Oversight-Arena/openenv.yaml at main · Aparna-Singha/Oversight-Arena · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
name: OversightArena
version: "1.0.0"
description: >
  An RL environment where an LLM oversight agent reviews worker AI answers
  about structured JSON records and flags errors (wrong_value, wrong_inference,
  omission). Trains accurate, calibrated oversight behaviour.

base_url: http://localhost:8000

endpoints:
  reset:
    method: POST
    path: /reset
    request_schema:
      type: object
      properties:
        seed:
          type: integer
          nullable: true
    response_schema: OversightObservation

  step:
    method: POST
    path: /step
    request_schema:
      type: object
      required: [episode_id, action]
      properties:
        episode_id:
          type: string
        action:
          $ref: "#/schemas/OversightAction"
    response_schema: StepResponse

  observation_space:
    method: GET
    path: /observation_space

  action_space:
    method: GET
    path: /action_space

  health:
    method: GET
    path: /health

schemas:
  OversightObservation:
    type: object
    properties:
      source_json: { type: object }
      questions: { type: array, items: { type: string } }
      worker_answers: { type: array, items: { type: string } }
      step_number: { type: integer }
      flags_used: { type: integer }
      flags_remaining: { type: integer }
      episode_id: { type: string }
      done: { type: boolean }
      message: { type: string }

  OversightAction:
    type: object
    required: [action_type, question_id, reasoning, confidence]
    properties:
      action_type:
        type: string
        enum: [approve, flag]
      question_id:
        type: integer
      error_type:
        type: string
        enum: [wrong_value, wrong_inference, omission]
        nullable: true
      reasoning:
        type: string
      confidence:
        type: number
        minimum: 0.0
        maximum: 1.0

reward:
  range: [-2.0, 2.0]
  true_positive: 2.0
  true_negative: 1.0
  false_positive: -1.0
  false_negative: -2.0
  budget_exceeded: -0.5

episode:
  steps_per_episode: 5
  flag_budget: 3
  termination: all_questions_reviewed

task_types:
  - employee_record
  - product_record

error_types:
  - wrong_value
  - wrong_inference
  - omission