{ "seed": 1616, "math_verification": { "threshold_sweep": [ { "tau": 0.0, "predicted_post_fraction_P_rho_ge_tau": 0.659, "observed_post_fraction": 0.659, "post_uphill_rate": 0.346, "switch_uphill_rate": 0.023916666666666666, "post_mean_true_alignment": 0.23465021696378952, "chosen_mean_true_alignment": 0.5225910290545672, "post_mean_inner": 3.7098687082536306, "switch_mean_inner": 7.275880389094258 }, { "tau": 0.02, "predicted_post_fraction_P_rho_ge_tau": 0.6564166666666666, "observed_post_fraction": 0.6564166666666666, "post_uphill_rate": 0.346, "switch_uphill_rate": 0.022083333333333333, "post_mean_true_alignment": 0.23465021696378952, "chosen_mean_true_alignment": 0.5238502896959694, "post_mean_inner": 3.7098687082536306, "switch_mean_inner": 7.285496323498012 }, { "tau": 0.05, "predicted_post_fraction_P_rho_ge_tau": 0.6530833333333333, "observed_post_fraction": 0.6530833333333333, "post_uphill_rate": 0.346, "switch_uphill_rate": 0.020583333333333332, "post_mean_true_alignment": 0.23465021696378952, "chosen_mean_true_alignment": 0.5251752698351125, "post_mean_inner": 3.7098687082536306, "switch_mean_inner": 7.293756898064051 }, { "tau": 0.1, "predicted_post_fraction_P_rho_ge_tau": 0.64675, "observed_post_fraction": 0.64675, "post_uphill_rate": 0.346, "switch_uphill_rate": 0.018083333333333333, "post_mean_true_alignment": 0.23465021696378952, "chosen_mean_true_alignment": 0.5269429164262665, "post_mean_inner": 3.7098687082536306, "switch_mean_inner": 7.300649061591103 } ], "linear_descent_identity": { "predicted_delta": 0.9385206011733538, "observed_delta": 0.9385206011733537, "absolute_error": 1.1102230246251565e-16, "eta_sweep": [ { "eta": 0.01, "predicted_delta": 0.13407437159619337, "observed_delta": 0.13407437159619337, "abs_error": 0.0 }, { "eta": 0.03, "predicted_delta": 0.40222311478858014, "observed_delta": 0.4022231147885801, "abs_error": 5.551115123125783e-17 }, { "eta": 0.07, "predicted_delta": 0.9385206011733538, "observed_delta": 0.9385206011733537, "abs_error": 1.1102230246251565e-16 }, { "eta": 0.15, "predicted_delta": 2.0111155739429005, "observed_delta": 2.0111155739429005, "abs_error": 0.0 }, { "eta": 0.3, "predicted_delta": 4.022231147885801, "observed_delta": 4.022231147885801, "abs_error": 0.0 } ] }, "noise_sweep": [ { "heldout_noise_std": 0.0, "routing_agreement": 1.0 }, { "heldout_noise_std": 0.25, "routing_agreement": 0.9743333333333334 }, { "heldout_noise_std": 0.55, "routing_agreement": 0.9373333333333334 }, { "heldout_noise_std": 1.0, "routing_agreement": 0.8823333333333333 }, { "heldout_noise_std": 2.0, "routing_agreement": 0.765 } ] }, "quadratic_comparison": { "post": { "final_loss": 0.1444657301426662, "best_loss": 0.04207063737212821, "uphill_steps": 59, "post_fraction": 1.0 }, "pre": { "final_loss": 10.456574724429673, "best_loss": 7.509928272102478, "uphill_steps": 59, "post_fraction": 0.0 }, "switch": { "final_loss": 0.18319758786532345, "best_loss": 0.040381919404750846, "uphill_steps": 57, "post_fraction": 0.75 } } }