prompts/context_ladder.json

{
  "levels": [
    {
      "name": "L1_focused",
      "max_tokens": 4096,
      "corpus_filter": "relevant",
      "exhibit_filter": "relevant",
      "defendants_filter": "relevant",
      "cross_count": false,
      "description": "Performance ceiling — clean signal, no distractors"
    },
    {
      "name": "L2_exhibit_noise",
      "max_tokens": 4096,
      "corpus_filter": "relevant",
      "exhibit_filter": "all",
      "defendants_filter": "relevant",
      "cross_count": false,
      "description": "Exhibit selection pressure — can the model ignore irrelevant exhibits?"
    },
    {
      "name": "L3_full_noise",
      "max_tokens": 4096,
      "corpus_filter": "all",
      "exhibit_filter": "all",
      "defendants_filter": "all",
      "cross_count": false,
      "description": "Maximum context noise — both corpus and exhibits are saturated"
    },
    {
      "name": "L4_cross_count",
      "max_tokens": 4096,
      "corpus_filter": "all",
      "exhibit_filter": "all",
      "defendants_filter": "all",
      "cross_count": true,
      "description": "Cross-count confusion — competing legal theories in same prompt"
    }
  ]
}