{ "schema_version": "1.0", "generated_utc": "2026-10-27T18:21:39.533061+00:00", "scope": "Completed task-specific references actually used in the paper; separate from frozen neural results. No global reference score. The Uniform column of Table 3 is here too, under reference_id uniform_choice.", "row_count": 95, "precision_note": "87 rows preserve saved aggregate decimals; 7 rows preserve exactly the four-decimal values reported in the manuscript; the Horizon uniform mean reward preserves the one decimal printed in the paper. Rounded rows are explicit reported_rounded_source records and do not imply recovered full precision.", "rows": [ { "row_id": "rw3-heldout-legal-nll", "reference_id": "horizon_rw3", "model": "RW3 (local formula fit)", "task": "horizon", "mode": "heldout_human_history_prediction", "metric": "human_nll", "value": 0.400188875690336, "value_raw": "0.400188875690336", "value_status": "finite", "n": 6720, "aggregation": "response_weighted_mean", "unit": "nats_per_free_choice", "scope": "6_test_participants", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "parameter_fit": "25 original training participants; 28,000 free choices; forced choices retained for updates", "predictive_evaluation": "6 held-out participants; 6,720 free choices", "autonomous_evaluation": "31 original timelines including the 25 training-participant timelines; 34,720 free choices", "autonomous_action_rule": "Softmax sampling at temperature 1 with original participant seeds" }, "comparison_note": "Same Horizon target modes and published split; task fitting and softmax readout differ from frozen neural native-action protocols. Autonomous reward is not held-out reward evidence. The unavailable author executable and its archived NLL are distinct references.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rw3_scalar_replay_acceptance_20261027", "field": "/rw_common_legal_nll_eps1e8" }, "metric_variant": "legal_action_negative_log_p_plus_1e_minus_8", "epsilon": 1e-08, "n_participants": 6, "same_neural_mode": true }, { "row_id": "rw3-autonomous-reward", "reference_id": "horizon_rw3", "model": "RW3 (local formula fit)", "task": "horizon", "mode": "autonomous_generated_environment", "metric": "reward", "value": 54.559447004608295, "value_raw": "54.559447004608295", "value_status": "finite", "n": 34720, "aggregation": "free_choice_mean", "unit": "original_reward_units_per_free_choice", "scope": "all_31_original_timelines", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "parameter_fit": "25 original training participants; 28,000 free choices; forced choices retained for updates", "predictive_evaluation": "6 held-out participants; 6,720 free choices", "autonomous_evaluation": "31 original timelines including the 25 training-participant timelines; 34,720 free choices", "autonomous_action_rule": "Softmax sampling at temperature 1 with original participant seeds" }, "comparison_note": "Same Horizon target modes and published split; task fitting and softmax readout differ from frozen neural native-action protocols. Autonomous reward is not held-out reward evidence. The unavailable author executable and its archived NLL are distinct references.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rw3_accepted_summary_20261027", "field": "/generative/mean_reward_free" }, "n_timelines": 31, "same_neural_mode": true, "identical_action_readout": false }, { "row_id": "rw3-autonomous-optimal_action_rate", "reference_id": "horizon_rw3", "model": "RW3 (local formula fit)", "task": "horizon", "mode": "autonomous_generated_environment", "metric": "optimal_action_rate", "value": 0.7459389400921659, "value_raw": "0.7459389400921659", "value_status": "finite", "n": 34720, "aggregation": "free_choice_mean", "unit": "fraction", "scope": "all_31_original_timelines", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "parameter_fit": "25 original training participants; 28,000 free choices; forced choices retained for updates", "predictive_evaluation": "6 held-out participants; 6,720 free choices", "autonomous_evaluation": "31 original timelines including the 25 training-participant timelines; 34,720 free choices", "autonomous_action_rule": "Softmax sampling at temperature 1 with original participant seeds" }, "comparison_note": "Same Horizon target modes and published split; task fitting and softmax readout differ from frozen neural native-action protocols. Autonomous reward is not held-out reward evidence. The unavailable author executable and its archived NLL are distinct references.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rw3_accepted_summary_20261027", "field": "/generative/mean_optimal_choice_rate" }, "n_timelines": 31, "same_neural_mode": true, "identical_action_readout": false }, { "row_id": "rw3-autonomous-pseudo_regret", "reference_id": "horizon_rw3", "model": "RW3 (local formula fit)", "task": "horizon", "mode": "autonomous_generated_environment", "metric": "pseudo_regret", "value": 2.7289746543778803, "value_raw": "2.7289746543778803", "value_status": "finite", "n": 34720, "aggregation": "free_choice_mean", "unit": "latent_reward_gap_per_free_choice", "scope": "all_31_original_timelines", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "parameter_fit": "25 original training participants; 28,000 free choices; forced choices retained for updates", "predictive_evaluation": "6 held-out participants; 6,720 free choices", "autonomous_evaluation": "31 original timelines including the 25 training-participant timelines; 34,720 free choices", "autonomous_action_rule": "Softmax sampling at temperature 1 with original participant seeds" }, "comparison_note": "Same Horizon target modes and published split; task fitting and softmax readout differ from frozen neural native-action protocols. Autonomous reward is not held-out reward evidence. The unavailable author executable and its archived NLL are distinct references.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rw3_accepted_summary_20261027", "field": "/generative/mean_pseudo_regret_free" }, "n_timelines": 31, "same_neural_mode": true, "identical_action_readout": false }, { "row_id": "hybrid-ann-human_nll", "reference_id": "hybrid_end_to_end_ann", "model": "Independent end-to-end ANN (reimplementation)", "task": "hybrid_information_sampling", "mode": "heldout_within_participant_human_prediction", "metric": "human_nll", "value": 0.2639591167960235, "value_raw": "0.2639591167960235", "value_status": "finite", "n": 20, "aggregation": "equal_participant_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "participants": 20, "responses": 163135, "folds": 160, "folds_per_participant": 8, "iterations_per_fold": 3000, "split_unit": "Original response/trial fold within each participant; not held-out participants", "checkpoint_selection": "Lowest loss on that same validation fold across all 3,000 iterations; not nested model selection", "normalization": "Training-fold statistics only", "neural_models": "Frozen checkpoints without this task-specific fitting" }, "comparison_note": "Same 163,135 human action targets and original folds. ANN uses disclosed structured inputs and within-participant fitting; neural models use their frozen text/choice interfaces. Validation-fold checkpoint selection and training exposure differ, so this is not a matched-training superiority comparison.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "independent_ann_full_fit_acceptance_20261027", "field": "/subject_mean_response_nll" }, "n_responses": 163135, "n_folds": 160, "same_neural_mode": true, "metric_variant": "participant_mean_response_nll", "ci95": { "type": "pointwise Student-t over 20 participant means", "low": 0.2142866697685797, "high": 0.31363156382346735, "low_raw": "0.2142866697685797", "high_raw": "0.31363156382346735" } }, { "row_id": "hybrid-ann-human_agreement", "reference_id": "hybrid_end_to_end_ann", "model": "Independent end-to-end ANN (reimplementation)", "task": "hybrid_information_sampling", "mode": "heldout_within_participant_human_prediction", "metric": "human_agreement", "value": 0.9129627161693102, "value_raw": "0.9129627161693102", "value_status": "finite", "n": 20, "aggregation": "equal_participant_mean", "unit": "fraction", "scope": "all", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "participants": 20, "responses": 163135, "folds": 160, "folds_per_participant": 8, "iterations_per_fold": 3000, "split_unit": "Original response/trial fold within each participant; not held-out participants", "checkpoint_selection": "Lowest loss on that same validation fold across all 3,000 iterations; not nested model selection", "normalization": "Training-fold statistics only", "neural_models": "Frozen checkpoints without this task-specific fitting" }, "comparison_note": "Same 163,135 human action targets and original folds. ANN uses disclosed structured inputs and within-participant fitting; neural models use their frozen text/choice interfaces. Validation-fold checkpoint selection and training exposure differ, so this is not a matched-training superiority comparison.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "independent_ann_full_fit_acceptance_20261027", "field": "/participant_weighted_agreement" }, "n_responses": 163135, "n_folds": 160, "same_neural_mode": true, "metric_variant": "participant_weighted_agreement" }, { "row_id": "hybrid-ann-stop_probability", "reference_id": "hybrid_end_to_end_ann", "model": "Independent end-to-end ANN (reimplementation)", "task": "hybrid_information_sampling", "mode": "heldout_within_participant_human_prediction", "metric": "stop_probability", "value": 0.036404701456336064, "value_raw": "0.036404701456336064", "value_status": "finite", "n": 20, "aggregation": "equal_participant_mean", "unit": "probability_0_to_1", "scope": "all", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "participants": 20, "responses": 163135, "folds": 160, "folds_per_participant": 8, "iterations_per_fold": 3000, "split_unit": "Original response/trial fold within each participant; not held-out participants", "checkpoint_selection": "Lowest loss on that same validation fold across all 3,000 iterations; not nested model selection", "normalization": "Training-fold statistics only", "neural_models": "Frozen checkpoints without this task-specific fitting" }, "comparison_note": "Same 163,135 human action targets and original folds. ANN uses disclosed structured inputs and within-participant fitting; neural models use their frozen text/choice interfaces. Validation-fold checkpoint selection and training exposure differ, so this is not a matched-training superiority comparison.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "independent_ann_full_fit_acceptance_20261027", "field": "/participant_weighted_stop_probability" }, "n_responses": 163135, "n_folds": 160, "same_neural_mode": true, "metric_variant": "participant_weighted_stop_probability" }, { "row_id": "hybrid-ann-stop_readout_frequency", "reference_id": "hybrid_end_to_end_ann", "model": "Independent end-to-end ANN (reimplementation)", "task": "hybrid_information_sampling", "mode": "heldout_within_participant_human_prediction", "metric": "stop_readout_frequency", "value": 0.010760341242104746, "value_raw": "0.010760341242104746", "value_status": "finite", "n": 20, "aggregation": "equal_participant_mean", "unit": "fraction", "scope": "all", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "participants": 20, "responses": 163135, "folds": 160, "folds_per_participant": 8, "iterations_per_fold": 3000, "split_unit": "Original response/trial fold within each participant; not held-out participants", "checkpoint_selection": "Lowest loss on that same validation fold across all 3,000 iterations; not nested model selection", "normalization": "Training-fold statistics only", "neural_models": "Frozen checkpoints without this task-specific fitting" }, "comparison_note": "Same 163,135 human action targets and original folds. ANN uses disclosed structured inputs and within-participant fitting; neural models use their frozen text/choice interfaces. Validation-fold checkpoint selection and training exposure differ, so this is not a matched-training superiority comparison.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "independent_ann_full_fit_acceptance_20261027", "field": "/participant_weighted_stop_readout" }, "n_responses": 163135, "n_folds": 160, "same_neural_mode": true, "metric_variant": "participant_weighted_stop_readout" }, { "row_id": "mcpl-overall-test-nll", "reference_id": "multiple_cue_disclosed", "model": "Multiple-cue formula reimplementation", "task": "multiple_cue_judgment", "mode": "heldout_human_history_prediction", "metric": "human_nll", "value": 2.0290877869005786, "value_raw": "2.0290877869005786", "value_status": "finite", "n": 6136, "aggregation": "response_weighted_mean", "unit": "nats_per_response", "scope": "all_three_original_test_experiments", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "train_participants": 232, "train_responses": 52464, "heldout_test_participants": 28, "heldout_test_responses": 6136, "parameter_fits": "Each of three original experiments fitted separately on its training rows", "parameter_selection": "Held-out data not used for parameter selection", "feedback": "Update only after feedback actually visible to the human; retain no-feedback rows" }, "comparison_note": "Same original held-out human-response mode, split and all responses. Disclosed formula fit is distinct from unrecovered author fitting code and archived unpaired scores. Retain each output interface and NLL definition; shared task membership does not make integer, typed or binned likelihoods interchangeable.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "mcpl_full_fit_acceptance_20261028", "field": "/test/mean_nll_response_weighted" }, "n_participants": 28, "same_neural_mode": true }, { "row_id": "mcpl-exp1-test-nll", "reference_id": "multiple_cue_disclosed", "model": "Multiple-cue formula reimplementation", "task": "multiple_cue_judgment", "mode": "heldout_human_history_prediction", "metric": "human_nll", "value": 1.9985996178748873, "value_raw": "1.9985996178748873", "value_status": "finite", "n": 2800, "aggregation": "response_weighted_mean", "unit": "nats_per_response", "scope": "collsiöö2023MCPL/exp1.csv", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "train_participants": 232, "train_responses": 52464, "heldout_test_participants": 28, "heldout_test_responses": 6136, "parameter_fits": "Each of three original experiments fitted separately on its training rows", "parameter_selection": "Held-out data not used for parameter selection", "feedback": "Update only after feedback actually visible to the human; retain no-feedback rows" }, "comparison_note": "Same original held-out human-response mode, split and all responses. Disclosed formula fit is distinct from unrecovered author fitting code and archived unpaired scores. Retain each output interface and NLL definition; shared task membership does not make integer, typed or binned likelihoods interchangeable.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "mcpl_full_fit_acceptance_20261028", "field": "/experiments/0/test/mean_nll_response_weighted" }, "n_participants": 10, "same_neural_mode": true }, { "row_id": "mcpl-exp2-test-nll", "reference_id": "multiple_cue_disclosed", "model": "Multiple-cue formula reimplementation", "task": "multiple_cue_judgment", "mode": "heldout_human_history_prediction", "metric": "human_nll", "value": 2.0859402307757233, "value_raw": "2.0859402307757233", "value_status": "finite", "n": 1680, "aggregation": "response_weighted_mean", "unit": "nats_per_response", "scope": "collsiöö2023MCPL/exp2.csv", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "train_participants": 232, "train_responses": 52464, "heldout_test_participants": 28, "heldout_test_responses": 6136, "parameter_fits": "Each of three original experiments fitted separately on its training rows", "parameter_selection": "Held-out data not used for parameter selection", "feedback": "Update only after feedback actually visible to the human; retain no-feedback rows" }, "comparison_note": "Same original held-out human-response mode, split and all responses. Disclosed formula fit is distinct from unrecovered author fitting code and archived unpaired scores. Retain each output interface and NLL definition; shared task membership does not make integer, typed or binned likelihoods interchangeable.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "mcpl_full_fit_acceptance_20261028", "field": "/experiments/1/test/mean_nll_response_weighted" }, "n_participants": 6, "same_neural_mode": true }, { "row_id": "mcpl-exp3-test-nll", "reference_id": "multiple_cue_disclosed", "model": "Multiple-cue formula reimplementation", "task": "multiple_cue_judgment", "mode": "heldout_human_history_prediction", "metric": "human_nll", "value": 2.0229614388098134, "value_raw": "2.0229614388098134", "value_status": "finite", "n": 1656, "aggregation": "response_weighted_mean", "unit": "nats_per_response", "scope": "collsiöö2023MCPL/exp3.csv", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "train_participants": 232, "train_responses": 52464, "heldout_test_participants": 28, "heldout_test_responses": 6136, "parameter_fits": "Each of three original experiments fitted separately on its training rows", "parameter_selection": "Held-out data not used for parameter selection", "feedback": "Update only after feedback actually visible to the human; retain no-feedback rows" }, "comparison_note": "Same original held-out human-response mode, split and all responses. Disclosed formula fit is distinct from unrecovered author fitting code and archived unpaired scores. Retain each output interface and NLL definition; shared task membership does not make integer, typed or binned likelihoods interchangeable.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "mcpl_full_fit_acceptance_20261028", "field": "/experiments/2/test/mean_nll_response_weighted" }, "n_participants": 12, "same_neural_mode": true }, { "row_id": "rlwm-M9-human_nll", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "human_nll", "value": 0.4588994826234648, "value_raw": "0.4588994826234648", "value_status": "finite", "n": 79, "aggregation": "equal_participant_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_overall_20261027", "field": "/author_fitted_9/human_nll" }, "n_responses": 52911, "reference_variant": "M9: WM1 RL1 0", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant interval as saved", "low": 0.4348499806100895, "high": 0.4829489846368401, "low_raw": "0.4348499806100895", "high_raw": "0.4829489846368401" } }, { "row_id": "rlwm-M9-correct_action_probability", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "correct_action_probability", "value": 0.8008611307731598, "value_raw": "0.8008611307731598", "value_status": "finite", "n": 79, "aggregation": "equal_participant_mean", "unit": "probability_0_to_1", "scope": "all", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_overall_20261027", "field": "/author_fitted_9/probability_correct_action" }, "n_responses": 52911, "reference_variant": "M9: WM1 RL1 0", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant interval as saved", "low": 0.7909388198134125, "high": 0.8107834417329071, "low_raw": "0.7909388198134125", "high_raw": "0.8107834417329071" } }, { "row_id": "rlwm-M9-load-contrast", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "correct_action_probability_load_contrast", "value": -0.113117101356202, "value_raw": "-0.113117101356202", "value_status": "finite", "n": 79, "aggregation": "paired_participant_mean_difference_size6_minus_size2", "unit": "probability_difference", "scope": "set_size_6_minus_2", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_load_contrasts_20261027", "field": "/author_fitted_9/set_size_2_6/probability_correct_action_difference" }, "reference_variant": "M9: WM1 RL1 0", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant percentile bootstrap; 5,000 resamples", "low": -0.1271982867107257, "high": -0.0995069883604533, "low_raw": "-0.1271982867107257", "high_raw": "-0.0995069883604533" } }, { "row_id": "rlwm-M6-human_nll", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "human_nll", "value": 0.4490387769244143, "value_raw": "0.4490387769244143", "value_status": "finite", "n": 79, "aggregation": "equal_participant_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_overall_20261027", "field": "/author_fitted_6/human_nll" }, "n_responses": 52911, "reference_variant": "M6: WM=RL 0", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant interval as saved", "low": 0.4263442616348485, "high": 0.4717332922139801, "low_raw": "0.4263442616348485", "high_raw": "0.4717332922139801" } }, { "row_id": "rlwm-M6-correct_action_probability", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "correct_action_probability", "value": 0.8005620940472629, "value_raw": "0.8005620940472629", "value_status": "finite", "n": 79, "aggregation": "equal_participant_mean", "unit": "probability_0_to_1", "scope": "all", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_overall_20261027", "field": "/author_fitted_6/probability_correct_action" }, "n_responses": 52911, "reference_variant": "M6: WM=RL 0", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant interval as saved", "low": 0.7907658641403306, "high": 0.8103583239541952, "low_raw": "0.7907658641403306", "high_raw": "0.8103583239541952" } }, { "row_id": "rlwm-M6-load-contrast", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "correct_action_probability_load_contrast", "value": -0.1079334788546584, "value_raw": "-0.1079334788546584", "value_status": "finite", "n": 79, "aggregation": "paired_participant_mean_difference_size6_minus_size2", "unit": "probability_difference", "scope": "set_size_6_minus_2", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_load_contrasts_20261027", "field": "/author_fitted_6/set_size_2_6/probability_correct_action_difference" }, "reference_variant": "M6: WM=RL 0", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant percentile bootstrap; 5,000 resamples", "low": -0.1216905509333861, "high": -0.0949873075569775, "low_raw": "-0.1216905509333861", "high_raw": "-0.0949873075569775" } }, { "row_id": "rlwm-M2-human_nll", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "human_nll", "value": 0.4478506251677495, "value_raw": "0.4478506251677495", "value_status": "finite", "n": 79, "aggregation": "equal_participant_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_overall_20261027", "field": "/author_fitted_2/human_nll" }, "n_responses": 52911, "reference_variant": "M2: WM RL0 0", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant interval as saved", "low": 0.4252069262142459, "high": 0.4704943241212532, "low_raw": "0.4252069262142459", "high_raw": "0.4704943241212532" } }, { "row_id": "rlwm-M2-correct_action_probability", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "correct_action_probability", "value": 0.8008229721224966, "value_raw": "0.8008229721224966", "value_status": "finite", "n": 79, "aggregation": "equal_participant_mean", "unit": "probability_0_to_1", "scope": "all", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_overall_20261027", "field": "/author_fitted_2/probability_correct_action" }, "n_responses": 52911, "reference_variant": "M2: WM RL0 0", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant interval as saved", "low": 0.7911070556328368, "high": 0.8105388886121565, "low_raw": "0.7911070556328368", "high_raw": "0.8105388886121565" } }, { "row_id": "rlwm-M2-load-contrast", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "correct_action_probability_load_contrast", "value": -0.1121538225540942, "value_raw": "-0.1121538225540942", "value_status": "finite", "n": 79, "aggregation": "paired_participant_mean_difference_size6_minus_size2", "unit": "probability_difference", "scope": "set_size_6_minus_2", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_load_contrasts_20261027", "field": "/author_fitted_2/set_size_2_6/probability_correct_action_difference" }, "reference_variant": "M2: WM RL0 0", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant percentile bootstrap; 5,000 resamples", "low": -0.1259096132984118, "high": -0.0990597443647526, "low_raw": "-0.1259096132984118", "high_raw": "-0.0990597443647526" } }, { "row_id": "rlwm-M8-human_nll", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "human_nll", "value": 0.4515424926816752, "value_raw": "0.4515424926816752", "value_status": "finite", "n": 79, "aggregation": "equal_participant_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_overall_20261027", "field": "/author_fitted_8/human_nll" }, "n_responses": 52911, "reference_variant": "M8: WM1 RL0 0", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant interval as saved", "low": 0.4283535304253389, "high": 0.4747314549380114, "low_raw": "0.4283535304253389", "high_raw": "0.4747314549380114" } }, { "row_id": "rlwm-M8-correct_action_probability", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "correct_action_probability", "value": 0.8013964226069231, "value_raw": "0.8013964226069231", "value_status": "finite", "n": 79, "aggregation": "equal_participant_mean", "unit": "probability_0_to_1", "scope": "all", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_overall_20261027", "field": "/author_fitted_8/probability_correct_action" }, "n_responses": 52911, "reference_variant": "M8: WM1 RL0 0", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant interval as saved", "low": 0.7916630694893303, "high": 0.811129775724516, "low_raw": "0.7916630694893303", "high_raw": "0.811129775724516" } }, { "row_id": "rlwm-M8-load-contrast", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "correct_action_probability_load_contrast", "value": -0.1121696976578527, "value_raw": "-0.1121696976578527", "value_status": "finite", "n": 79, "aggregation": "paired_participant_mean_difference_size6_minus_size2", "unit": "probability_difference", "scope": "set_size_6_minus_2", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_load_contrasts_20261027", "field": "/author_fitted_8/set_size_2_6/probability_correct_action_difference" }, "reference_variant": "M8: WM1 RL0 0", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant percentile bootstrap; 5,000 resamples", "low": -0.1260077106880077, "high": -0.0988420578885478, "low_raw": "-0.1260077106880077", "high_raw": "-0.0988420578885478" } }, { "row_id": "rlwm-M26-human_nll", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "human_nll", "value": 0.446287122933602, "value_raw": "0.446287122933602", "value_status": "finite", "n": 79, "aggregation": "equal_participant_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_overall_20261027", "field": "/author_fitted_26/human_nll" }, "n_responses": 52911, "reference_variant": "M26: freq WM=RL 0", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant interval as saved", "low": 0.4239245585430343, "high": 0.4686496873241698, "low_raw": "0.4239245585430343", "high_raw": "0.4686496873241698" } }, { "row_id": "rlwm-M26-correct_action_probability", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "correct_action_probability", "value": 0.7997900616416654, "value_raw": "0.7997900616416654", "value_status": "finite", "n": 79, "aggregation": "equal_participant_mean", "unit": "probability_0_to_1", "scope": "all", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_overall_20261027", "field": "/author_fitted_26/probability_correct_action" }, "n_responses": 52911, "reference_variant": "M26: freq WM=RL 0", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant interval as saved", "low": 0.789775636502756, "high": 0.8098044867805749, "low_raw": "0.789775636502756", "high_raw": "0.8098044867805749" } }, { "row_id": "rlwm-M26-load-contrast", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "correct_action_probability_load_contrast", "value": -0.1183725235590325, "value_raw": "-0.1183725235590325", "value_status": "finite", "n": 79, "aggregation": "paired_participant_mean_difference_size6_minus_size2", "unit": "probability_difference", "scope": "set_size_6_minus_2", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_load_contrasts_20261027", "field": "/author_fitted_26/set_size_2_6/probability_correct_action_difference" }, "reference_variant": "M26: freq WM=RL 0", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant percentile bootstrap; 5,000 resamples", "low": -0.1331247029160829, "high": -0.1041904769798383, "low_raw": "-0.1331247029160829", "high_raw": "-0.1041904769798383" } }, { "row_id": "rlwm-M38-human_nll", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "human_nll", "value": 0.4472478381301806, "value_raw": "0.4472478381301806", "value_status": "finite", "n": 79, "aggregation": "equal_participant_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_overall_20261027", "field": "/author_fitted_38/human_nll" }, "n_responses": 52911, "reference_variant": "M38: WM1 RL1 r0", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant interval as saved", "low": 0.4245758315826092, "high": 0.469919844677752, "low_raw": "0.4245758315826092", "high_raw": "0.469919844677752" } }, { "row_id": "rlwm-M38-correct_action_probability", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "correct_action_probability", "value": 0.8000421520856016, "value_raw": "0.8000421520856016", "value_status": "finite", "n": 79, "aggregation": "equal_participant_mean", "unit": "probability_0_to_1", "scope": "all", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_overall_20261027", "field": "/author_fitted_38/probability_correct_action" }, "n_responses": 52911, "reference_variant": "M38: WM1 RL1 r0", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant interval as saved", "low": 0.7899031854129458, "high": 0.8101811187582574, "low_raw": "0.7899031854129458", "high_raw": "0.8101811187582574" } }, { "row_id": "rlwm-M38-load-contrast", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "correct_action_probability_load_contrast", "value": -0.120294444984818, "value_raw": "-0.120294444984818", "value_status": "finite", "n": 79, "aggregation": "paired_participant_mean_difference_size6_minus_size2", "unit": "probability_difference", "scope": "set_size_6_minus_2", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_load_contrasts_20261027", "field": "/author_fitted_38/set_size_2_6/probability_correct_action_difference" }, "reference_variant": "M38: WM1 RL1 r0", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant percentile bootstrap; 5,000 resamples", "low": -0.1357290105003483, "high": -0.1055493800380204, "low_raw": "-0.1357290105003483", "high_raw": "-0.1055493800380204" } }, { "row_id": "rlwm-M42-human_nll", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "human_nll", "value": 0.4472334201585796, "value_raw": "0.4472334201585796", "value_status": "finite", "n": 79, "aggregation": "equal_participant_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_overall_20261027", "field": "/author_fitted_42/human_nll" }, "n_responses": 52911, "reference_variant": "M42: WM RL0 0-PC", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant interval as saved", "low": 0.424576273130584, "high": 0.4698905671865752, "low_raw": "0.424576273130584", "high_raw": "0.4698905671865752" } }, { "row_id": "rlwm-M42-correct_action_probability", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "correct_action_probability", "value": 0.8010999267299779, "value_raw": "0.8010999267299779", "value_status": "finite", "n": 79, "aggregation": "equal_participant_mean", "unit": "probability_0_to_1", "scope": "all", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_overall_20261027", "field": "/author_fitted_42/probability_correct_action" }, "n_responses": 52911, "reference_variant": "M42: WM RL0 0-PC", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant interval as saved", "low": 0.791395891917879, "high": 0.8108039615420768, "low_raw": "0.791395891917879", "high_raw": "0.8108039615420768" } }, { "row_id": "rlwm-M42-load-contrast", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "correct_action_probability_load_contrast", "value": -0.1105077497235097, "value_raw": "-0.1105077497235097", "value_status": "finite", "n": 79, "aggregation": "paired_participant_mean_difference_size6_minus_size2", "unit": "probability_difference", "scope": "set_size_6_minus_2", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Same broad offline human-response prediction family, but same-participant fitting and different history exposure. These are neither held-out predictions nor matched-training/identical-history superiority comparisons. Correct-action probability is distinct from an argmax action.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "rlwm_accepted_load_contrasts_20261027", "field": "/author_fitted_42/set_size_2_6/probability_correct_action_difference" }, "reference_variant": "M42: WM RL0 0-PC", "same_neural_mode": true, "identical_history": false, "ci95": { "type": "pointwise participant percentile bootstrap; 5,000 resamples", "low": -0.1242812665586301, "high": -0.0973078430324937, "low_raw": "-0.1242812665586301", "high_raw": "-0.0973078430324937" } }, { "row_id": "wcst-sl-1-objective_accuracy", "reference_id": "wcst_sl", "model": "SL (local refit)", "task": "wcst", "mode": "human_history_prediction", "metric": "objective_accuracy", "value": 0.604, "value_raw": "0.6040", "value_status": "finite", "n": 75, "aggregation": "equal_participant_mean", "unit": "fraction", "scope": "all", "provenance_type": "official-code", "fitting_exposure": { "fit_participants": 300, "test_participants": 75, "trials_per_test_mode": 12754, "split": "First 75 original source participants are test; remaining 300 are training" }, "comparison_note": "Same test participants and task-mode targets; training exposure and recorded inference rules differ from frozen neural checkpoints. Autonomous modes replay human-source rule-switch times. The four-decimal paper table is the source precision shown here.", "exact_precision": false, "precision": "reported_four_decimal_source", "status": "reported_rounded_source", "source": { "record": "wcst_completed_reference_manuscript_table", "field": "/SL_local_refit/column_1" }, "n_trials": 12754, "same_neural_mode": true }, { "row_id": "wcst-sl-2-human_nll", "reference_id": "wcst_sl", "model": "SL (local refit)", "task": "wcst", "mode": "human_history_prediction", "metric": "human_nll", "value": 0.8402, "value_raw": "0.8402", "value_status": "finite", "n": 75, "aggregation": "equal_participant_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "official-code", "fitting_exposure": { "fit_participants": 300, "test_participants": 75, "trials_per_test_mode": 12754, "split": "First 75 original source participants are test; remaining 300 are training" }, "comparison_note": "Same test participants and task-mode targets; training exposure and recorded inference rules differ from frozen neural checkpoints. Autonomous modes replay human-source rule-switch times. The four-decimal paper table is the source precision shown here.", "exact_precision": false, "precision": "reported_four_decimal_source", "status": "reported_rounded_source", "source": { "record": "wcst_completed_reference_manuscript_table", "field": "/SL_local_refit/column_2" }, "n_trials": 12754, "same_neural_mode": true }, { "row_id": "wcst-sl-3-objective_accuracy", "reference_id": "wcst_sl", "model": "SL (local refit)", "task": "wcst", "mode": "autonomous_replay_human_source_switch_schedule", "metric": "objective_accuracy", "value": 0.4892, "value_raw": "0.4892", "value_status": "finite", "n": 75, "aggregation": "equal_participant_mean", "unit": "fraction", "scope": "all", "provenance_type": "official-code", "fitting_exposure": { "fit_participants": 300, "test_participants": 75, "trials_per_test_mode": 12754, "split": "First 75 original source participants are test; remaining 300 are training" }, "comparison_note": "Same test participants and task-mode targets; training exposure and recorded inference rules differ from frozen neural checkpoints. Autonomous modes replay human-source rule-switch times. The four-decimal paper table is the source precision shown here.", "exact_precision": false, "precision": "reported_four_decimal_source", "status": "reported_rounded_source", "source": { "record": "wcst_completed_reference_manuscript_table", "field": "/SL_local_refit/column_3" }, "n_trials": 12754, "same_neural_mode": true }, { "row_id": "wcst-sl-4-perseveration_error", "reference_id": "wcst_sl", "model": "SL (local refit)", "task": "wcst", "mode": "autonomous_replay_human_source_switch_schedule", "metric": "perseveration_error", "value": 0.0947, "value_raw": "0.0947", "value_status": "finite", "n": 75, "aggregation": "equal_participant_mean", "unit": "fraction", "scope": "all", "provenance_type": "official-code", "fitting_exposure": { "fit_participants": 300, "test_participants": 75, "trials_per_test_mode": 12754, "split": "First 75 original source participants are test; remaining 300 are training" }, "comparison_note": "Same test participants and task-mode targets; training exposure and recorded inference rules differ from frozen neural checkpoints. Autonomous modes replay human-source rule-switch times. The four-decimal paper table is the source precision shown here.", "exact_precision": false, "precision": "reported_four_decimal_source", "status": "reported_rounded_source", "source": { "record": "wcst_completed_reference_manuscript_table", "field": "/SL_local_refit/column_4" }, "n_trials": 12754, "same_neural_mode": true }, { "row_id": "wcst-sl-5-set_loss_error", "reference_id": "wcst_sl", "model": "SL (local refit)", "task": "wcst", "mode": "autonomous_replay_human_source_switch_schedule", "metric": "set_loss_error", "value": 0.0463, "value_raw": "0.0463", "value_status": "finite", "n": 75, "aggregation": "equal_participant_mean", "unit": "fraction", "scope": "all", "provenance_type": "official-code", "fitting_exposure": { "fit_participants": 300, "test_participants": 75, "trials_per_test_mode": 12754, "split": "First 75 original source participants are test; remaining 300 are training" }, "comparison_note": "Same test participants and task-mode targets; training exposure and recorded inference rules differ from frozen neural checkpoints. Autonomous modes replay human-source rule-switch times. The four-decimal paper table is the source precision shown here.", "exact_precision": false, "precision": "reported_four_decimal_source", "status": "reported_rounded_source", "source": { "record": "wcst_completed_reference_manuscript_table", "field": "/SL_local_refit/column_5" }, "n_trials": 12754, "same_neural_mode": true }, { "row_id": "mouselab-metagreedy-1", "reference_id": "mouselab_metagreedy", "model": "MetaGreedy (official implementation)", "task": "mouselab_tree", "mode": "within_participant_heldout_human_action_prediction", "metric": "human_nll_total", "value": 27835.4919, "value_raw": "27835.4919", "value_status": "finite", "n": 20363, "aggregation": "sum_over_heldout_decisions", "unit": "nats_total", "scope": "all", "provenance_type": "official-code", "fitting_exposure": { "participants": 95, "heldout_trials": 2375, "heldout_action_distributions": 20363, "fit": "Five-fold cross-validation within each participant, with 100 successful Sobol restarts per fit", "same_participant_exposure": true, "heldout_unit": "Trials/action distributions within participants, not unseen participants" }, "comparison_note": "Matches the human-action prediction mode only. Its human-response NLL is not comparable numerically with the 23,750-episode neural autonomous reveals or returns. Project-local Julia 1.4-compatible dependencies differ from the historical author archive environment.", "exact_precision": false, "precision": "reported_four_decimal_source", "status": "reported_rounded_source", "source": { "record": "mouselab_completed_metagreedy_manuscript_description", "field": "/reported/human_nll_total" }, "n_participants": 95, "n_trials": 2375, "same_neural_mode": true, "not_comparable_with_neural_autonomous_return": true }, { "row_id": "mouselab-metagreedy-2", "reference_id": "mouselab_metagreedy", "model": "MetaGreedy (official implementation)", "task": "mouselab_tree", "mode": "within_participant_heldout_human_action_prediction", "metric": "human_nll", "value": 1.367, "value_raw": "1.3670", "value_status": "finite", "n": 20363, "aggregation": "response_weighted_mean", "unit": "nats_per_decision", "scope": "all", "provenance_type": "official-code", "fitting_exposure": { "participants": 95, "heldout_trials": 2375, "heldout_action_distributions": 20363, "fit": "Five-fold cross-validation within each participant, with 100 successful Sobol restarts per fit", "same_participant_exposure": true, "heldout_unit": "Trials/action distributions within participants, not unseen participants" }, "comparison_note": "Matches the human-action prediction mode only. Its human-response NLL is not comparable numerically with the 23,750-episode neural autonomous reveals or returns. Project-local Julia 1.4-compatible dependencies differ from the historical author archive environment.", "exact_precision": false, "precision": "reported_four_decimal_source", "status": "reported_rounded_source", "source": { "record": "mouselab_completed_metagreedy_manuscript_description", "field": "/reported/human_nll" }, "n_participants": 95, "n_trials": 2375, "same_neural_mode": true, "not_comparable_with_neural_autonomous_return": true }, { "row_id": "probability_coherence_reference_20261030-mean_numeric_absolute_error", "reference_id": "probability_coherence_reference_20261030", "model": "Median report of training participants", "task": "probability_coherence_judgment", "mode": "heldout_human_history_prediction", "metric": "mae", "value": 16.541056910569107, "value_raw": "16.541056910569107", "value_status": "finite", "n": 2460, "n_detail": "2,460 test responses from 15 Psych-101 test participants", "aggregation": "response_weighted_mean", "unit": "probability_judgment_percentage_points_0_to_100", "scope": "all", "provenance_type": "independent-fit", "fitting_exposure": { "fit": "19,740 reports from the 128 Psych-101 training participants of the same source experiments", "evaluation": "15 disjoint test participants; 2,460 responses" }, "comparison_note": "Predicts each test response with the median report of the training participants for the same query in the same experiment.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "values_file": "table2_values_20261030.json (main results table, paper commit 5a0f491)", "record": "runs/probes_20261028/analysis/reference_fill_20261030/probability_coherence/PROBABILITY_COHERENCE_REFERENCE.json", "field": "/value" }, "metric_variant": "mean_numeric_absolute_error", "table2_reference_type": "human fit", "table2_mark": "¶", "table2_metric": "Human MAE (pp)" }, { "row_id": "instrumental_learning_reference_20261030-original_latent_mean_optimal_action_rate", "reference_id": "instrumental_learning_reference_20261030", "model": "Rescorla-Wagner with separate learning rates", "task": "instrumental_learning", "mode": "autonomous_generated_environment", "metric": "accuracy", "value": 0.9489583333333333, "value_raw": "0.9489583333333333", "value_status": "finite", "n": 960, "n_detail": "10 original test runs of 96 trials; 911 correct", "aggregation": "response_weighted_mean", "unit": "fraction", "scope": "all", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "fit": "Learning rates 0.08 and 0.06 and inverse temperature 794.328 selected by mean earned reward on 100 independent training environments; no human data", "evaluation": "The ten original test environments, after selection" }, "comparison_note": "CogBench model with separate learning rates for positive and negative prediction errors. Reads only visible information.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "values_file": "table2_values_20261030.json (main results table, paper commit 5a0f491)", "record": "runs/probes_20261028/analysis/reference_fill_20261030/instrumental_learning/REFERENCE_RESULT.json", "field": "/value" }, "metric_variant": "original_latent_mean_optimal_action_rate", "table2_reference_type": "algorithm", "table2_mark": "¶", "table2_metric": "Better arm (%)" }, { "row_id": "bart_reference_20261030-mean_reward", "reference_id": "bart_reference_20261030", "model": "Fixed pump threshold per balloon letter", "task": "bart", "mode": "autonomous_generated_environment", "metric": "reward", "value": 10.493333333333334, "value_raw": "10.493333333333334", "value_status": "finite", "n": 900, "n_detail": "20 original test runs of 45 balloons", "aggregation": "response_weighted_mean", "unit": "task_specific_reward_units", "scope": "all", "provenance_type": "independent-fit", "fitting_exposure": { "fit": "Thresholds 61, 63 and 66 selected by mean reward on 100 independent training environments; no human data", "evaluation": "The 20 original test runs, after selection" }, "comparison_note": "The original runner reassigns the three maxima to the letters in every run, so these visible-information thresholds stay below the full-information optimum of 14.00, which requires the hidden maxima.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "values_file": "table2_values_20261030.json (main results table, paper commit 5a0f491)", "record": "runs/probes_20261028/analysis/reference_fill_20261030/bart/REFERENCE_RESULT.json", "field": "/value" }, "metric_variant": "mean_reward", "table2_reference_type": "algorithm", "table2_mark": "¶", "table2_metric": "Mean reward" }, { "row_id": "temporal_discounting_reference_20261030-delayed_reward_choice_rate", "reference_id": "temporal_discounting_reference_20261030", "model": "Later-reward rate of Psych-101 participants", "task": "temporal_discounting", "mode": "autonomous_generated_environment", "metric": "preference", "value": 0.41509744894531775, "value_raw": "0.41509744894531775", "value_status": "finite", "n": 13232, "n_detail": "13,232 participants (11,937 training and 1,295 test); 121,662 reward questions", "aggregation": "participant_mean", "unit": "fraction", "scope": "reward_choice_questions", "provenance_type": "human-data-summary", "fitting_exposure": { "fit": "None; descriptive summary of human sessions", "evaluation": "All 13,232 participants on the same instrument" }, "comparison_note": "Applies the realized-question rule of the model rows. Human sessions keep four anomaly items that the CogBench path skips after three immediate choices; dropping them gives 39.8%.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "values_file": "table2_values_20261030.json (main results table, paper commit 5a0f491)", "record": "runs/probes_20261028/analysis/reference_fill_20261030/temporal_discounting/TEMPORAL_DISCOUNTING_REFERENCE.json", "field": "/value" }, "metric_variant": "delayed_reward_choice_rate", "table2_reference_type": "human", "table2_mark": "¶", "table2_metric": "Delayed reward (%)" }, { "row_id": "two_step_reference_20261030-mean_reward", "reference_id": "two_step_reference_20261030", "model": "Hybrid model fits to CogBench participants (Daw et al. 2011)", "task": "two_step", "mode": "autonomous_generated_environment", "metric": "reward", "value": 0.51955, "value_raw": "0.51955", "value_status": "finite", "n": 2000, "n_detail": "100 original test runs of 20 trials for each of 40 parameter sets; 80,000 simulated trials", "aggregation": "mean_over_40_fitted_parameter_sets", "unit": "task_specific_reward_units", "scope": "all", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "fit": "Seven parameters by maximum likelihood on each of the 40 CogBench human participants", "evaluation": "Identical seeded reward-rate path of the five models" }, "comparison_note": "Twenty-five of the 40 fits have at least one parameter at a bound, so the parameters are a reference policy, not estimates of individual learning.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "values_file": "table2_values_20261030.json (main results table, paper commit 5a0f491)", "record": "runs/probes_20261028/analysis/reference_fill_20261030/two_step/REFERENCE_RESULT.json", "field": "/value" }, "metric_variant": "mean_reward", "table2_reference_type": "human fit", "table2_mark": "¶", "table2_metric": "Mean reward" }, { "row_id": "politeness_reference_20261030-human_response_weighted_mean_nll", "reference_id": "politeness_reference_20261030", "model": "Other participants, leave one out", "task": "politeness_pragmatics", "mode": "aggregate_human_distribution_comparison", "metric": "aggregate_distribution_nll", "value": 0.5872864995872029, "value_raw": "0.5872864995872029", "value_status": "finite", "n": 128, "n_detail": "128 trials; 21,056 human responses", "aggregation": "response_weighted_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "human-data-summary", "fitting_exposure": { "fit": "None; the other participants' responses to the same item, add-one smoothing", "evaluation": "Each released human response" }, "comparison_note": "Predicts each human response from the other participants' responses to the same item.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "values_file": "table2_values_20261030.json (main results table, paper commit 5a0f491)", "record": "runs/probes_20261028/analysis/reference_fill_20261030/politeness/politeness_reference_20261030.json", "field": "/value" }, "metric_variant": "human_response_weighted_mean_nll", "table2_reference_type": "human", "table2_mark": "¶", "table2_metric": "Aggregate NLL" }, { "row_id": "pragmatic_interpretation_reference_20261030-human_response_weighted_mean_nll", "reference_id": "pragmatic_interpretation_reference_20261030", "model": "Other participants, leave one out", "task": "pragmatic_language_interpretation", "mode": "aggregate_human_distribution_comparison", "metric": "aggregate_distribution_nll", "value": 0.41570917438542526, "value_raw": "0.41570917438542526", "value_status": "finite", "n": 170, "n_detail": "170 trials; 63,580 human responses", "aggregation": "response_weighted_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "human-data-summary", "fitting_exposure": { "fit": "None; the other participants' responses to the same item, add-one smoothing", "evaluation": "Each released human response" }, "comparison_note": "Predicts each human response from the other participants' responses to the same item.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "values_file": "table2_values_20261030.json (main results table, paper commit 5a0f491)", "record": "runs/probes_20261028/analysis/reference_fill_20261030/pragmatic_interpretation/pragmatic_interpretation_reference_20261030.json", "field": "/value" }, "metric_variant": "human_response_weighted_mean_nll", "table2_reference_type": "human", "table2_mark": "¶", "table2_metric": "Aggregate NLL" }, { "row_id": "nonliteral_number_reference_20261030-absolute_error_to_human_mean_mean", "reference_id": "nonliteral_number_reference_20261030", "model": "Human split-half", "task": "nonliteral_number_interpretation", "mode": "aggregate_human_distribution_comparison", "metric": "mae", "value": 4.976033702069305, "value_raw": "4.976033702069305", "value_status": "finite", "n": 330, "n_detail": "330 queries; 1,000 seeded splits, SD 0.22", "aggregation": "response_weighted_mean", "unit": "reported_percentage_points_0_to_100", "scope": "all", "provenance_type": "human-data-summary", "fitting_exposure": { "fit": "None; mean reports of two random halves of each item's respondents", "evaluation": "All 330 queries" }, "comparison_note": "A half-panel mean lies about twice as far from the other half as from the full-panel mean, so this value is a conservative human reference.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "values_file": "table2_values_20261030.json (main results table, paper commit 5a0f491)", "record": "runs/probes_20261028/analysis/reference_fill_20261030/nonliteral_number/nonliteral_number_reference_20261030.json", "field": "/value" }, "metric_variant": "absolute_error_to_human_mean_mean", "table2_reference_type": "human", "table2_mark": "¶", "table2_metric": "MAE to human mean" }, { "row_id": "self_assessment_lookup_reference_20261030-i_within_logic_10fold_crossfit", "reference_id": "self_assessment_lookup_reference_20261030", "model": "Lookup table", "task": "performance_self_assessment", "mode": "heldout_human_history_prediction", "metric": "human_nll", "value": 1.6193064420134906, "value_raw": "1.6193064420134906", "value_status": "finite", "n": 108248, "n_detail": "9,044 grammar and 99,204 logic responses", "aggregation": "response_weighted_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "fit": "Response-frequency tables with add-alpha smoothing. Grammar tables on 3,192 training sessions; logic by tenfold cross-fitting across logic participants", "evaluation": "323 grammar test sessions; held-out logic folds" }, "comparison_note": "Lookup-table family of Binz et al. (2025). The source defines no baseline for the logic set. Transferring the grammar tables to logic instead gives 1.917.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "values_file": "table2_values_20261030.json (main results table, paper commit 5a0f491)", "record": "runs/probes_20261028/analysis/reference_fill_20261030/self_assessment/SELF_ASSESSMENT_REFERENCE.json", "field": "/variants/i_within_logic_10fold_crossfit/value" }, "metric_variant": "i_within_logic_10fold_crossfit", "table2_reference_type": "human fit", "table2_mark": "¶", "table2_metric": "Human NLL" }, { "row_id": "uniform-bayesian_urn_inference", "reference_id": "uniform_choice", "model": "Uniform", "task": "bayesian_urn_inference", "mode": "generated_explicit_posterior", "protocol": "cogbench_generated_posterior_200x2", "target": "analytic_Bayes_posterior", "metric": "mae", "value": 0.34946797528825263, "value_raw": "0.34946797528825263", "value_status": "finite", "n": 400, "aggregation": "response_weighted_mean", "unit": "probability_0_to_1", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Continuous uniform report on [0, 1]. The expected absolute error for posterior p is (p^2 + (1 - p)^2) / 2, averaged over the 400 posterior targets.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/0/uniform" }, "metric_variant": "posterior_mae", "table2_column": "Uniform", "mode_key": "bayesian_urn_inference", "reproduced_by_uniform_cli_run": true }, { "row_id": "uniform-multiple_cue_judgment", "reference_id": "uniform_choice", "model": "Uniform", "task": "multiple_cue_judgment", "mode": "heldout_human_history_prediction", "protocol": "compact_full_history", "target": "observed_human_response", "metric": "human_nll", "value": 2.1972245773362196, "value_raw": "2.1972245773362196", "value_status": "finite", "n": 6136, "aggregation": "response_weighted_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Uniform over 9 legal responses: log 9.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/1/uniform" }, "metric_variant": "typed_choice_mean_nll", "table2_column": "Uniform", "mode_key": "multiple_cue_judgment", "reproduced_by_uniform_cli_run": true }, { "row_id": "uniform-multi_attribute_choice", "reference_id": "uniform_choice", "model": "Uniform", "task": "multi_attribute_choice", "mode": "heldout_human_history_prediction", "protocol": "original_full_prefix_typed", "target": "observed_human_response", "metric": "human_nll", "value": 0.6931471805599453, "value_raw": "0.6931471805599453", "value_status": "finite", "n": 576, "aggregation": "response_weighted_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Uniform over 2 legal responses: log 2.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/2/uniform" }, "metric_variant": "typed_choice_mean_nll", "table2_column": "Uniform", "mode_key": "multi_attribute_choice", "reproduced_by_uniform_cli_run": true }, { "row_id": "uniform-probability_coherence_judgment", "reference_id": "uniform_choice", "model": "Uniform", "task": "probability_coherence_judgment", "mode": "heldout_human_history_prediction", "protocol": "full-original-prefix-integer-and-self-assessment-v1", "target": "original_human_numeric_response", "metric": "mae", "value": 32.13385253159464, "value_raw": "32.13385253159464", "value_status": "finite", "n": 2460, "aggregation": "response_weighted_mean", "unit": "probability_judgment_percentage_points_0_to_100", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Absolute loss averaged over all 101 legal integer reports for each of the 2,460 saved human targets.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/3/uniform" }, "metric_variant": "mean_numeric_absolute_error", "table2_column": "Uniform", "mode_key": "probability_coherence_judgment", "reproduced_by_uniform_cli_run": true }, { "row_id": "uniform-shj_category_learning", "reference_id": "uniform_choice", "model": "Uniform", "task": "shj_category_learning", "mode": "heldout_human_history_prediction", "protocol": "original_full_prefix_typed", "target": "original_objective_answer_or_optimal_action_label", "metric": "accuracy", "value": 0.5, "value_raw": "0.5", "value_status": "finite", "n": 3664, "aggregation": "response_weighted_mean", "unit": "fraction", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Uniform over 2 legal responses: 1/2.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/4/uniform" }, "metric_variant": "category_accuracy_under_human_history", "table2_column": "Uniform", "mode_key": "shj_category_learning", "reproduced_by_uniform_cli_run": true }, { "row_id": "uniform-medin_category_learning", "reference_id": "uniform_choice", "model": "Uniform", "task": "medin_category_learning", "mode": "heldout_human_history_prediction", "protocol": "original_full_prefix_typed", "target": "original_objective_answer_or_optimal_action_label", "metric": "accuracy", "value": 0.5, "value_raw": "0.5", "value_status": "finite", "n": 6300, "aggregation": "response_weighted_mean", "unit": "fraction", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Uniform over 2 legal responses: 1/2.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/5/uniform" }, "metric_variant": "objective_label_accuracy_under_human_history", "table2_column": "Uniform", "mode_key": "medin_category_learning", "reproduced_by_uniform_cli_run": true }, { "row_id": "uniform-weather_prediction", "reference_id": "uniform_choice", "model": "Uniform", "task": "weather_prediction", "mode": "heldout_human_history_prediction", "protocol": "original_full_prefix_typed", "target": "original_objective_answer_or_optimal_action_label", "metric": "accuracy", "value": 0.5, "value_raw": "0.5", "value_status": "finite", "n": 400, "aggregation": "response_weighted_mean", "unit": "fraction", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Uniform over 2 legal responses: 1/2.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/6/uniform" }, "metric_variant": "objective_label_accuracy_under_human_history", "table2_column": "Uniform", "mode_key": "weather_prediction", "reproduced_by_uniform_cli_run": true }, { "row_id": "uniform-rlwm", "reference_id": "uniform_choice", "model": "Uniform", "task": "rlwm", "mode": "original_human_history_replay_no_official_heldout_split", "protocol": "original_full_prefix_typed", "target": "original_objective_answer_or_optimal_action_label", "metric": "accuracy", "value": 0.3333333333333333, "value_raw": "0.3333333333333333", "value_status": "finite", "n": 52911, "aggregation": "response_weighted_mean", "unit": "fraction", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Uniform over 3 legal responses: 1/3.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/7/uniform" }, "metric_variant": "objective_label_accuracy_under_human_history", "table2_column": "Uniform", "mode_key": "rlwm", "reproduced_by_uniform_cli_run": true }, { "row_id": "uniform-recent_probes", "reference_id": "uniform_choice", "model": "Uniform", "task": "recent_probes", "mode": "heldout_human_history_prediction", "protocol": "original_full_prefix_typed", "target": "original_objective_answer_or_optimal_action_label", "metric": "accuracy", "value": 0.5, "value_raw": "0.5", "value_status": "finite", "n": 3774, "aggregation": "response_weighted_mean", "unit": "fraction", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Uniform over 2 legal responses: 1/2.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/8/uniform" }, "metric_variant": "objective_label_accuracy_under_human_history", "table2_column": "Uniform", "mode_key": "recent_probes", "reproduced_by_uniform_cli_run": true }, { "row_id": "uniform-wcst-human_predictive", "reference_id": "uniform_choice", "model": "Uniform", "task": "wcst", "mode": "predictive", "protocol": "wcst_predictive_75_test_participants", "target": "observed_human_response", "metric": "human_nll", "value": 1.3862943611198906, "value_raw": "1.3862943611198906", "value_status": "finite", "n": 75, "aggregation": "participant_mean", "unit": "nats_per_response", "scope": "predictive", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Uniform over 4 legal responses: log 4.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/9/uniform" }, "metric_variant": "participant_mean", "table2_column": "Uniform", "mode_key": "wcst:human_predictive", "reproduced_by_uniform_cli_run": true }, { "row_id": "uniform-wcst-autonomous_generative", "reference_id": "uniform_choice", "model": "Uniform", "task": "wcst", "mode": "generative", "protocol": "wcst_generative_75_test_participants", "target": "original_rule_correct_card", "metric": "accuracy", "value": 0.25, "value_raw": "0.25", "value_status": "finite", "n": 75, "aggregation": "participant_mean", "unit": "fraction", "scope": "generative", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Expected accuracy of uniform card choices: 1/4.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/10/uniform" }, "metric_variant": "participant_mean", "table2_column": "Uniform", "mode_key": "wcst:autonomous_generative", "reproduced_by_uniform_cli_run": false }, { "row_id": "uniform-nonstationary_bandit", "reference_id": "uniform_choice", "model": "Uniform", "task": "nonstationary_bandit", "mode": "autonomous_generated_environment", "protocol": "cogbench_restless_action_and_confidence_10x80", "target": "original_objective_answer_or_optimal_action_label", "metric": "accuracy", "value": 0.5, "value_raw": "0.5", "value_status": "finite", "n": 800, "aggregation": "mean_of_10_equal_length_80_trial_run_scores", "unit": "fraction", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Expected better-arm rate of uniform choices between 2 arms: 1/2.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/11/uniform" }, "metric_variant": "accuracy_post_step_original", "table2_column": "Uniform", "mode_key": "nonstationary_bandit", "reproduced_by_uniform_cli_run": false }, { "row_id": "uniform-instrumental_learning", "reference_id": "uniform_choice", "model": "Uniform", "task": "instrumental_learning", "mode": "autonomous_generated_environment", "protocol": "original_environment", "target": "maximal_original_latent_arm_mean", "metric": "accuracy", "value": 0.75, "value_raw": "0.75", "value_status": "finite", "n": 960, "aggregation": "response_weighted_mean", "unit": "fraction", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Credits both actions on tied-mean trials, over the original trial-weighted scorer.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/12/uniform" }, "metric_variant": "original_latent_mean_optimal_action_rate", "table2_column": "Uniform", "mode_key": "instrumental_learning", "reproduced_by_uniform_cli_run": false }, { "row_id": "uniform-horizon-human_predictive", "reference_id": "uniform_choice", "model": "Uniform", "task": "horizon", "mode": "heldout_human_history_prediction", "protocol": "namazova_human_predictive_6x1120", "target": "observed_human_response", "metric": "human_nll", "value": 0.6931471805599453, "value_raw": "0.6931471805599453", "value_status": "finite", "n": 6720, "aggregation": "response_weighted_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Uniform over 2 legal responses: log 2.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/13/uniform" }, "metric_variant": "legal_nll_eps1e8", "table2_column": "Uniform", "mode_key": "horizon:human_predictive", "reproduced_by_uniform_cli_run": true }, { "row_id": "uniform-horizon-autonomous_generative", "reference_id": "uniform_choice", "model": "Uniform", "task": "horizon", "mode": "autonomous_generative", "protocol": "namazova_generative_31_timelines_uniform_expectation", "target": "latent_mean_optimal_arm", "metric": "accuracy", "value": 0.5, "value_raw": "0.5", "value_status": "finite", "n": 34720, "aggregation": "response_weighted_mean", "unit": "fraction", "scope": "free_choices", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Expected optimal rate of uniform choices on the 34,720 original free choices. The two arm means differ on every one.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/14/uniform" }, "metric_variant": "optimal_action_rate", "table2_column": "Uniform", "mode_key": "horizon:autonomous_generative", "reproduced_by_uniform_cli_run": false }, { "row_id": "uniform-stationary_multiarmed_bandit", "reference_id": "uniform_choice", "model": "Uniform", "task": "stationary_multiarmed_bandit", "mode": "autonomous", "protocol": "official_code_fullbench_summary_history_16_conditions", "target": "maximal_original_latent_arm_mean", "metric": "accuracy", "value": 0.125, "value_raw": "0.125", "value_status": "finite", "n": 312000, "aggregation": "response_weighted_mean", "unit": "proportion", "scope": "all_16_original_conditions", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Tied-maximal-arm probability under uniform actions, weighted over all 312,000 trials and 16 original conditions.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/15/uniform" }, "metric_variant": "accuracy", "table2_column": "Uniform", "mode_key": "stationary_multiarmed_bandit", "reproduced_by_uniform_cli_run": false }, { "row_id": "uniform-hybrid_information_sampling", "reference_id": "uniform_choice", "model": "Uniform", "task": "hybrid_information_sampling", "mode": "heldout_human_history_prediction", "protocol": "original_20_subjects_8fold_frozen_prediction", "target": "observed_human_response", "metric": "human_nll", "value": 1.3862943611198906, "value_raw": "1.3862943611198906", "value_status": "finite", "n": 163135, "aggregation": "response_weighted_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Uniform over 4 legal responses: log 4.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/16/uniform" }, "metric_variant": "pooled_action_nll", "table2_column": "Uniform", "mode_key": "hybrid_information_sampling", "reproduced_by_uniform_cli_run": true }, { "row_id": "uniform-mouselab_tree-human_predictive", "reference_id": "uniform_choice", "model": "Uniform", "task": "mouselab_tree", "mode": "heldout_human_history_prediction", "protocol": "original_human_predictive_95x25", "target": "observed_human_response", "metric": "human_nll", "value": 1.6336348460418304, "value_raw": "1.6336348460418304", "value_status": "finite", "n": 20363, "aggregation": "response_weighted_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Log of the legal-option count, averaged over the 20,363 original held-out decisions.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/17/uniform" }, "metric_variant": "mean_nll", "table2_column": "Uniform", "mode_key": "mouselab_tree:human_predictive", "reproduced_by_uniform_cli_run": true }, { "row_id": "uniform-mouselab_tree-autonomous_generative", "reference_id": "uniform_choice", "model": "Uniform", "task": "mouselab_tree", "mode": "autonomous_generative", "protocol": "original_metamdp_generative_23750_uniform_expectation", "target": "best_path_expected_value_at_STOP", "metric": "reward", "value": 7.007757929108834, "value_raw": "7.007757929108834", "value_status": "finite", "n": 23750, "aggregation": "episode_mean", "unit": "task_specific_reward_units", "scope": "generating_distribution_expectation", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Exact expectation over the 1,296 reachable known-node sets under uniform legal reveal and STOP choices, with the original independent four-point node prior and zero observation cost. A generating-distribution expectation, not a score on the models' sampled trees. No Monte Carlo rollouts.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/18/uniform" }, "metric_variant": "mean_terminal_expected_reward", "table2_column": "Uniform", "mode_key": "mouselab_tree:autonomous_generative", "reproduced_by_uniform_cli_run": false }, { "row_id": "uniform-bart", "reference_id": "uniform_choice", "model": "Uniform", "task": "bart", "mode": "autonomous_generated_environment", "protocol": "original_environment", "target": "autonomous_environment_reward", "metric": "reward", "value": 0.8377278646682195, "value_raw": "0.8377278646682195", "value_status": "finite", "n": 900, "aggregation": "response_weighted_mean", "unit": "task_specific_reward_units", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Exact expected reward of independent 1/2 skip and inflate decisions under the original hazard 1/(M - successful pumps), averaged over M = 128, 32 and 8.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/19/uniform" }, "metric_variant": "mean_reward", "table2_column": "Uniform", "mode_key": "bart", "reproduced_by_uniform_cli_run": false }, { "row_id": "uniform-temporal_discounting", "reference_id": "uniform_choice", "model": "Uniform", "task": "temporal_discounting", "mode": "autonomous_generated_environment", "protocol": "original_environment", "target": "later_receipt_among_reward_choice_questions", "metric": "preference", "value": 0.5, "value_raw": "0.5", "value_status": "finite", "n": null, "aggregation": "response_weighted_mean", "unit": "fraction", "scope": "reward_choice_questions", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Equal probability of either legal option on each realized reward question. A preference summary, not an accuracy.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/20/uniform" }, "metric_variant": "delayed_reward_choice_rate", "table2_column": "Uniform", "mode_key": "temporal_discounting", "reproduced_by_uniform_cli_run": false }, { "row_id": "uniform-two_step", "reference_id": "uniform_choice", "model": "Uniform", "task": "two_step", "mode": "autonomous_generated_environment", "protocol": "original_CogBench_v1_environment_with_new_typed_choice_adapter", "target": "autonomous_environment_reward", "metric": "reward", "value": 0.49628932868351744, "value_raw": "0.49628932868351744", "value_status": "finite", "n": 2000, "aggregation": "response_weighted_mean", "unit": "task_specific_reward_units", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Uniform first-stage choices put equal mass on both states, and uniform second-stage choices yield the mean of the four current reward rates, on the identical 2,000-trial seeded rate schedule of the five open models.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/21/uniform" }, "metric_variant": "mean_reward", "table2_column": "Uniform", "mode_key": "two_step", "reproduced_by_uniform_cli_run": false }, { "row_id": "uniform-false_belief_story_qa", "reference_id": "uniform_choice", "model": "Uniform", "task": "false_belief_story_qa", "mode": "public_scenario_qa", "protocol": "social-public-full-materials-typed-v1", "target": "original_objective_answer_or_optimal_action_label", "metric": "accuracy", "value": 0.5, "value_raw": "0.5", "value_status": "finite", "n": 5994, "aggregation": "response_weighted_mean", "unit": "fraction", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Uniform over the original 2 visible-container choices: 1/2.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/22/uniform" }, "metric_variant": "original_answer_exact_match_mean", "table2_column": "Uniform", "mode_key": "false_belief_story_qa", "reproduced_by_uniform_cli_run": true }, { "row_id": "uniform-politeness_pragmatics", "reference_id": "uniform_choice", "model": "Uniform", "task": "politeness_pragmatics", "mode": "aggregate_human_distribution_comparison", "protocol": "social-public-full-materials-typed-v1", "target": "public_aggregate_human_distribution_not_individual_holdout_response", "metric": "aggregate_distribution_nll", "value": 1.0139120992233244, "value_raw": "1.0139120992233244", "value_status": "finite", "n": 128, "aggregation": "response_weighted_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Log of the legal-option count, weighted by original human response counts.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/23/uniform" }, "metric_variant": "human_response_weighted_mean_nll", "table2_column": "Uniform", "mode_key": "politeness_pragmatics", "reproduced_by_uniform_cli_run": true }, { "row_id": "uniform-pragmatic_language_interpretation", "reference_id": "uniform_choice", "model": "Uniform", "task": "pragmatic_language_interpretation", "mode": "aggregate_human_distribution_comparison", "protocol": "social-public-full-materials-typed-v1", "target": "public_aggregate_human_distribution_not_individual_holdout_response", "metric": "aggregate_distribution_nll", "value": 1.2822683175124885, "value_raw": "1.2822683175124885", "value_status": "finite", "n": 170, "aggregation": "response_weighted_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Log of the legal-option count, weighted by original human response counts.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/24/uniform" }, "metric_variant": "human_response_weighted_mean_nll", "table2_column": "Uniform", "mode_key": "pragmatic_language_interpretation", "reproduced_by_uniform_cli_run": true }, { "row_id": "uniform-nonliteral_number_interpretation", "reference_id": "uniform_choice", "model": "Uniform", "task": "nonliteral_number_interpretation", "mode": "aggregate_human_distribution_comparison", "protocol": "social-public-full-materials-typed-v1", "target": "public_human_mean_slider_judgment", "metric": "mae", "value": 34.76903931818182, "value_raw": "34.76903931818182", "value_status": "finite", "n": 330, "aggregation": "response_weighted_mean", "unit": "reported_percentage_points_0_to_100", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Continuous uniform report on [0, 100]. The expected absolute error for human mean h is (h^2 + (100 - h)^2) / 200, averaged over the 330 human means.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/25/uniform" }, "metric_variant": "absolute_error_to_human_mean_mean", "table2_column": "Uniform", "mode_key": "nonliteral_number_interpretation", "reproduced_by_uniform_cli_run": true }, { "row_id": "uniform-performance_self_assessment", "reference_id": "uniform_choice", "model": "Uniform", "task": "performance_self_assessment", "mode": "heldout_human_history_prediction", "protocol": "full-original-prefix-integer-and-self-assessment-v1", "target": "observed_human_response", "metric": "human_nll", "value": 2.039272330321602, "value_raw": "2.039272330321602", "value_status": "finite", "n": 108248, "aggregation": "response_weighted_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Pooled strict NLL on each question's complete support: 5 answers, 21 count reports, 101 percentile reports and 11 difficulty reports. Not the NLL of an eleven-bin display.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/26/uniform" }, "metric_variant": "mean_typed_nll", "table2_column": "Uniform", "mode_key": "performance_self_assessment", "reproduced_by_uniform_cli_run": true }, { "row_id": "multi_attribute_wadd-typed_choice_mean_nll", "reference_id": "multi_attribute_wadd", "model": "WADD", "task": "multi_attribute_choice", "mode": "heldout_human_history_prediction", "metric": "human_nll", "value": 0.19221569848548428, "value_raw": "0.19221569848548428", "value_status": "finite", "n": 576, "n_detail": "576 responses from six held-out participants", "aggregation": "response_weighted_mean", "unit": "nats_per_response", "scope": "all", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "fit": "Four-weight paper-form reimplementation; 7,008 responses from 73 training participants", "evaluation": "Six held-out participants" }, "comparison_note": "Follows the published model family of Binz et al. (2025). An existing fitted output, aligned to the current target with no new model call.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_revision_b_20261029/EXISTING_REFERENCE_ROW_ALIGNMENT_20261029.json", "field": "/rows/0/value" }, "metric_variant": "typed_choice_mean_nll", "table2_reference_type": "human fit", "table2_mark": "†", "table2_metric": "Human NLL" }, { "row_id": "shj_gcm-category_accuracy_under_human_history", "reference_id": "shj_gcm", "model": "GCM", "task": "shj_category_learning", "mode": "heldout_human_history_prediction", "metric": "accuracy", "value": 0.9383187772925764, "value_raw": "0.9383187772925764", "value_status": "finite", "n": 3664, "n_detail": "3,664 responses from ten held-out participants", "aggregation": "response_weighted_mean", "unit": "fraction", "scope": "all", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "fit": "Paper-form reimplementation; 29,776 responses from 85 training participants", "evaluation": "Ten held-out participants" }, "comparison_note": "Follows the published model family of Binz et al. (2025). Compares each saved legal argmax to the original objective label under human history; not a human-likelihood or autonomous-reward score.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_revision_b_20261029/EXISTING_REFERENCE_ROW_ALIGNMENT_20261029.json", "field": "/rows/1/value" }, "metric_variant": "category_accuracy_under_human_history", "table2_reference_type": "human fit", "table2_mark": "†", "table2_metric": "Accuracy (%)" }, { "row_id": "medin_gcm-objective_label_accuracy_under_human_history", "reference_id": "medin_gcm", "model": "GCM", "task": "medin_category_learning", "mode": "heldout_human_history_prediction", "metric": "accuracy", "value": 0.9661904761904762, "value_raw": "0.9661904761904762", "value_status": "finite", "n": 6300, "n_detail": "6,300 learning classifications from 42 held-out participant sessions", "aggregation": "response_weighted_mean", "unit": "fraction", "scope": "all", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "fit": "Train-only refit on 34,200 learning responses across the two source experiments", "evaluation": "42 held-out participant sessions" }, "comparison_note": "Uses the local train-only refit, not the separately archived author-parameter replay. The separate 336 no-feedback classifications and 336 typicality reports are not counted, and the full 6,972-response human-likelihood task is not claimed complete.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_revision_b_20261029/EXISTING_REFERENCE_ROW_ALIGNMENT_20261029.json", "field": "/rows/2/value" }, "metric_variant": "objective_label_accuracy_under_human_history", "table2_reference_type": "human fit", "table2_mark": "†", "table2_metric": "Accuracy (%)" }, { "row_id": "weather_gcm-objective_label_accuracy_under_human_history", "reference_id": "weather_gcm", "model": "GCM", "task": "weather_prediction", "mode": "heldout_human_history_prediction", "metric": "accuracy", "value": 0.7825, "value_raw": "0.7825", "value_status": "finite", "n": 400, "n_detail": "400 responses from two held-out participants", "aggregation": "response_weighted_mean", "unit": "fraction", "scope": "all", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "fit": "Paper-form reimplementation; 4,600 responses from 23 training participants", "evaluation": "Two held-out participants" }, "comparison_note": "Follows the published model family of Binz et al. (2025). Compares each saved legal argmax to the original objective label under human history.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_revision_b_20261029/EXISTING_REFERENCE_ROW_ALIGNMENT_20261029.json", "field": "/rows/3/value" }, "metric_variant": "objective_label_accuracy_under_human_history", "table2_reference_type": "human fit", "table2_mark": "†", "table2_metric": "Accuracy (%)" }, { "row_id": "recent_probes_rational-objective_label_accuracy_under_human_history", "reference_id": "recent_probes_rational", "model": "Rational conditional-response model", "task": "recent_probes", "mode": "heldout_human_history_prediction", "metric": "accuracy", "value": 1.0, "value_raw": "1.0", "value_status": "finite", "n": 3774, "n_detail": "3,774 responses from 51 held-out participants", "aggregation": "response_weighted_mean", "unit": "fraction", "scope": "all", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "fit": "Conditional-response matrix fitted on 34,714 responses from 471 training participants", "evaluation": "51 held-out participants" }, "comparison_note": "The prompt supplies the current letter set, and the fitted model selects the correct visible-set membership category on every item. This does not demonstrate an internal working-memory mechanism.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_revision_b_20261029/EXISTING_REFERENCE_ROW_ALIGNMENT_20261029.json", "field": "/rows/4/value" }, "metric_variant": "objective_label_accuracy_under_human_history", "table2_reference_type": "human fit", "table2_mark": "†", "table2_metric": "Accuracy (%)" }, { "row_id": "stationary_thompson_family-accuracy", "reference_id": "stationary_thompson_family", "model": "Thompson family", "task": "stationary_multiarmed_bandit", "mode": "autonomous", "metric": "accuracy", "value": 0.6809038461538461, "value_raw": "0.6809038461538461", "value_status": "finite", "n": 312000, "n_detail": "312,000 steps; 16 conditions, 30 seeds per condition", "aggregation": "response_weighted_mean", "unit": "proportion", "scope": "all_16_original_conditions", "provenance_type": "official-code-and-disclosed-reimplementation", "fitting_exposure": { "fit": "None; the policies observe only their own past selected arms and rewards", "evaluation": "All 16 original conditions" }, "comparison_note": "Official Beta(1,1) Thompson on eight Bernoulli conditions and disclosed normal-inverse-gamma Thompson on eight Gaussian conditions. Same original conditions, horizons and environment seeds, with separate logged action RNGs. Step-weighted; no reward units are pooled.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_revision_b_20261029/EXISTING_BANDIT_REFERENCE_TARGET_ALIGNMENT_20261029.json", "field": "/value" }, "metric_variant": "accuracy", "table2_reference_type": "algorithm", "table2_mark": "†", "table2_metric": "Optimal arm (%)" }, { "row_id": "restless_qlearning-accuracy_post_step_original", "reference_id": "restless_qlearning", "model": "Two-parameter Q-learning", "task": "nonstationary_bandit", "mode": "autonomous_generated_environment", "metric": "accuracy", "value": 0.6875, "value_raw": "0.6875", "value_status": "finite", "n": 800, "n_detail": "10 original test runs of 80 trials; 550 correct", "aggregation": "mean_of_10_equal_length_80_trial_run_scores", "unit": "fraction", "scope": "all", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "fit": "Alpha 0.44 and beta 100 selected by mean earned reward on 100 independent 80-trial environments; no human data", "evaluation": "The ten original test environments, after selection" }, "comparison_note": "Original choice equations in the retained CogBench environment. Mean reward 54.6679. No confidence mapping is fitted, and softmax choice probabilities are not read as confidence.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/authorized_reference_and_matched_20261029/restless_qlearning/SUMMARY.json", "field": "/accuracy_original" }, "metric_variant": "accuracy_post_step_original", "table2_reference_type": "algorithm", "table2_mark": "∗", "table2_metric": "Better arm (%)" }, { "row_id": "mouselab_metagreedy_autonomous-mean_terminal_expected_reward", "reference_id": "mouselab_metagreedy_autonomous", "model": "MetaGreedy, archived fitted parameters", "task": "mouselab_tree", "mode": "autonomous_generative", "metric": "reward", "value": 11.77221052631579, "value_raw": "11.77221052631579", "value_status": "finite", "n": 23750, "n_detail": "23,750 archived synthetic trees; 164,765 selected actions", "aggregation": "episode_mean", "unit": "task_specific_reward_units", "scope": "all_23750_archived_trees", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "fit": "All 95 archived participant-specific fitted parameter sets; no fitting on the generated evaluation trees", "evaluation": "The same trees as the five open models" }, "comparison_note": "Python port checked against all 20,363 saved Julia probability vectors. The sampler matches the original distribution, not Julia's random stream. Distinct from the archived human-prediction NLL of 1.3670.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/authorized_reference_and_matched_20261029/mouselab_metagreedy/SUMMARY.json", "field": "/mean_terminal_expected_reward" }, "metric_variant": "mean_terminal_expected_reward", "table2_reference_type": "human fit", "table2_mark": "∗", "table2_metric": "Expected return" }, { "row_id": "bayesian_urn_analytic-posterior_mae", "reference_id": "bayesian_urn_analytic", "model": "Exact Bayes", "task": "bayesian_urn_inference", "mode": "generated_explicit_posterior", "metric": "mae", "value": 0.0, "value_raw": "0.0", "value_status": "finite", "n": 400, "n_detail": "The 400 generated queries", "aggregation": "response_weighted_mean", "unit": "probability_0_to_1", "scope": "all", "provenance_type": "analytic-solution", "fitting_exposure": { "fit": "None; the analytic posterior from the visible priors, likelihoods and observations", "evaluation": "The 400 generated queries" }, "comparison_note": "Zero by construction. examples/bayes_event_agent.py computes it through the original scorer.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json", "field": "/rows/0/reference" }, "metric_variant": "posterior_mae", "table2_reference_type": "optimal", "table2_mark": "", "table2_metric": "Posterior MAE" }, { "row_id": "false_belief_answer_key-original_answer_exact_match_mean", "reference_id": "false_belief_answer_key", "model": "ToMi answer key", "task": "false_belief_story_qa", "mode": "public_scenario_qa", "metric": "accuracy", "value": 1.0, "value_raw": "1.0", "value_status": "finite", "n": 5994, "n_detail": "5,994 questions", "aggregation": "response_weighted_mean", "unit": "fraction", "scope": "all", "provenance_type": "answer-key", "fitting_exposure": { "fit": "None; the answer key itself", "evaluation": "By definition" }, "comparison_note": "ToMi has no human responses, so the reference is the answer key itself; it is 100 by definition, not a measured performance.", "exact_precision": true, "precision": "by_definition", "status": "accepted_completed_reference", "source": { "record": null, "field": null, "note": "Author decision of 2026-10-30. Nothing was computed." }, "metric_variant": "original_answer_exact_match_mean", "table2_reference_type": "optimal", "table2_mark": null, "table2_metric": "Accuracy (%)" }, { "row_id": "rlwm-M9-argmax-accuracy", "reference_id": "rlwm_archived_models", "model": "RLWM archived fitted cognitive models", "task": "rlwm", "mode": "same_participant_fitted_human_history_replay", "metric": "accuracy", "value": 0.893368108710854, "value_raw": "0.893368108710854", "value_status": "finite", "n": 52911, "n_detail": "52,911 responses from 79 participants", "aggregation": "response_weighted_mean", "unit": "fraction", "scope": "all", "provenance_type": "archived-replay", "fitting_exposure": { "participants": 79, "scored_responses": 52911, "fitting": "Archived individual parameters fitted on these same 79 participants", "heldout": false, "history": "Archive uses RT-filtered histories; neural inputs retain earlier fast responses and omissions", "models_in_archive_order": [ 9, 6, 2, 8, 26, 38, 42 ] }, "comparison_note": "Derived argmax of the archived M9 fit, first legal action on an exact tie, not generated author actions. By participant 0.8928 (95% CI 0.8896 to 0.8959). Mean correct-action probability is the separate row rlwm-M9-correct_action_probability.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "backups/vislab_project_snapshot_20261028/runs/first_paper_24_20261026/analysis/rlwm_conditions_20261026/overall.csv", "field": "source=author_fitted_9, choice_accuracy_response_weighted" }, "metric_variant": "objective_label_accuracy_under_human_history", "table2_reference_type": "human fit", "table2_mark": "§", "table2_metric": "Accuracy (%)" }, { "reference_id": "horizon_rw3", "model": "RW3 (local formula fit)", "task": "horizon", "mode": "autonomous_generated_environment", "value_status": "finite", "n": 34720, "aggregation": "free_choice_mean", "scope": "all_31_original_timelines", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "parameter_fit": "25 original training participants; 28,000 free choices; forced choices retained for updates", "predictive_evaluation": "6 held-out participants; 6,720 free choices", "autonomous_evaluation": "31 original timelines including the 25 training-participant timelines; 34,720 free choices", "autonomous_action_rule": "Softmax sampling at temperature 1 with original participant seeds" }, "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "row_id": "rw3-autonomous-greedy-optimal_action_rate", "metric": "optimal_action_rate", "value": 0.8488767281105991, "value_raw": "0.8488767281105991", "unit": "fraction", "comparison_note": "Greedy replay of the same fit on the same 31 timelines, with no refit; exact ties take the first legal action. Table 3 prints the softmax sampling value.", "source": { "record": "runs/probes_20261028/analysis/reference_fill_20261030/horizon_rw3_greedy/RW3_GREEDY_RESULT.json", "field": "/greedy/optimal_rate" } }, { "reference_id": "horizon_rw3", "model": "RW3 (local formula fit)", "task": "horizon", "mode": "autonomous_generated_environment", "value_status": "finite", "n": 34720, "aggregation": "free_choice_mean", "scope": "all_31_original_timelines", "provenance_type": "disclosed-reimplementation", "fitting_exposure": { "parameter_fit": "25 original training participants; 28,000 free choices; forced choices retained for updates", "predictive_evaluation": "6 held-out participants; 6,720 free choices", "autonomous_evaluation": "31 original timelines including the 25 training-participant timelines; 34,720 free choices", "autonomous_action_rule": "Softmax sampling at temperature 1 with original participant seeds" }, "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "row_id": "rw3-autonomous-greedy-reward", "metric": "reward", "value": 55.99043778801843, "value_raw": "55.99043778801843", "unit": "original_reward_units_per_free_choice", "comparison_note": "Greedy replay of the same fit on the same 31 timelines, with no refit; exact ties take the first legal action. Table 3 prints the softmax sampling value.", "source": { "record": "runs/probes_20261028/analysis/reference_fill_20261030/horizon_rw3_greedy/RW3_GREEDY_RESULT.json", "field": "/greedy/mean_reward" } }, { "row_id": "uniform-wcst-autonomous_perseveration_error", "reference_id": "uniform_choice", "model": "Uniform", "task": "wcst", "mode": "generative", "protocol": "wcst_generative_75_test_participants", "target": "task_metrics_definition", "metric": "perseveration_error", "value": 0.17127027135271494, "value_raw": "0.17127027135271494", "value_status": "finite", "n": 75, "aggregation": "participant_mean", "unit": "fraction", "scope": "generative", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Exact expectation of task_metrics under uniform card choice, enumerating key pairs on consecutive trials.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "wcst_uniform_errors_20261031/UNIFORM_ERRORS.json", "field": "perseveration_error" }, "metric_variant": "participant_mean", "reproduced_by_uniform_cli_run": false }, { "row_id": "uniform-wcst-autonomous_set_loss_error", "reference_id": "uniform_choice", "model": "Uniform", "task": "wcst", "mode": "generative", "protocol": "wcst_generative_75_test_participants", "target": "task_metrics_definition", "metric": "set_loss_error", "value": 0.17127027135271494, "value_raw": "0.17127027135271494", "value_status": "finite", "n": 75, "aggregation": "participant_mean", "unit": "fraction", "scope": "generative", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Exact expectation of task_metrics under uniform card choice, enumerating key pairs on consecutive trials.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "wcst_uniform_errors_20261031/UNIFORM_ERRORS.json", "field": "set_loss_error" }, "metric_variant": "participant_mean", "reproduced_by_uniform_cli_run": false }, { "row_id": "human-wcst-accuracy", "reference_id": "human_participants", "model": "Human", "task": "wcst", "mode": "generative", "protocol": "wcst_observed_75_test_participants", "target": "observed_human_choice", "metric": "accuracy", "value": 0.500006279755749, "value_raw": "0.500006279755749", "value_status": "finite", "n": 75, "aggregation": "participant_mean", "unit": "fraction", "scope": "generative", "provenance_type": "human-data-summary", "fitting_exposure": { "fit": "None; observed human responses", "evaluation": "task_metrics in src/wcst_task.py on the 75 test participants" }, "comparison_note": "Participant mean of the observed human choices.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "wcst_uniform_errors_20261031/UNIFORM_ERRORS.json", "field": "human_participant_mean/accuracy" }, "metric_variant": "participant_mean", "reproduced_by_uniform_cli_run": false }, { "row_id": "human-wcst-perseveration_error", "reference_id": "human_participants", "model": "Human", "task": "wcst", "mode": "generative", "protocol": "wcst_observed_75_test_participants", "target": "observed_human_choice", "metric": "perseveration_error", "value": 0.0731772546148336, "value_raw": "0.0731772546148336", "value_status": "finite", "n": 75, "aggregation": "participant_mean", "unit": "fraction", "scope": "generative", "provenance_type": "human-data-summary", "fitting_exposure": { "fit": "None; observed human responses", "evaluation": "task_metrics in src/wcst_task.py on the 75 test participants" }, "comparison_note": "Participant mean of the observed human choices.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "wcst_uniform_errors_20261031/UNIFORM_ERRORS.json", "field": "human_participant_mean/perseveration_error" }, "metric_variant": "participant_mean", "reproduced_by_uniform_cli_run": false }, { "row_id": "human-wcst-set_loss_error", "reference_id": "human_participants", "model": "Human", "task": "wcst", "mode": "generative", "protocol": "wcst_observed_75_test_participants", "target": "observed_human_choice", "metric": "set_loss_error", "value": 0.01876096126383518, "value_raw": "0.01876096126383518", "value_status": "finite", "n": 75, "aggregation": "participant_mean", "unit": "fraction", "scope": "generative", "provenance_type": "human-data-summary", "fitting_exposure": { "fit": "None; observed human responses", "evaluation": "task_metrics in src/wcst_task.py on the 75 test participants" }, "comparison_note": "Participant mean of the observed human choices.", "exact_precision": true, "precision": "saved_aggregate_decimal", "status": "accepted_completed_reference", "source": { "record": "wcst_uniform_errors_20261031/UNIFORM_ERRORS.json", "field": "human_participant_mean/set_loss_error" }, "metric_variant": "participant_mean", "reproduced_by_uniform_cli_run": false }, { "row_id": "uniform-horizon-autonomous_mean_reward", "reference_id": "uniform_choice", "model": "Uniform", "task": "horizon", "mode": "autonomous_generative", "protocol": "namazova_generative_31_timelines_uniform_expectation", "target": "latent_mean_optimal_arm", "metric": "mean_reward", "value": 50.0, "value_raw": "50.0", "value_status": "finite", "n": 34720, "aggregation": "response_weighted_mean", "unit": "reward_points", "scope": "free_choices", "provenance_type": "analytic-expectation", "fitting_exposure": { "fit": "None; no model and no fitting", "evaluation": "Expectation under the original scoring rule" }, "comparison_note": "Arm means averaged over all 34,720 original free choices. The same in every readout. Printed to one decimal.", "exact_precision": false, "precision": "reported_one_decimal_source", "status": "reported_rounded_source", "source": { "record": "Horizon readout table of the paper (Table 4)", "field": "Uniform, Mean reward" }, "metric_variant": "arm_mean_over_free_choices", "reproduced_by_uniform_cli_run": false } ], "table2_extension": { "source_acceptance": "table2_values_20261030.json (main results table, paper commit 5a0f491)", "rows_added": 9 }, "table2_remaining": { "uniform_rows": 27, "reference_rows": 10 } }