CubicQuant bounded evaluation (arXiv:2608.06763v1)

Verdict

Supported for bounded scalar-reconstruction research; delegate model quality and hardware performance. The independently read fixture produces a deterministic 24-row ledger (three public synthetic distributions x W1-W8). integrate means CubicQuant beat both tuned baselines in this fixture; otherwise the row says abstain. This is not a universal format selection.

Pinned contract and provenance

Dimension Pin
Research artifact arxiv:2608.06763v1, https://arxiv.org/abs/2608.06763v1
Artifact digest v1 PDF SHA-256 245a523fd1b06203c123e6b03c39ed8e3cef107dd7cc33917b280511e71c9df0
Fixture internal/cubicquanteval/testdata/evaluation-v1.json; schema fak.cubicquanteval.fixture/v1
Recipe cubicquant-bounded-reconstruction-v1@1; seed 42; 384 samples/distribution; groups of 128; W1-W8; shape grid step 0.25; 17 scale candidates
Runtime fak/internal/cubicquanteval@contract-v1; Go stdlib CPU delegate
Model synthetic-reference-distributions@splitmix64-boxmuller-inversecdf-v1; no weights (weights_sha256=none)
Fixture digest SHA-256 9e6cab3b9593d08157113e99349241a3b658216122ff8f550de0417e9d4d02a8

The paper’s format is preserved rather than replaced by a fak-only artifact: W1 is the paper’s separately defined symmetric binary case. For W2-W8, M=2^(B-1)-1, t=i/M, and reconstruction magnitudes are s[a*t + b*t^2 + (1-a-b)*t^3]; the implementation rejects shape candidates whose derivative is negative on [0,1]. The paper and this fixture both use group size 128 and seed 42. This bounded fixture deliberately lowers the paper’s 15,360 samples per distribution to 384 (three groups) so the complete W1-W8 contract remains fast and reviewable.

Evidence taxonomy

Baselines and decision rule

“Tuned uniform” searches clipping scale for the cubic family’s exact uniform special case (a,b)=(1,0). “Tuned non-uniform” is a symmetric per-group Lloyd-Max reconstruction oracle initialized from magnitude quantiles; it is a quality reference, not an executable format or performance baseline. CubicQuant searches the pinned feasible cubic shape/scale grid. All three use each same group. A row is integrate only when cubic RMSE is strictly lower than both tuned baselines; ties and losses abstain. Percentage is (baseline-cubic)/baseline*100; negative means cubic lost.

Independent result ledger

Distribution Width Cubic RMSE Tuned uniform RMSE Tuned non-uniform RMSE vs uniform vs non-uniform Decision
uniform W1 0.286632 0.286632 0.286632 0.00% 0.00% abstain
uniform W2 0.190458 0.190458 0.190126 0.00% -0.17% abstain
uniform W3 0.078525 0.080111 0.078138 1.98% -0.49% abstain
uniform W4 0.035324 0.037172 0.034541 4.97% -2.27% abstain
uniform W5 0.017559 0.018147 0.016731 3.24% -4.95% abstain
uniform W6 0.008603 0.009110 0.008120 5.57% -5.94% abstain
uniform W7 0.004177 0.004561 0.003234 8.42% -29.17% abstain
uniform W8 0.002139 0.002216 0.000139 3.44% -1,442.72% abstain
gaussian W1 0.591627 0.591627 0.591627 0.00% 0.00% abstain
gaussian W2 0.481985 0.481985 0.436324 0.00% -10.47% abstain
gaussian W3 0.200548 0.213962 0.199902 6.27% -0.32% abstain
gaussian W4 0.089293 0.105564 0.102086 15.41% 12.53% integrate
gaussian W5 0.044038 0.054152 0.058876 18.68% 25.20% integrate
gaussian W6 0.020757 0.026265 0.033748 20.97% 38.50% integrate
gaussian W7 0.010377 0.012873 0.013135 19.39% 21.00% integrate
gaussian W8 0.005021 0.006460 0.000357 22.28% -1,307.53% abstain
laplace W1 0.708023 0.708023 0.708023 0.00% 0.00% abstain
laplace W2 0.544137 0.544137 0.503585 0.00% -8.05% abstain
laplace W3 0.225873 0.257572 0.228876 12.31% 1.31% integrate
laplace W4 0.107411 0.130767 0.134655 17.86% 20.23% integrate
laplace W5 0.047483 0.066849 0.087658 28.97% 45.83% integrate
laplace W6 0.023619 0.032619 0.064932 27.59% 63.63% integrate
laplace W7 0.011637 0.016069 0.030279 27.58% 61.57% integrate
laplace W8 0.005897 0.008083 0.000099 27.05% -5,871.86% abstain

The ledger’s narrow envelope is intentional: 384 synthetic values per distribution on the runtime named by the witness, with no model weights and no accelerator kernel. In particular, high-width percentage deltas can look large when the oracle denominator is near zero; the absolute RMSE columns are authoritative and no throughput or model-quality inference is permitted.

Typed boundary behavior

Input Outcome / reason
Pinned reconstruction fixture supported / CUBICQUANT_EVALUATED
Unknown schema unsupported / UNKNOWN_SCHEMA_VERSION
Tampered artifact, recipe, runtime, or model provenance unsupported / PROVENANCE_MISMATCH
Unknown scope or unsupported recipe combination unsupported / CUBICQUANT_COMBINATION_REJECTED
Model-quality request delegate / CUBICQUANT_QUALITY_REROUTE
Hardware-performance request delegate / CUBICQUANT_ACCELERATOR_REROUTE

There is no silent fallback. The result repeats artifact, recipe, runtime, and model provenance, labels evidence as observed over modeled-synthetic inputs, reports its exact OS/architecture/Go envelope, and carries the fixture digest.

Reproduction and witness

From the repository root:

go test ./internal/cubicquanteval -run TestPinnedFixtureProducesCompleteObservedLedger -v
# Windows host policy: run the package/full validation under WSL as required by AGENTS.md.
fak validate --mine internal/cubicquanteval --mine docs/research/quantization/cubicquant.md
fak buildcheck --vet --mine internal/cubicquanteval/contract.go --mine internal/cubicquanteval/doc.go

TestPinnedFixtureProducesCompleteObservedLedger independently reads the committed JSON fixture, requires all 24 rows and pinned provenance, and rejects any result not labeled observed-over-modeled. Boundary tests prove unknown schema, tampered provenance, unsupported scope, and both delegation paths. The deterministic read-back test evaluates the fixture twice and compares the serialized ledgers.