{
  "title": "Two illustrative prompts through one finite-choice application",
  "origin": "This authored example illustrates a Jev-style finite-choice interface: a selected option and a probability for each option. The stated toy weights produce the intent probabilities. The two next-token rows are authored separately. No live model was called and no evaluation job ran.",
  "intent_rule": {
    "base_weights": {
      "retrieval": 1,
      "other": 1,
      "unspecified": 3
    },
    "named_title_bonus": 15,
    "normalization": "Divide each weight by their sum to obtain the displayed probability. The highest-scoring candidate is proposed.",
    "named_titles": {
      "retrieval": "retrieval eval",
      "other": "ranking eval"
    }
  },
  "application_rule": "Ask which eval when the result is unspecified. A named result may resume its saved checkpoint when the requested title matches and a checkpoint exists.",
  "checkpoint": {
    "eval": "retrieval eval",
    "completed_batch": 12,
    "next_batch": 13
  },
  "text_generation": "The two illustrative iterations use specified toy next-token probabilities conditioned on the prompt, app result and any earlier chosen token. Greedy decoding takes the largest probability in each row. A leading space is part of a token. Percentages in the finite-choice table have a different meaning from the token distributions.",
  "cases": [
    {
      "id": "ambiguous",
      "conversation": [
        "Resume yesterday’s run"
      ],
      "prompt": "Resume yesterday’s run",
      "intent_weights": {
        "retrieval": 1,
        "other": 1,
        "unspecified": 3
      },
      "intent_probabilities": {
        "retrieval": 0.2,
        "other": 0.2,
        "unspecified": 0.6
      },
      "proposed": "unspecified",
      "app_action": "Ask which eval",
      "token_iterations": [
        {
          "iteration": 1,
          "prefix_before": "",
          "candidates": {
            "Which": 0.6,
            "Can": 0.25,
            "I": 0.15
          },
          "selected": "Which",
          "prefix_after": "Which"
        },
        {
          "iteration": 2,
          "prefix_before": "Which",
          "candidates": {
            " eval": 0.7,
            " one": 0.2,
            " run": 0.1
          },
          "selected": " eval",
          "prefix_after": "Which eval"
        }
      ]
    },
    {
      "id": "clarified",
      "conversation": [
        "Resume yesterday’s run",
        "I meant the retrieval eval"
      ],
      "prompt": "I meant the retrieval eval",
      "intent_weights": {
        "retrieval": 16,
        "other": 1,
        "unspecified": 3
      },
      "intent_probabilities": {
        "retrieval": 0.8,
        "other": 0.05,
        "unspecified": 0.15
      },
      "proposed": "retrieval",
      "app_action": "Resume checkpoint",
      "token_iterations": [
        {
          "iteration": 1,
          "prefix_before": "",
          "candidates": {
            "Resuming": 0.7,
            "Opening": 0.2,
            "I": 0.1
          },
          "selected": "Resuming",
          "prefix_after": "Resuming"
        },
        {
          "iteration": 2,
          "prefix_before": "Resuming",
          "candidates": {
            " the": 0.65,
            " retrieval": 0.25,
            " your": 0.1
          },
          "selected": " the",
          "prefix_after": "Resuming the"
        }
      ]
    }
  ]
}
