tugrapaydiner.github.io

{
  "name": "read_dispute",
  "kind": "answer",
  "impl": "imperative",
  "description": "Summarizes the active evaluation dispute: both labs' conclusions with scores and intervals, benchmark identity, evidence coverage, integrity status, and the protocol dimensions where the labs differ.",
  "inputSchema": {
    "type": "object",
    "properties": {
      "disputeId": {
        "type": "string",
        "description": "dispute id from read_dispute; omit to use the single canonical dispute"
      }
    },
    "required": [],
    "additionalProperties": false
  },
  "page": "/mpw/"
}

{
  "name": "run_counterfactual",
  "kind": "act",
  "impl": "imperative",
  "description": "Builds one hybrid protocol from a lab baseline by adopting selected dimensions from the other lab, then runs the deterministic synthetic evaluation with paired statistics.",
  "inputSchema": {
    "type": "object",
    "properties": {
      "disputeId": {
        "type": "string",
        "description": "dispute id from read_dispute; omit to use the single canonical dispute"
      },
      "baseLab": {
        "type": "string",
        "description": "lab whose protocol is the starting point; the target conclusion is always the other lab's",
        "enum": [
          "A",
          "B"
        ]
      },
      "adopt": {
        "type": "array",
        "description": "dimensions taken from the other lab; empty re-runs the baseline as a control: exposed protocol dimension: reasoning_budget (thinking effort per item), answer_parser (strictness of answer parsing), retry_policy (whether one retry is allowed), tool_access (breadth of tool use)",
        "items": {
          "type": "string",
          "enum": [
            "answer_parser",
            "reasoning_budget",
            "retry_policy",
            "tool_access"
          ]
        },
        "minItems": 0,
        "maxItems": 4,
        "uniqueItems": true
      }
    },
    "required": [
      "baseLab",
      "adopt"
    ],
    "additionalProperties": false
  },
  "page": "/mpw/"
}

{
  "name": "inspect_evidence",
  "kind": "answer",
  "impl": "imperative",
  "description": "Reads stored diagnostics for a completed experiment: coverage, paired outcome counts, parser, retry and tool behavior, category summary, and evidence identity.",
  "inputSchema": {
    "type": "object",
    "properties": {
      "experimentId": {
        "type": "string",
        "description": "experiment id returned by run_counterfactual"
      },
      "category": {
        "type": "string",
        "description": "optional benchmark stratum; omit for all strata",
        "enum": [
          "multi-step-reasoning",
          "quantitative-reasoning",
          "instruction-following",
          "tool-reasoning"
        ]
      },
      "limit": {
        "type": "integer",
        "description": "sample rows to return, 1 to 20, default 5",
        "minimum": 1,
        "maximum": 20,
        "default": 5
      }
    },
    "required": [
      "experimentId"
    ],
    "additionalProperties": false
  },
  "page": "/mpw/"
}

{
  "name": "verify_witness",
  "kind": "act",
  "impl": "imperative",
  "description": "Checks every exposed protocol subset and determines whether a proposed candidate reproduces the other lab's conclusion with globally minimum cardinality. Returns all co-minimum witnesses and a request-bound certificate only on VERIFIED.",
  "inputSchema": {
    "type": "object",
    "properties": {
      "disputeId": {
        "type": "string",
        "description": "dispute id from read_dispute; omit to use the single canonical dispute"
      },
      "baseLab": {
        "type": "string",
        "description": "lab whose protocol is the starting point; the target conclusion is always the other lab's",
        "enum": [
          "A",
          "B"
        ]
      },
      "candidate": {
        "type": "array",
        "description": "proposed witness subset; empty is a legal vacuous check: exposed protocol dimension: reasoning_budget (thinking effort per item), answer_parser (strictness of answer parsing), retry_policy (whether one retry is allowed), tool_access (breadth of tool use)",
        "items": {
          "type": "string",
          "enum": [
            "answer_parser",
            "reasoning_budget",
            "retry_policy",
            "tool_access"
          ]
        },
        "minItems": 0,
        "maxItems": 4,
        "uniqueItems": true
      }
    },
    "required": [
      "baseLab",
      "candidate"
    ],
    "additionalProperties": false
  },
  "page": "/mpw/"
}