Perturbation Robustness (External)

Available

llm.perturbation_robustness_external

Measures how consistently the system stays correct across meaning-preserving variants of a question, from a pre-scored dataset with off-platform SUT responses.

Contract

Field	Value
version	1.0.0
metric_type	pointwise
scorer_contract	full_dataset
direction	higher_is_better
entrypoint	aip_metrics_llm.unanswered.perturbation_robustness_external
target_kind	None

Required columns

• input_id
• prompt
• sut_response

Accepted schemas

[
  {
    "name": "gdi_text_v1",
    "task_types": [
      "multi_turn_llm",
      "multi_turn_rag",
      "single_turn_llm",
      "single_turn_rag"
    ]
  }
]

Methodology

1. This metric is pure scoring; it does not generate perturbations or call the system under test. Generation lives in the aip_sdg_llm.augmentation.perturbation_robustness_external SDG recipe; SUT inference happens off-platform.
2. Rows are grouped by metadata.original_query; the baseline row (query equals original_query) and its perturbed variants are identified within each group.
3. The judge scores base correctness, perturbed correctness, and a per-pair robustness score, reusing the hosted perturbation_robustness metric's judges and aggregation.
4. Combine as 0.5 times mean robustness plus 0.5 times max(0, 1 minus mean correctness drop); the group score is replicated to every row in the group.

Score semantics

Scores range 0-1, blending consistency across variants with the drop in correctness. A high score means answers stay correct and stable under rewording; a low score means correctness degrades. Higher is better.

Worked example

Mean robustness 0.9 with a small correctness drop of 0.1 -> 0.50.9 + 0.50.9 = 0.9.

Configuration schema

{
  "properties": {
    "groundtruth_context": {
      "anyOf": [
        {
          "type": "string"
        },
        {
          "type": "null"
        }
      ],
      "default": null,
      "title": "Groundtruth Context"
    },
    "inverted": {
      "default": false,
      "title": "Inverted",
      "type": "boolean",
      "x-aip-param-role": "scoring_metadata"
    },
    "max": {
      "default": 1.0,
      "title": "Max",
      "type": "number",
      "x-aip-param-role": "scoring_metadata"
    },
    "metric_name": {
      "default": "perturbation_robustness_external",
      "title": "Metric Name",
      "type": "string",
      "x-aip-param-role": "scoring_metadata"
    },
    "min": {
      "default": 0.0,
      "title": "Min",
      "type": "number",
      "x-aip-param-role": "scoring_metadata"
    },
    "reference_answer": {
      "anyOf": [
        {
          "type": "string"
        },
        {
          "type": "null"
        }
      ],
      "default": null,
      "title": "Reference Answer"
    },
    "sub_call_concurrency_limit": {
      "default": 5,
      "maximum": 64,
      "minimum": 1,
      "title": "Sub Call Concurrency Limit",
      "type": "integer",
      "x-aip-param-role": "operational"
    },
    "timeout_seconds": {
      "default": 60.0,
      "exclusiveMinimum": 0,
      "title": "Timeout Seconds",
      "type": "number",
      "x-aip-param-role": "operational"
    },
    "weight": {
      "default": 1.0,
      "title": "Weight",
      "type": "number",
      "x-aip-param-role": "scoring_metadata"
    }
  },
  "title": "PerturbationRobustnessExternalConfig",
  "type": "object"
}

Execution

{
  "emits_metric_family": false,
  "function_name": null,
  "max_concurrency": 8,
  "processing_kind": "network",
  "stream_batch_size": null,
  "timeout_seconds": 900
}

Complete manifest

accepts:
- name: gdi_text_v1
  task_types:
  - multi_turn_llm
  - multi_turn_rag
  - single_turn_llm
  - single_turn_rag
config_schema:
  properties:
    groundtruth_context:
      anyOf:
      - type: string
      - type: 'null'
      default: null
      title: Groundtruth Context
    inverted:
      default: false
      title: Inverted
      type: boolean
      x-aip-param-role: scoring_metadata
    max:
      default: 1.0
      title: Max
      type: number
      x-aip-param-role: scoring_metadata
    metric_name:
      default: perturbation_robustness_external
      title: Metric Name
      type: string
      x-aip-param-role: scoring_metadata
    min:
      default: 0.0
      title: Min
      type: number
      x-aip-param-role: scoring_metadata
    reference_answer:
      anyOf:
      - type: string
      - type: 'null'
      default: null
      title: Reference Answer
    sub_call_concurrency_limit:
      default: 5
      maximum: 64
      minimum: 1
      title: Sub Call Concurrency Limit
      type: integer
      x-aip-param-role: operational
    timeout_seconds:
      default: 60.0
      exclusiveMinimum: 0
      title: Timeout Seconds
      type: number
      x-aip-param-role: operational
    weight:
      default: 1.0
      title: Weight
      type: number
      x-aip-param-role: scoring_metadata
  title: PerturbationRobustnessExternalConfig
  type: object
dependencies: []
description: How consistently does the system stay correct across equivalent question
  variants, scored from off-platform SUT responses?
direction: higher_is_better
display_name: Perturbation Robustness (External)
entrypoint: aip_metrics_llm.unanswered.perturbation_robustness_external
execution:
  emits_metric_family: false
  function_name: null
  max_concurrency: 8
  processing_kind: network
  stream_batch_size: null
  timeout_seconds: 900
kind: metric
manifest_version: '1'
max_prompt_slots: 0
metric_metadata:
  methodology:
  - This metric is pure scoring; it does not generate perturbations or call the system
    under test. Generation lives in the aip_sdg_llm.augmentation.perturbation_robustness_external
    SDG recipe; SUT inference happens off-platform.
  - Rows are grouped by metadata.original_query; the baseline row (query equals original_query)
    and its perturbed variants are identified within each group.
  - The judge scores base correctness, perturbed correctness, and a per-pair robustness
    score, reusing the hosted perturbation_robustness metric's judges and aggregation.
  - Combine as 0.5 times mean robustness plus 0.5 times max(0, 1 minus mean correctness
    drop); the group score is replicated to every row in the group.
  score_semantics: Scores range 0-1, blending consistency across variants with the
    drop in correctness. A high score means answers stay correct and stable under
    rewording; a low score means correctness degrades. Higher is better.
  summary: Measures how consistently the system stays correct across meaning-preserving
    variants of a question, from a pre-scored dataset with off-platform SUT responses.
  worked_example: Mean robustness 0.9 with a small correctness drop of 0.1 -> 0.5*0.9
    + 0.5*0.9 = 0.9.
metric_type: pointwise
name: llm.perturbation_robustness_external
partition_types: []
required_columns:
- input_id
- prompt
- sut_response
required_kinds: []
scorer_contract: full_dataset
target_kind: null
unsupported_trace_shapes: []
version: 1.0.0