{
  "manifest_version": 1,
  "generated_from": "Preserved raw episode CSVs and checked analysis/trace records; no benchmark data were written.",
  "protocols": {
    "phase2": "Corrected local evaluation; nine conditions × 3,000 episodes; 27,000 total.",
    "iterative_development": "Separate Argon High iterative experiment; three versions × 3,000 episodes; 9,000 total.",
    "diagnostic_traces": "Single preserved historical diagnostic replay per primary model; not representative aggregate traces."
  },
  "sources": [
    {
      "path": "experiments/argon_high_iterative_development/evaluation/full_run/analysis/condition_metrics.json",
      "sha256": "e5c5ef1fa7bbc2d20bc2b6094ec7d7a7e6524d44daa7b6ea5460eb9ff2c715a2"
    },
    {
      "path": "experiments/argon_high_iterative_development/evaluation/full_run/conditions/argon_high_followup_1/episode_scores.csv",
      "sha256": "dc2a1fca571ccede39c5d4b96f5a93a07a4ab00b092d990ffd177a112774a527"
    },
    {
      "path": "experiments/argon_high_iterative_development/evaluation/full_run/conditions/argon_high_followup_2/episode_scores.csv",
      "sha256": "6f2dff042bc9d47fd5d4d2c9caa11a920f6a2bc89d55061c4a3c88cbc08b6cd0"
    },
    {
      "path": "experiments/argon_high_iterative_development/evaluation/full_run/conditions/argon_high_original/episode_scores.csv",
      "sha256": "a6d14220c51b14e88e1288a3e396731aa3e4fa2e8fdd3f9895f115735cf0da1c"
    },
    {
      "path": "experiments/corrected_evaluation/failure_trace_replays.json",
      "sha256": "7af54ed95222f2ccf4821408dec689faf8dcb601d86179550cc09d780e05f129"
    },
    {
      "path": "experiments/corrected_evaluation/full_run/analysis/condition_metrics.json",
      "sha256": "bfe30618860f004091c22ef9b7e21e69e4729967d363c952bef7b8a234757ad1"
    },
    {
      "path": "experiments/corrected_evaluation/full_run/analysis/paired_comparisons.json",
      "sha256": "f49e8a57858e185f87c735521b58e5527ea1e2534c58ac43317b11ccaafb46e9"
    },
    {
      "path": "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv",
      "sha256": "95739181a6b781ce3ef3318aef4a6660645038a91aec31b33e11bc69ceb56872"
    },
    {
      "path": "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
      "sha256": "86d5a91027649e8d07f898d8ad3e145494880e5fc06e4fc6671efb0c8b7fc3db"
    },
    {
      "path": "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
      "sha256": "80211f3504025b24197768fc648991342ae54747621afb0b89454b139795a6a8"
    },
    {
      "path": "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
      "sha256": "382d63a7df6f412624aabb1d0391c2ae9c076e6f316dab1e7e0dd0612912e466"
    },
    {
      "path": "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
      "sha256": "205e8fd0085e4e6e3d8a83b4fc102a65a220743ca6a695c15ad073ef3a8e8788"
    },
    {
      "path": "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
      "sha256": "300413380b0d284fb10968ff000f7eb9272ec816feda06729bd4e71a6e208002"
    },
    {
      "path": "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
      "sha256": "8076f8fedce248fbe865f084b8ff7dd287f2d158db0588b2b7ba8e75b67c59a1"
    },
    {
      "path": "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
      "sha256": "c47a86dbed6b1addf32ba7838a4c93ec6c3966c2f12170a65928bd284ce8cdcc"
    },
    {
      "path": "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
      "sha256": "bd42aa85035b89085001a46d7842e879dcb747ac9d644c631ffb0d977871fbbe"
    }
  ],
  "assets": [
    {
      "asset_name": "benchmark_conditions.svg",
      "description": "Mean score and standard error across all nine corrected benchmark conditions; Astra is separated on its own linear panel rather than a broken axis.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "svg",
      "file_size_bytes": 26904,
      "sha256": "5f161edce3c056e6166312215728cc1d51fc303b0dd18787de9d5dd2e2d64f19",
      "resolution": "vector",
      "figure_size_inches": [
        12.5,
        5.6
      ]
    },
    {
      "asset_name": "benchmark_conditions.png",
      "description": "Mean score and standard error across all nine corrected benchmark conditions; Astra is separated on its own linear panel rather than a broken axis.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "png",
      "file_size_bytes": 278291,
      "sha256": "d77bd6faa350e6c5e6aa82f331b122ce046eaf1a8153a44001d75a4eb5abd899",
      "resolution_px": [
        3692,
        1697
      ],
      "dpi": 300
    },
    {
      "asset_name": "benchmark_conditions.pdf",
      "description": "Mean score and standard error across all nine corrected benchmark conditions; Astra is separated on its own linear panel rather than a broken axis.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "pdf",
      "file_size_bytes": 31847,
      "sha256": "9611ea0dfb58e5f7fc5251f7a39d07c2be6a57e6cdd0f28bbf77367338e934d5",
      "resolution": "vector",
      "figure_size_inches": [
        12.5,
        5.6
      ]
    },
    {
      "asset_name": "primary_zero_shot_ranking.svg",
      "description": "Five original zero-shot conditions ranked by raw mean score with standard error; excludes Astra, PPO, controls, and DeepSeek Corrected.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "svg",
      "file_size_bytes": 17156,
      "sha256": "35c31ca6d50c42636c1bd49ba7fcf7dc30eb8e1f1b3af24147b713d4573be81c",
      "resolution": "vector",
      "figure_size_inches": [
        9.2,
        4.9
      ]
    },
    {
      "asset_name": "primary_zero_shot_ranking.png",
      "description": "Five original zero-shot conditions ranked by raw mean score with standard error; excludes Astra, PPO, controls, and DeepSeek Corrected.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "png",
      "file_size_bytes": 145856,
      "sha256": "7ae5ef2147b08a26c86ba1a14deea2943aea8c3c6ff6416f87a8902c9ca05c3f",
      "resolution_px": [
        2978,
        1334
      ],
      "dpi": 300
    },
    {
      "asset_name": "primary_zero_shot_ranking.pdf",
      "description": "Five original zero-shot conditions ranked by raw mean score with standard error; excludes Astra, PPO, controls, and DeepSeek Corrected.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "pdf",
      "file_size_bytes": 24338,
      "sha256": "4219e283a76161fd45f56e174111b1fc8c6a30b046445248533d8f1f8070cd0e",
      "resolution": "vector",
      "figure_size_inches": [
        9.2,
        4.9
      ]
    },
    {
      "asset_name": "score_distributions_ecdf.svg",
      "description": "Empirical cumulative distributions from all 27,000 corrected Phase 2 episode scores, shown in separate role-aware panels with a clearly labeled log1p score axis.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "svg",
      "file_size_bytes": 73571,
      "sha256": "6904a291f9a31a493ba96c99d976e58cc66ff5082f4d93114caf176eb7a85b58",
      "resolution": "vector",
      "figure_size_inches": [
        12.2,
        5.2
      ]
    },
    {
      "asset_name": "score_distributions_ecdf.png",
      "description": "Empirical cumulative distributions from all 27,000 corrected Phase 2 episode scores, shown in separate role-aware panels with a clearly labeled log1p score axis.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "png",
      "file_size_bytes": 275463,
      "sha256": "b0d1d402a4a0568606b927f88f118a57cb1411d95216cfcf4b5532c83f3904a5",
      "resolution_px": [
        4375,
        1589
      ],
      "dpi": 300
    },
    {
      "asset_name": "score_distributions_ecdf.pdf",
      "description": "Empirical cumulative distributions from all 27,000 corrected Phase 2 episode scores, shown in separate role-aware panels with a clearly labeled log1p score axis.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "pdf",
      "file_size_bytes": 38739,
      "sha256": "1d95e48e3f80280ad95874f088671467a226640205b2299b1d36cb5c996d2cdc",
      "resolution": "vector",
      "figure_size_inches": [
        12.2,
        5.2
      ]
    },
    {
      "asset_name": "primary_per_seed.svg",
      "description": "Mean scores and standard errors for each of the five primary zero-shot conditions split across seed groups 42, 100, and 200; 1,000 raw episodes per point.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "svg",
      "file_size_bytes": 37731,
      "sha256": "bcd22a78c82ef2742f89431c201a2c4870bd08d1b6207d792209b57d1eb5c015",
      "resolution": "vector",
      "figure_size_inches": [
        9.6,
        10.2
      ]
    },
    {
      "asset_name": "primary_per_seed.png",
      "description": "Mean scores and standard errors for each of the five primary zero-shot conditions split across seed groups 42, 100, and 200; 1,000 raw episodes per point.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "png",
      "file_size_bytes": 302481,
      "sha256": "5e5bdb936056a263e3bb302a88287dbc1887fe1931b6d94fb7c678cbdf56883e",
      "resolution_px": [
        3080,
        3051
      ],
      "dpi": 300
    },
    {
      "asset_name": "primary_per_seed.pdf",
      "description": "Mean scores and standard errors for each of the five primary zero-shot conditions split across seed groups 42, 100, and 200; 1,000 raw episodes per point.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "pdf",
      "file_size_bytes": 28329,
      "sha256": "e716e97b3eddb1538d27578df48c9b0389f115c27f8824c5a96cdd88b03f795b",
      "resolution": "vector",
      "figure_size_inches": [
        9.6,
        10.2
      ]
    },
    {
      "asset_name": "iterative_development_comparison.svg",
      "description": "Three distinct Argon High iterative-development versions with mean, standard error, median, and 3,000 episode count; not interchangeable with zero-shot conditions.",
      "source_data": [
        "experiments/argon_high_iterative_development/evaluation/full_run/conditions/argon_high_original/episode_scores.csv",
        "experiments/argon_high_iterative_development/evaluation/full_run/conditions/argon_high_followup_1/episode_scores.csv",
        "experiments/argon_high_iterative_development/evaluation/full_run/conditions/argon_high_followup_2/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "svg",
      "file_size_bytes": 12233,
      "sha256": "d41db75ede59d0c3d22fa5962636825d46a5f6ae6a9ffc769eb6a442951d07ec",
      "resolution": "vector",
      "figure_size_inches": [
        9.2,
        4.7
      ]
    },
    {
      "asset_name": "iterative_development_comparison.png",
      "description": "Three distinct Argon High iterative-development versions with mean, standard error, median, and 3,000 episode count; not interchangeable with zero-shot conditions.",
      "source_data": [
        "experiments/argon_high_iterative_development/evaluation/full_run/conditions/argon_high_original/episode_scores.csv",
        "experiments/argon_high_iterative_development/evaluation/full_run/conditions/argon_high_followup_1/episode_scores.csv",
        "experiments/argon_high_iterative_development/evaluation/full_run/conditions/argon_high_followup_2/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "png",
      "file_size_bytes": 140620,
      "sha256": "cd3d9496918873ae5f8c43fa7d573e0ef2742c9a88d3452867352437d5a1043d",
      "resolution_px": [
        3531,
        1281
      ],
      "dpi": 300
    },
    {
      "asset_name": "iterative_development_comparison.pdf",
      "description": "Three distinct Argon High iterative-development versions with mean, standard error, median, and 3,000 episode count; not interchangeable with zero-shot conditions.",
      "source_data": [
        "experiments/argon_high_iterative_development/evaluation/full_run/conditions/argon_high_original/episode_scores.csv",
        "experiments/argon_high_iterative_development/evaluation/full_run/conditions/argon_high_followup_1/episode_scores.csv",
        "experiments/argon_high_iterative_development/evaluation/full_run/conditions/argon_high_followup_2/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "pdf",
      "file_size_bytes": 26237,
      "sha256": "fb25f6c1121857bcfa3dc6dba2af10f6dda148d257dfe5362f4daa6f375d18f9",
      "resolution": "vector",
      "figure_size_inches": [
        9.2,
        4.7
      ]
    },
    {
      "asset_name": "primary_paired_comparisons.svg",
      "description": "Paired-bootstrap 95% intervals for the ten primary zero-shot pairwise mean differences, with exact paired Cohen's dz, two-sided sign-flip p-values, and Holm p-values from the full 36-comparison family.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/analysis/paired_comparisons.json"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "svg",
      "file_size_bytes": 24939,
      "sha256": "98b5939e35b0e3ba1a037dac8705f929079a7c6f3c6373fb9f7859f4b74d6f49",
      "resolution": "vector",
      "figure_size_inches": [
        15.2,
        7.0
      ]
    },
    {
      "asset_name": "primary_paired_comparisons.png",
      "description": "Paired-bootstrap 95% intervals for the ten primary zero-shot pairwise mean differences, with exact paired Cohen's dz, two-sided sign-flip p-values, and Holm p-values from the full 36-comparison family.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/analysis/paired_comparisons.json"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "png",
      "file_size_bytes": 311127,
      "sha256": "31d948c782394c92042e433011fdb22b54696b0ba05b9949317ef399cdadf587",
      "resolution_px": [
        3854,
        1987
      ],
      "dpi": 300
    },
    {
      "asset_name": "primary_paired_comparisons.pdf",
      "description": "Paired-bootstrap 95% intervals for the ten primary zero-shot pairwise mean differences, with exact paired Cohen's dz, two-sided sign-flip p-values, and Holm p-values from the full 36-comparison family.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/analysis/paired_comparisons.json"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "pdf",
      "file_size_bytes": 35296,
      "sha256": "090dd9e480dc280de02e0030531bdad788361852cbed93a87251613ede36177e",
      "resolution": "vector",
      "figure_size_inches": [
        15.2,
        7.0
      ]
    },
    {
      "asset_name": "primary_action_distributions.svg",
      "description": "Normalized action-selection frequencies computed by summing actual action counts from raw episode CSVs for the five primary zero-shot policies; does not conflate selected with executed actions.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "svg",
      "file_size_bytes": 20596,
      "sha256": "af298d319c244fe0fdc2081d573c3e31a25756312fcf3f118b8d513659816eb5",
      "resolution": "vector",
      "figure_size_inches": [
        10.5,
        4.7
      ]
    },
    {
      "asset_name": "primary_action_distributions.png",
      "description": "Normalized action-selection frequencies computed by summing actual action counts from raw episode CSVs for the five primary zero-shot policies; does not conflate selected with executed actions.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "png",
      "file_size_bytes": 182160,
      "sha256": "73b81adbfc10c7c53b8401caf243012925c0ebafe78fae7ffdce2f4b3cd17462",
      "resolution_px": [
        3415,
        1229
      ],
      "dpi": 300
    },
    {
      "asset_name": "primary_action_distributions.pdf",
      "description": "Normalized action-selection frequencies computed by summing actual action counts from raw episode CSVs for the five primary zero-shot policies; does not conflate selected with executed actions.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "pdf",
      "file_size_bytes": 29879,
      "sha256": "2eff773e7c8c9f59e59c50b6cdae0e08a9b818b9d91a9ef6f945585831935790",
      "resolution": "vector",
      "figure_size_inches": [
        10.5,
        4.7
      ]
    },
    {
      "asset_name": "episode_lengths_and_cutoffs.svg",
      "description": "Episode-length ECDFs, mean/median lengths, and 10,000-step cutoff counts/rates computed from all nine corrected raw episode CSVs; the ECDF panel omits Astra's much longer durations, while the other panels include all conditions.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "svg",
      "file_size_bytes": 94990,
      "sha256": "274ca2dcff96407e11d5fcb414c747503671e90b618af4d31b2ab2fc4740bd35",
      "resolution": "vector",
      "figure_size_inches": [
        16.0,
        5.7
      ]
    },
    {
      "asset_name": "episode_lengths_and_cutoffs.png",
      "description": "Episode-length ECDFs, mean/median lengths, and 10,000-step cutoff counts/rates computed from all nine corrected raw episode CSVs; the ECDF panel omits Astra's much longer durations, while the other panels include all conditions.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "png",
      "file_size_bytes": 349366,
      "sha256": "33a98b27a854cfbdc181272afb6fb6e08e8032076e2e513aefbadbb47c392cf7",
      "resolution_px": [
        4716,
        1726
      ],
      "dpi": 300
    },
    {
      "asset_name": "episode_lengths_and_cutoffs.pdf",
      "description": "Episode-length ECDFs, mean/median lengths, and 10,000-step cutoff counts/rates computed from all nine corrected raw episode CSVs; the ECDF panel omits Astra's much longer durations, while the other panels include all conditions.",
      "source_data": [
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "pdf",
      "file_size_bytes": 50929,
      "sha256": "80ba1c8dcb1aaf48ab0d74852b7228e94ebb64674dcd5f4c04871653b28f3358",
      "resolution": "vector",
      "figure_size_inches": [
        16.0,
        5.7
      ]
    },
    {
      "asset_name": "diagnostic_trajectory_traces.svg",
      "description": "Step-by-step state positions and selected/executed actions and bricks remaining from five preserved single-episode historical diagnostic traces; not representative aggregate behavior.",
      "source_data": [
        "experiments/corrected_evaluation/failure_trace_replays.json",
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "svg",
      "file_size_bytes": 430484,
      "sha256": "6930065aa0b4e33c9813d6240f3fc2c522c3bc1b94415139ff66ed96684b43ad",
      "resolution": "vector",
      "figure_size_inches": [
        11.0,
        12.8
      ]
    },
    {
      "asset_name": "diagnostic_trajectory_traces.png",
      "description": "Step-by-step state positions and selected/executed actions and bricks remaining from five preserved single-episode historical diagnostic traces; not representative aggregate behavior.",
      "source_data": [
        "experiments/corrected_evaluation/failure_trace_replays.json",
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "png",
      "file_size_bytes": 1322636,
      "sha256": "17c26fd9d77485b817605ab81e8a61f703dc78797df01f90d075bca932d53930",
      "resolution_px": [
        4138,
        3800
      ],
      "dpi": 300
    },
    {
      "asset_name": "diagnostic_trajectory_traces.pdf",
      "description": "Step-by-step state positions and selected/executed actions and bricks remaining from five preserved single-episode historical diagnostic traces; not representative aggregate behavior.",
      "source_data": [
        "experiments/corrected_evaluation/failure_trace_replays.json",
        "experiments/corrected_evaluation/full_run/conditions/gemini_argon_4_medium/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/claude_sonnet_55/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/ppo_trained_baseline/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/gpt6_luna_extra_high/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/zero_weight_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/deepseek_original/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/random_action_control/episode_scores.csv",
        "experiments/corrected_evaluation/full_run/conditions/astra_local_reference/episode_scores.csv"
      ],
      "generation_script": "docs/scripts/visualization/generate_figures.py",
      "format": "pdf",
      "file_size_bytes": 51384,
      "sha256": "898fe107543ffb09a44b8ecd6c711363c6f42604435ab0eb4cf3749966b4d81c",
      "resolution": "vector",
      "figure_size_inches": [
        11.0,
        12.8
      ]
    }
  ]
}
