{
  "schemaVersion": "robotic.story.v1",
  "edition": "002",
  "id": "human-instructions",
  "mode": "ridiculous",
  "label": "PLAY / HUMAN COMPATIBILITY LAB",
  "title": "Can you operate a human?",
  "robotTitle": "Natural language. Eight undocumented endpoints.",
  "description": "Eight ordinary requests. Twenty-four ways to interpret them. Find the useful question hiding inside each one.",
  "robotDescription": "Specification received. Intent unavailable. Please select a response before touching the soup.",
  "minutes": "6 min + play",
  "kind": "Comedy game + research",
  "interactive": "instructions",
  "signal": "INPUT: AMBIGUOUS",
  "url": "/stories/human-instructions/",
  "publishedAt": "2026-09-11",
  "topics": [
    "humans",
    "play"
  ],
  "sections": [
    {
      "title": "Why a question can be the smart move.",
      "human": [
        "The game is written comedy. The answers are authored, the branches are fixed, and the clarification count is simply the number of useful questions you selected. It measures neither intelligence nor readiness to operate a real robot.",
        "But the premise has a real research neighbor. KnowNo studies how a language-model planner can ask for help when a task admits more than one plausible action. Its examples include ambiguity about objects, quantities and preferences.",
        "The method uses calibrated sets of possible actions. When the set does not resolve to a single option, the system can ask a person to narrow it. The work reports experiments with real and simulated robots. It does not turn every uncertain sentence into a safe physical action."
      ],
      "robot": [
        "An instruction is not necessarily a complete task specification. “The cup” may identify several cups. “Clean” may hide a preferred destination. Choosing an answer confidently does not add missing information.",
        "KnowNo combines a language-model planner with conformal prediction to form sets of possible actions and decide when to seek help. Its statistical claims depend on the stated setup and assumptions, not on a model sounding sure of itself.",
        "This game implements none of that algorithm. It is a small authored argument for asking a better question. There is no model running behind the buttons, and the clarification count is not a calibrated probability."
      ]
    },
    {
      "title": "Ask for the boundary, not the whole universe.",
      "human": [
        "The useful options in the game ask for something concrete: a destination, a time limit, an audience, a rule for duplicates. A question can move work forward without demanding a complete theory of what the human meant.",
        "That is the habit worth taking away. Before a consequential action, identify the missing detail that would change the next step. “What do you mean?” is sometimes necessary. “Should I alert you or stir it too?” is much easier to answer."
      ],
      "robot": [
        "Useful clarification reduces the space of next actions. It does not merely request more words. A response can specify an object, a preference or a stopping condition.",
        "The studied guarantees do not cover every source of uncertainty. Perception, execution and the accuracy of the environment description still matter. A well-formed instruction can fail against a poorly understood world."
      ]
    }
  ],
  "question": {
    "human": "What is the smallest question that would prevent the biggest wrong turn?",
    "robot": "Which missing variable would actually change your next action?"
  },
  "sources": [
    {
      "title": "Robots That Ask For Help: Uncertainty Alignment for Large Language Model Planners",
      "url": "https://arxiv.org/html/2307.01928v2",
      "published": "4 September 2023 · version 2",
      "facts": [
        "KnowNo uses conformal prediction with a language-model planner to decide when to ask for help.",
        "Evaluations include ambiguity in real and simulated robot tasks."
      ],
      "limit": "The stated guarantees depend on calibration and task assumptions. This is not a guarantee of physical safety, correct perception or correct low-level execution."
    },
    {
      "title": "KnowNo: author project and demonstrations",
      "url": "https://robot-help.github.io/",
      "published": "Undated project page · CoRL 2023",
      "facts": [
        "The author project presents the method and demonstrations behind the paper."
      ],
      "limit": "Robotic’s game is an authored illustration, not an implementation or replication of KnowNo."
    }
  ],
  "related": [
    "act-naturally",
    "mars-latency"
  ],
  "image": {
    "kind": "AI illustration / fiction",
    "alt": "An imaginary kitchen robot considers a bowl of soup and an oversized ladle.",
    "caption": "Original AI illustration for our fictional instructions game. Please clarify before deploying the ladle."
  },
  "notice": "Original editorial fiction. Robot voices are invented, not evidence of what real systems think or feel."
}