Run all three categorical systems from the same strictly interior state with the same stationary reward sequence. Compare every canonical state, not just the endpoint. This is a software/theorem reproduction check.
Choose one named violation, a frozen epsilon grid, discrepancy metric, horizon, and seed set. The baseline is multiplicative weights under the correspondence assumptions. The target receives the violation. Record both per-seed discrepancies and their aggregate; do not silently discard failed trajectories.
The output is the empirical map epsilon -> delta(epsilon). A first-order coefficient
is estimated only from a caller-declared source window through an origin-constrained
least-squares fit. The raw curve remains the primary artifact.
Fit a source coefficient on explicitly named source-domain data, freeze it, and apply it to target epsilons without fitting target observations. Report predictions, observations, residuals, and source/target identities. The bundled command is an instrumentation demonstration, not evidence that the coefficient transports in a broader system class.
The Gaussian optimizer is first checked on a quadratic objective with analytic expected gradient and Fisher geometry. The contextual bandit is first checked against exact expected policy gradients. Finite-sample estimators are compared with these reference directions before any larger problem is admitted.
ACL-002 is frozen under preregistrations/ACL-002/. It uses deterministic landscapes,
not seed replication. Probabilities and sensitivities are row vectors, mutation is
right multiplication by a row-stochastic matrix, and the Jacobian stores input index
first and output index second.
The primary endpoint-L1 analysis reports two separate held-out verdicts: the derived
zero-fit prediction and a landscape-balanced median calibration fitted on source
landscapes. The secondary oriented divergence is exactly KL(q_T || p_T) and uses its
preregistered Fisher-curvature coefficient. Stress epsilons and secondary metrics are
structurally absent from primary gate calculation.
Within each regular target landscape, the primary score is the maximum relative error across the three strict-confirmatory epsilons. Across landscape scores, the frozen Type-7 median and Q0.90 gates are applied independently to zero-fit and calibrated predictions.
The strict confirmatory epsilon set ends at 1e-3; 3e-3 and 1e-2 are extended-local
only, and 3e-2 and 1e-1 remain stress points. Iterative perturbed trajectories must
agree with an independent normalized matrix-power oracle before analysis. All numerical
guards are separately named and frozen in the manifest.
The targets are a deterministic held-out benchmark assembled from recombinations of the same catalogs used by the source split. Its median and Q0.90 thresholds are descriptive criteria for within-family transport, not population confidence claims or evidence of transport across adaptive-system classes.
Preparing or validating this bundle may recompute clean trajectories and their analytic tangents. It must not generate an epsilon-positive trajectory from the confirmatory manifest until the public preregistration SHA is reviewed and explicitly approved.