHA-256 or Git-blob SHA-1.
• —Post-assembly verification: all 18,432 selected expert-matrix reconstructions were installed, and the reloaded checkpoint reproduced the sealed 2,047-position student logit tensor exactly (max_abs_delta = 0).

The receipt describes and verifies the immutable model revision above. The repository head may advance for documentation without changing that sealed content revision.

Historical fixed-Hadamard control

The first 48-layer corrected-R10 K3/K4 control remains preserved as a separate historical experiment:

• —exact routed-expert rate: 9,216 K3 plus 9,216 K4 matrix choices;
• —2,047-position GLM-style mean KLD: 0.06335949321311507;
• —historical allocation: 5edda4ab29f9961f6f2ab5be3aab6cc0aaa3817afa325b40fd5063fef05de034;
• —KLD report: 3685fa5eac6064aa6bd5e51ddc53fb10d791c0744c955f551283fdcad62469ff.

Uniform K3/K4 endpoints and five score-blind exact-3.5 controls are retained under results/qwen3-30b-a3b-v1. Those earlier panels used their documented attention backend and must not be silently mixed with the matched SDPA causal comparison above.

Reconstructed paper panel

The reconstructed Hill-style BFCL/RULER panel measured 0.018260861970005038 over 32,752 positions with the predecessor pipeline. It is 48.3% below the paper's reported 0.0353 additive row and 57.4% below its 0.0429 ModelOpt row, but it is cross-system evidence, not a strict allocator head-to-head: exact author tokens were unavailable, the parent and arithmetic differ, and the new causal allocation has not been measured on that panel.

Format and runtime boundary

The exact validation model installs selected codec-produced reconstructions into a Transformers BF16 checkpoint. It is intentionally an expanded validation reconstruction, not a compact 3.5-bpw checkpoint. Current official BTX couples gate/up rates, while this allocation selects those matrices independently. CUDA-graph, packed-runtime, and throughput qualification are not claimed by these KLD artifacts.

Attribution

• —Base model: the Qwen team.
• —EXL3/TRELLIS and MCG codec lineage: turboderp and ExLlamaV3 contributors.
• —Aumann–Shapley quantization precedent: Joshua Hill, *Saturation Makes Quantization Error Additive*, and NVIDIA Model Optimizer PR #2183.
• —Routed Fisher/Jacobian attribution, explicit residual reconciliation, MCG integration, experiments, and publication: Brandon Music / ShapleyMCG.

See the repository's docs/REFERENCES.md, THIRD_PARTY_NOTICES.md, and THIRD_PARTY_LICENSES for complete citations and license boundaries.