Upper bound
0 / 163,840
Epsilon-regret violations across four finite-model families and five epsilon/delta settings.
The Optimal Sample Complexity of Linear Contracts
Clean-room CPU reproduction · OpenReview ry5HitnXzc · exact finite evidence at the paper's stated constants
Failure bound
0.0003657
Largest one-sided 95% upper bound, below the smallest tested delta=0.01.
Verification
15 / 15
Deterministic tests passing on local CPU.
Matching lower-bound mechanism
The cited two-type/two-action construction was rebuilt directly. Both wrong-region gaps equal 2 epsilon. Across 16 exact Bayes tests, N epsilon^2 / ln(1/delta) stays between 0.00834 and 0.02153, exercising the claimed rate.
Negative control
Shared latent types violate i.i.d. sampling and fail, showing the positive certificate is not an always-success check.
Scope boundary
Finite executable certificates support the theorem mechanisms, constants, and scaling; they cannot prove a universal distribution-free theorem. The paper states constant 6912 but later uses 13824 in its proof, and the cited lower-bound probabilities require epsilon below 1/8.
Reproduction bundle
Source, configurations, tests, deterministic CSV/JSON results, requirements, and source audit are packaged together.