{
  "article_slug": "offline-rl-support-constraints",
  "figures": [
    {
      "id": "backup-error",
      "kind": "analytical_curve",
      "caption": "Exact author-constructed persistent perturbation recurrence e_(t+1)=.1+gamma e_t, e0=0. No agent training or measured return.",
      "x": {
        "name": "Backup iterations t",
        "values": [
          0,
          1,
          2,
          5,
          10,
          20,
          50,
          100
        ]
      },
      "series": [
        {
          "name": "gamma=0.5",
          "values": [
            0.0,
            0.1,
            0.15000000000000002,
            0.19375,
            0.19980468750000002,
            0.19999980926513672,
            0.19999999999999984,
            0.2
          ]
        },
        {
          "name": "gamma=0.9",
          "values": [
            0.0,
            0.1,
            0.19,
            0.40951000000000004,
            0.6513215599000001,
            0.878423345409431,
            0.9948462247926801,
            0.9999734386011127
          ]
        },
        {
          "name": "gamma=0.99",
          "values": [
            0.0,
            0.1,
            0.19900000000000012,
            0.4900995010000005,
            0.9561792499119551,
            1.8209306240276908,
            3.9499393286246334,
            6.339676587267704
          ]
        }
      ]
    },
    {
      "id": "importance-variance",
      "kind": "analytical_curve",
      "caption": "Exact author-constructed one-step importance-estimator variance, fixed reward one, target action a1, behavior probability epsilon>0. No trajectory benchmark.",
      "x": {
        "name": "Behavior probability epsilon",
        "values": [
          0.01,
          0.02,
          0.05,
          0.1,
          0.2,
          0.5,
          1.0
        ]
      },
      "series": [
        {
          "name": "One-observation estimator variance",
          "values": [
            99.0,
            49.0,
            19.0,
            9.0,
            4.0,
            1.0,
            0.0
          ]
        }
      ]
    }
  ],
  "cover_prompt": "Premium scientific editorial cover, landscape 3:2, no words. A polished jade path is densely supported by warm ivory archival tiles and ends at a fork: one golden bridge rests on visible evidence columns while a second translucent bridge reaches into deep emerald mist with its far supports absent. Fine brass decision geometry, quiet precise architecture, editorial realism, generous negative space. Conceptual offline-data support and unknown actions, not a measured reward chart.",
  "numeric_data_provenance": "Computed and checked by the corresponding standalone standard-library script; constructed, not empirical."
}