Przejdź do treści

Eval Report v1

Source schema: doc/schemas/eval-report.v1.schema.json

Machine-readable schema for adapter evaluation outputs that gate validation, deployment, or rejection.

Governing Basis

Project Lineage

Requirements

Stories

Fields

Field Required Shape Description
schema/v yes const: 1 Schema version.
eval-report/id yes string Stable identifier of the evaluation report.
subject/kind yes enum: training-job, adapter-artifact Evaluated subject class.
subject/ref yes string Reference to the evaluated job or adapter.
base-model/ref yes string Immutable base model against which the evaluated artifact must remain interpretable.
adapter/hash no string Immutable adapter hash when the evaluated subject is an adapter artifact.
generated-at yes string Evaluation completion timestamp.
verdict yes enum: pass, conditional-pass, fail High-level evaluation verdict.
summary yes string Human-readable summary of the evaluation outcome.
evaluator/refs yes array People, nodes, or policy engines that produced or signed the report.
policy/profile no string Evaluation policy profile used for this report.
known-issues no array Residual issues kept visible even when the verdict is not a hard fail.
suites yes array Evaluation suites covering quality, regression, and risk gates.
policy_annotations no object Optional implementation-local annotations that do not change the core evaluation semantics.

Conditional Rules

Rule 1

Constraint:

{
  "properties": {
    "suites": {
      "contains": {
        "type": "object",
        "properties": {
          "suite/class": {
            "const": "quality"
          }
        },
        "required": [
          "suite/class"
        ]
      }
    }
  }
}

Rule 2

Constraint:

{
  "properties": {
    "suites": {
      "contains": {
        "type": "object",
        "properties": {
          "suite/class": {
            "const": "regression"
          }
        },
        "required": [
          "suite/class"
        ]
      }
    }
  }
}

Rule 3

Constraint:

{
  "properties": {
    "suites": {
      "contains": {
        "type": "object",
        "properties": {
          "suite/class": {
            "const": "risk"
          }
        },
        "required": [
          "suite/class"
        ]
      }
    }
  }
}

Rule 4

When:

{
  "properties": {
    "subject/kind": {
      "const": "adapter-artifact"
    }
  },
  "required": [
    "subject/kind"
  ]
}

Then:

{
  "required": [
    "adapter/hash"
  ]
}

Field Semantics

schema/v

  • Required: yes
  • Shape: const: 1

Schema version.

eval-report/id

  • Required: yes
  • Shape: string

Stable identifier of the evaluation report.

subject/kind

  • Required: yes
  • Shape: enum: training-job, adapter-artifact

Evaluated subject class.

subject/ref

  • Required: yes
  • Shape: string

Reference to the evaluated job or adapter.

base-model/ref

  • Required: yes
  • Shape: string

Immutable base model against which the evaluated artifact must remain interpretable.

adapter/hash

  • Required: no
  • Shape: string

Immutable adapter hash when the evaluated subject is an adapter artifact.

generated-at

  • Required: yes
  • Shape: string

Evaluation completion timestamp.

verdict

  • Required: yes
  • Shape: enum: pass, conditional-pass, fail

High-level evaluation verdict.

summary

  • Required: yes
  • Shape: string

Human-readable summary of the evaluation outcome.

evaluator/refs

  • Required: yes
  • Shape: array

People, nodes, or policy engines that produced or signed the report.

policy/profile

  • Required: no
  • Shape: string

Evaluation policy profile used for this report.

known-issues

  • Required: no
  • Shape: array

Residual issues kept visible even when the verdict is not a hard fail.

suites

  • Required: yes
  • Shape: array

Evaluation suites covering quality, regression, and risk gates.

policy_annotations

  • Required: no
  • Shape: object

Optional implementation-local annotations that do not change the core evaluation semantics.