{ "status": "planned; not existing HF datasets", "evidence": "experiments/tracking/data/01a/divl_frozen_actor/full_protocol.yaml", "additional_inputs": [ "experiments/tracking/data/01a/grid.csv", "experiments/tracking/01a_dagger_mining/paper_baselines.md", "experiments/tracking/data/01a/divl_frozen_actor/full_campaign/results.csv", "experiments/tracking/data/01a/divl_frozen_actor/mainline_outputs.json" ], "bundles": [ { "destination_repo": "mulligan/sim-square-narrow-r00-r03-eval", "action": "create-from-artifacts", "cells": [ { "cell": "sq_d0_r1_auto_iql_n32", "family": "autonomous_iql", "round": "r1", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-08-03/01a_sim_autonomous_baselines/grid/sq_d0/r1/auto_iql_n32/seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r1_auto_iql_n32/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r1_auto_iql_success_bc_n32", "family": "autonomous_iql", "round": "r1", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-08-03/01a_sim_autonomous_baselines/grid/sq_d0/r1/auto_iql_success_bc_n32/seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r1_auto_iql_success_bc_n32/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r2_auto_iql_n32", "family": "autonomous_iql", "round": "r2", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-08-03/01a_sim_autonomous_baselines/grid/sq_d0/r2/auto_iql_n32/seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r2_auto_iql_n32/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r2_auto_iql_success_bc_n32", "family": "autonomous_iql", "round": "r2", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-08-03/01a_sim_autonomous_baselines/grid/sq_d0/r2/auto_iql_success_bc_n32/seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r2_auto_iql_success_bc_n32/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r3_auto_iql_n32", "family": "autonomous_iql", "round": "r3", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-08-03/01a_sim_autonomous_baselines/grid/sq_d0/r3/auto_iql_n32/seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r3_auto_iql_n32/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r3_auto_iql_success_bc_n32", "family": "autonomous_iql", "round": "r3", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-08-03/01a_sim_autonomous_baselines/grid/sq_d0/r3/auto_iql_success_bc_n32/seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r3_auto_iql_success_bc_n32/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r0_baseline_uniform", "family": "historical_hil", "round": "r0", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_equal_tile_points", "path": "experiments/2026-05-23/init_states/sq_d0_01a_equal_tile_sobol_100pcell_seed20260524.json", "file_sha256": "dc802983cf54a32d7819400ec993887da9276b68ebcd5de221555078364ebe8e", "manifest_hash": "38bcc681b4baf571e87668b8f1ab200421b91fbdb599c95dec34e64d476f47cb", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-23/11_eval_sq_d0_01a_equal_tile_points.sh", "parent_result_path_template": "plots/2026-05-23/01a_sq_d0_equal_tile_grid/baseline_uniform_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r0_baseline_uniform/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r0_ours_sobol", "family": "historical_hil", "round": "r0", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_equal_tile_points", "path": "experiments/2026-05-23/init_states/sq_d0_01a_equal_tile_sobol_100pcell_seed20260524.json", "file_sha256": "dc802983cf54a32d7819400ec993887da9276b68ebcd5de221555078364ebe8e", "manifest_hash": "38bcc681b4baf571e87668b8f1ab200421b91fbdb599c95dec34e64d476f47cb", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-23/11_eval_sq_d0_01a_equal_tile_points.sh", "parent_result_path_template": "plots/2026-05-23/01a_sq_d0_equal_tile_grid/ours_sobol_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r0_ours_sobol/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r1_baseline_uniform_nocf_human_only", "family": "historical_hil", "round": "r1", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-05-25/01a_sq_d0_r1_beta05_valid_sobol8k/baseline_uniform_nocf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r1_baseline_uniform_nocf_human_only/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r1_baseline_uniform_nocf_straddled_auto_success", "family": "historical_hil", "round": "r1", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-05-25/01a_sq_d0_r1_beta05_valid_sobol8k/baseline_uniform_nocf_straddled_auto_success_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r1_baseline_uniform_nocf_straddled_auto_success/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r1_ours_mining_beta05_freecf_human_only", "family": "historical_hil", "round": "r1", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-05-25/01a_sq_d0_r1_beta05_valid_sobol8k/ours_mining_beta05_freecf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-29/01a_divl_frozen_actor/grid/sq_d0/seed{seed}/results.json" ], "source_runtime_root": "/iris/u/ankile/self-improving-robots-workspace/self-improving-robots/worktrees/divl-frozen-actor-da3e", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r1_ours_mining_beta05_nocf_human_only", "family": "historical_hil", "round": "r1", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-05-25/01a_sq_d0_r1_beta05_valid_sobol8k/ours_mining_beta05_nocf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r1_ours_mining_beta05_nocf_human_only/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r1_ours_sobol_freecf_human_only", "family": "historical_hil", "round": "r1", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-05-25/01a_sq_d0_r1_beta05_valid_sobol8k/ours_sobol_freecf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r1_ours_sobol_freecf_human_only/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r1_ours_sobol_nocf_human_only", "family": "historical_hil", "round": "r1", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-05-25/01a_sq_d0_r1_beta05_valid_sobol8k/ours_sobol_nocf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r1_ours_sobol_nocf_human_only/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r2_baseline_uniform_nocf_human_only", "family": "historical_hil", "round": "r2", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-05-28/01a_sq_d0_r2_mining_shape_valid_sobol8k/baseline_uniform_nocf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r2_baseline_uniform_nocf_human_only/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r2_ours_mining_shape_beta05_freecf_human_only", "family": "historical_hil", "round": "r2", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-05-28/01a_sq_d0_r2_mining_shape_valid_sobol8k/ours_mining_shape_beta05_freecf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r2_ours_mining_shape_beta05_freecf_human_only/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r2_ours_mining_shape_beta05_nocf_human_only", "family": "historical_hil", "round": "r2", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-05-28/01a_sq_d0_r2_mining_shape_valid_sobol8k/ours_mining_shape_beta05_nocf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r2_ours_mining_shape_beta05_nocf_human_only/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r3_baseline_uniform_nocf_human_only", "family": "historical_hil", "round": "r3", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-06-04/01a_sq_d0_r3_valid_sobol8k/baseline_uniform_nocf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r3_baseline_uniform_nocf_human_only/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r3_ours_grid_cell_bonus_b1p0_freecf_human_only", "family": "historical_hil", "round": "r3", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-06-04/01a_sq_d0_r3_valid_sobol8k/ours_grid_cell_bonus_b1p0_freecf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r3_ours_grid_cell_bonus_b1p0_freecf_human_only/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r3_ours_grid_cell_bonus_b1p0_nocf_human_only", "family": "historical_hil", "round": "r3", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-06-04/01a_sq_d0_r3_valid_sobol8k/ours_grid_cell_bonus_b1p0_nocf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r3_ours_grid_cell_bonus_b1p0_nocf_human_only/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r3_repair_baseline_uniform_nocf_human_only_r3roll", "family": "historical_hil", "round": "r3_symmetry_repair", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-06-05/01a_sq_d0_r3_symmetry_repair_valid_sobol8k/baseline_uniform_nocf_human_only_r3roll_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r3_repair_baseline_uniform_nocf_human_only_r3roll/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d0_r3_repair_ours_grid_cell_bonus_b1p0_freecf_human_only_no_r3roll", "family": "historical_hil", "round": "r3_symmetry_repair", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d0_valid_sobol8k_points", "path": "experiments/2026-05-24/init_states/sq_d0_valid_sobol8k_seed2026052402.json", "file_sha256": "93489d5e02043ee5f0f2f3ec4f1a36ade60b156114f822b82eb41d0d3aee57cc", "manifest_hash": "b17ba12eee08a6a83eb85379449c657d3f2bdbe9dad8b9cc211d8752e8daa499", "num_points": 8000, "num_cells": 80, "shards": 1, "launch_script": "experiments/2026-05-25/05_eval_sq_d0_01a_valid_sobol8k_points.sh", "parent_result_path_template": "plots/2026-06-05/01a_sq_d0_r3_symmetry_repair_valid_sobol8k/ours_grid_cell_bonus_b1p0_freecf_human_only_no_r3roll_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d0_r3_repair_ours_grid_cell_bonus_b1p0_freecf_human_only_no_r3roll/seed{seed}/results.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." } ] }, { "destination_repo": "mulligan/sim-square-broad-r00-r03-eval", "action": "create-from-artifacts", "cells": [ { "cell": "sq_d1_r1_auto_iql_n32", "family": "autonomous_iql", "round": "r1", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-08-03/01a_sim_autonomous_baselines/grid/sq_d1/r1/auto_iql_n32/seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_auto_iql_n32/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_auto_iql_n32/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_auto_iql_n32/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_auto_iql_n32/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r1_auto_iql_success_bc_n32", "family": "autonomous_iql", "round": "r1", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-08-03/01a_sim_autonomous_baselines/grid/sq_d1/r1/auto_iql_success_bc_n32/seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_auto_iql_success_bc_n32/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_auto_iql_success_bc_n32/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_auto_iql_success_bc_n32/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_auto_iql_success_bc_n32/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r2_auto_iql_n32", "family": "autonomous_iql", "round": "r2", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-08-03/01a_sim_autonomous_baselines/grid/sq_d1/r2/auto_iql_n32/seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_auto_iql_n32/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_auto_iql_n32/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_auto_iql_n32/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_auto_iql_n32/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r2_auto_iql_success_bc_n32", "family": "autonomous_iql", "round": "r2", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-08-03/01a_sim_autonomous_baselines/grid/sq_d1/r2/auto_iql_success_bc_n32/seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_auto_iql_success_bc_n32/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_auto_iql_success_bc_n32/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_auto_iql_success_bc_n32/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_auto_iql_success_bc_n32/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r3_auto_iql_n32", "family": "autonomous_iql", "round": "r3", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-08-03/01a_sim_autonomous_baselines/grid/sq_d1/r3/auto_iql_n32/seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_auto_iql_n32/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_auto_iql_n32/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_auto_iql_n32/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_auto_iql_n32/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r3_auto_iql_success_bc_n32", "family": "autonomous_iql", "round": "r3", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-08-03/01a_sim_autonomous_baselines/grid/sq_d1/r3/auto_iql_success_bc_n32/seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_auto_iql_success_bc_n32/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_auto_iql_success_bc_n32/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_auto_iql_success_bc_n32/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_auto_iql_success_bc_n32/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r0_baseline_uniform", "family": "historical_hil", "round": "r0", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-05-24/01a_sq_d1_valid_sobol30k/full200/baseline_uniform_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r0_baseline_uniform/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r0_baseline_uniform/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r0_baseline_uniform/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r0_baseline_uniform/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r0_ours_sobol", "family": "historical_hil", "round": "r0", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-05-24/01a_sq_d1_valid_sobol30k/full200/ours_sobol_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r0_ours_sobol/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r0_ours_sobol/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r0_ours_sobol/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r0_ours_sobol/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r0_first100_baseline_uniform", "family": "historical_hil", "round": "r0_first100", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-05-24/01a_sq_d1_valid_sobol30k/first100/baseline_uniform_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r0_first100_baseline_uniform/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r0_first100_baseline_uniform/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r0_first100_baseline_uniform/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r0_first100_baseline_uniform/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r0_first100_ours_sobol", "family": "historical_hil", "round": "r0_first100", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-05-24/01a_sq_d1_valid_sobol30k/first100/ours_sobol_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r0_first100_ours_sobol/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r0_first100_ours_sobol/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r0_first100_ours_sobol/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r0_first100_ours_sobol/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r1_baseline_uniform_nocf_human_only", "family": "historical_hil", "round": "r1", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-05-28/01a_sq_d1_r1_valid_sobol30k/baseline_uniform_nocf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_baseline_uniform_nocf_human_only/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_baseline_uniform_nocf_human_only/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_baseline_uniform_nocf_human_only/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_baseline_uniform_nocf_human_only/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r1_baseline_uniform_nocf_straddled_auto_success", "family": "historical_hil", "round": "r1", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-05-28/01a_sq_d1_r1_valid_sobol30k/baseline_uniform_nocf_straddled_auto_success_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_baseline_uniform_nocf_straddled_auto_success/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_baseline_uniform_nocf_straddled_auto_success/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_baseline_uniform_nocf_straddled_auto_success/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_baseline_uniform_nocf_straddled_auto_success/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r1_ours_mining_freecf_human_only", "family": "historical_hil", "round": "r1", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-05-28/01a_sq_d1_r1_valid_sobol30k/ours_mining_freecf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-29/01a_divl_frozen_actor/grid/sq_d1/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-29/01a_divl_frozen_actor/grid/sq_d1/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-29/01a_divl_frozen_actor/grid/sq_d1/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-29/01a_divl_frozen_actor/grid/sq_d1/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "/iris/u/ankile/self-improving-robots-workspace/self-improving-robots/worktrees/divl-frozen-actor-da3e", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r1_ours_mining_nocf_human_only", "family": "historical_hil", "round": "r1", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-05-28/01a_sq_d1_r1_valid_sobol30k/ours_mining_nocf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_ours_mining_nocf_human_only/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_ours_mining_nocf_human_only/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_ours_mining_nocf_human_only/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_ours_mining_nocf_human_only/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r1_ours_sobol_freecf_human_only", "family": "historical_hil", "round": "r1", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-05-28/01a_sq_d1_r1_valid_sobol30k/ours_sobol_freecf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_ours_sobol_freecf_human_only/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_ours_sobol_freecf_human_only/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_ours_sobol_freecf_human_only/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_ours_sobol_freecf_human_only/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r1_ours_sobol_nocf_human_only", "family": "historical_hil", "round": "r1", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-05-28/01a_sq_d1_r1_valid_sobol30k/ours_sobol_nocf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_ours_sobol_nocf_human_only/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_ours_sobol_nocf_human_only/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_ours_sobol_nocf_human_only/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r1_ours_sobol_nocf_human_only/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r2_baseline_uniform_nocf_human_only", "family": "historical_hil", "round": "r2", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-06-04/01a_sq_d1_r2_valid_sobol30k/baseline_uniform_nocf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_baseline_uniform_nocf_human_only/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_baseline_uniform_nocf_human_only/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_baseline_uniform_nocf_human_only/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_baseline_uniform_nocf_human_only/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r2_ours_mining_freecf_human_only", "family": "historical_hil", "round": "r2", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-06-04/01a_sq_d1_r2_valid_sobol30k/ours_mining_freecf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_ours_mining_freecf_human_only/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_ours_mining_freecf_human_only/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_ours_mining_freecf_human_only/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_ours_mining_freecf_human_only/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r2_ours_mining_nocf_human_only", "family": "historical_hil", "round": "r2", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-06-04/01a_sq_d1_r2_valid_sobol30k/ours_mining_nocf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_ours_mining_nocf_human_only/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_ours_mining_nocf_human_only/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_ours_mining_nocf_human_only/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_ours_mining_nocf_human_only/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r2_ours_sobol_freecf_human_only", "family": "historical_hil", "round": "r2", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-06-04/01a_sq_d1_r2_valid_sobol30k/ours_sobol_freecf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_ours_sobol_freecf_human_only/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_ours_sobol_freecf_human_only/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_ours_sobol_freecf_human_only/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_ours_sobol_freecf_human_only/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r2_ours_sobol_nocf_human_only", "family": "historical_hil", "round": "r2", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-06-04/01a_sq_d1_r2_valid_sobol30k/ours_sobol_nocf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_ours_sobol_nocf_human_only/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_ours_sobol_nocf_human_only/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_ours_sobol_nocf_human_only/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r2_ours_sobol_nocf_human_only/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r3_baseline_uniform_nocf_human_only", "family": "historical_hil", "round": "r3", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-06-11/01a_sq_d1_r3_valid_sobol30k/baseline_uniform_nocf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_baseline_uniform_nocf_human_only/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_baseline_uniform_nocf_human_only/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_baseline_uniform_nocf_human_only/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_baseline_uniform_nocf_human_only/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r3_ours_mining_freecf_human_only", "family": "historical_hil", "round": "r3", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-06-11/01a_sq_d1_r3_valid_sobol30k/ours_mining_freecf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_ours_mining_freecf_human_only/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_ours_mining_freecf_human_only/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_ours_mining_freecf_human_only/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_ours_mining_freecf_human_only/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." }, { "cell": "sq_d1_r3_ours_mining_nocf_human_only", "family": "historical_hil", "round": "r3", "seeds": [ 1, 2, 3, 4, 5 ], "grid": { "kind": "sq_d1_valid_sobol30k_points", "path": "experiments/2026-05-24/init_states/sq_d1_valid_sobol30k_seed2026052499.json", "file_sha256": "cc9adb9c2d25a8e89c0e849abaf699310415f7edd7d727ca4ac5606d3ee1cb84", "manifest_hash": "a5754fa5b686c239255be76b2ca47c4e19924cc46b8fca629395a02b93525737", "num_points": 30000, "num_cells": 720, "shards": 4, "launch_script": "experiments/2026-05-24/03_eval_sq_d1_01a_valid_sobol30k_points.sh", "parent_result_path_template": "plots/2026-06-11/01a_sq_d1_r3_valid_sobol30k/ours_mining_nocf_human_only_seed{seed}/results.json" }, "divl_raw_path_templates": [ "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_ours_mining_nocf_human_only/seed{seed}/results_shard_1_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_ours_mining_nocf_human_only/seed{seed}/results_shard_2_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_ours_mining_nocf_human_only/seed{seed}/results_shard_3_of_4.json", "plots/2026-08-31/01a_divl_frozen_actor_full/grid/sq_d1_r3_ours_mining_nocf_human_only/seed{seed}/results_shard_4_of_4.json" ], "source_runtime_root": "repo-relative S3 results identity", "availability": "Paths verified in ingestion code; hydrate and hash-check every raw result before packaging." } ] } ] }