{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "title": "S1CogBench public aggregate result row",
  "type": "object",
  "required": [
    "row_id",
    "task",
    "model",
    "protocol",
    "mode",
    "metric",
    "metric_variant",
    "scope",
    "target",
    "unit",
    "aggregation",
    "n",
    "value",
    "value_raw",
    "value_status",
    "status",
    "source"
  ],
  "properties": {
    "row_id": {
      "type": "string"
    },
    "task": {
      "type": "string"
    },
    "model": {
      "type": "string"
    },
    "protocol": {
      "type": "string"
    },
    "mode": {
      "type": "string"
    },
    "metric": {
      "type": "string"
    },
    "metric_variant": {
      "type": "string"
    },
    "scope": {
      "type": "string"
    },
    "target": {
      "type": "string"
    },
    "unit": {
      "type": "string"
    },
    "aggregation": {
      "type": "string"
    },
    "value_raw": {
      "type": "string"
    },
    "n_raw": {
      "type": "string"
    },
    "status": {
      "type": "string",
      "description": "Row state. Accepted cells: available_full_protocol and the complete_* source states. Table 3 cells without a value: not_applicable_readout (n.a.: no option probabilities for an NLL target), partial_records_retained (part.), not_evaluated (n.e.: outside the measured scope), non_full_prefix_only (n.f.), cancelled_before_generation (canc.), all_responses_invalid (inv.). Local runs of python -m s1cogbench: user_supplied_full_protocol, user_supplied_incomplete."
    },
    "notes": {
      "type": "string"
    },
    "n": {
      "type": [
        "integer",
        "null"
      ]
    },
    "value": {
      "oneOf": [
        {
          "type": "number"
        },
        {
          "enum": [
            "Infinity",
            "-Infinity"
          ]
        },
        {
          "type": "null"
        }
      ]
    },
    "value_status": {
      "enum": [
        "finite",
        "defined_positive_infinity",
        "undefined"
      ]
    },
    "source": {
      "type": "object"
    },
    "runtime_context": {
      "type": "object"
    },
    "common_metric": {
      "type": "boolean"
    },
    "steps": {
      "type": "integer"
    },
    "table2_marker": {
      "description": "The marker printed in the Table 3 cell when the row has no value; null otherwise.",
      "enum": [
        "n.a.",
        "part.",
        "n.e.",
        "n.f.",
        "canc.",
        "inv.",
        null
      ]
    }
  },
  "additionalProperties": true,
  "comparison_context": [
    "task",
    "protocol",
    "mode",
    "metric",
    "metric_variant",
    "scope",
    "target",
    "unit",
    "aggregation"
  ],
  "notes": [
    "Context keys define comparable views; do not collapse distinct protocols, reward families, horizons, units or targets.",
    "Latency additionally requires matching runtime_context and notes. Shared-GPU timings are not isolated model speed or human reaction time.",
    "n is the original aggregation denominator and can mean responses, participants or episodes; interpret with aggregation.",
    "No overall cognitive score or cross-task normalized ranking is supplied.",
    "A row with a table2_marker has value null and value_status undefined; its status names the marker. Do not read it as a zero or a missing run of an applicable target.",
    "API rows (model_access api in results.json) keep the provider readout in runtime_context.readout: native typed choice (Jev), generated FINAL line (Claude; human-NLL targets n.a.), or stated probabilities (Claude Haiku on four modes). Stated probabilities are not internal model probabilities."
  ]
}