squaredcuber's picture
|
download
raw
2.17 kB

Deterministic experiment plan

The full matrix contains 14,000 independently identified tasks:

Suite Tasks Role
Gaussian portfolio, paper settings 5,000 Main C1/C2/C3 evidence
Gaussian portfolio, lambda_p=0 5,000 Coverage and conservatism ablation
Absolute-error regression 500 Main C1/C2/C3 evidence
Ten-dimensional Gaussian portfolio 1,000 Appendix F.1
Discrete portfolio 1,000 Appendix F.2.1
Gaussian-mixture portfolio 1,000 Appendix F.2.2
Squared-error regression 500 Appendix F.3

Every task identity freezes suite, distribution, dataset replicate, sample size, loss, transport geometry, hyperparameters, and six independent RNG streams. Distribution parameters remain constant across replicates and sample sizes. A dataset replicate is nested across sample sizes so n=10 is an exact prefix of n=20; bootstrap and optimizer streams remain task-specific.

The runner must preserve scalar objective, penalty, gradient norm, stopping diagnostics, solver residual maxima, and wall time for every iteration. It preserves complete decisions, metric matrices, ambiguity radii, gradients, solver receipts, and optimizer state at iteration 0, every 100 iterations, the first stop trigger, and the terminal state. Both streams are content-hashed and mutually bound. A terminal summary without the scalar trace and checkpoint stream is invalid. This retains scientific auditability without creating a terabyte-scale duplicate of unchanged matrices.

The official repository is a pinned reference only. Its public examples use five or ten bootstraps and a hardcoded Adam path, while Algorithm 2 and Table 1 specify constant-step gradient descent. The independent primary route follows Algorithm 2. A paired released-source Adam route is required as an implementation sensitivity and cannot promote a paper claim by itself.

No paper-scale run is authorized. The fixed-100 and fixed-1,000 CPU gates measured update cost but did not establish primary stopping times. A paired optimizer probe and a compact-trace stopping-time escalation must pass before CPU Jobs approval.

Xet Storage Details

Size:
2.17 kB
·
Xet hash:
e644a33bbdfd12ba8265e7144aae4d1f8ca09fbb721baea3208f886ebb87a309

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.