openenv.yaml

name: synapse-x
version: "1.0.0"
description: >
  SYNAPSE-X is a reproducible benchmark for real-world operational decision-making
  under uncertainty, deadlines, and constrained resources. It models choices an
  autonomous agent faces during incident response, escalation handling, and
  moderation triage: not just what has highest value, but what should be done now,
  with limited time, imperfect information, and failure spillover across dependent tasks.

domain: decision-intelligence
tags:
  - openenv
  - scheduling
  - risk-management
  - uncertainty
  - resource-allocation
  - real-world
  - content-moderation
  - incident-response

author: "SYNAPSE-X Team"
license: MIT
entry_point: env.environment:SynapseXEnvironment

# ---------------------------------------------------------------------------
# Environment class
# ---------------------------------------------------------------------------
environment:
  class: SynapseXEnvironment
  module: env.environment

  observation:
    type: Observation
    description: >
      A snapshot of all tasks and episode-level state. Tasks are sorted by id.
    fields:
      tasks:
        type: List[Task]
        description: Sorted list of tasks with risk, deadlines, ECHO predictions and cascade state
      time:
        type: int
        range: [0, 30]
        description: Current episode timestep
      resources:
        type: float
        range: [0.0, 1.0]
        description: Available resource pool (normalized)
      episode_done:
        type: bool
        description: True when episode has terminated

  action:
    type: Action
    description: JSON object selecting an action type and target task id
    fields:
      action_type:
        type: str
        choices: [execute, delay, reallocate]
        description: >
          execute — attempt task completion via PRISM;
          delay   — defer work, reduce deadline, accumulate delay penalty;
          reallocate — recover up to 0.2 resources before a future execution
      task_id:
        type: int
        description: ID of the target task (matches Task.id in observation)

  reward:
    type: float
    description: >
      Dense, continuous, multi-objective reward. Clamped to [-2.0, 2.0] per step.
      Terminal episode bonus of +2.0 when all tasks complete successfully.
      Grader score is always in [0.0, 1.0].
    range: [-2.0, 2.0]
    terminal_bonus: 2.0

# ---------------------------------------------------------------------------
# Tasks with agent graders
# ---------------------------------------------------------------------------
tasks:
  - name: easy
    description: >
      Low-risk tasks with generous deadlines. Tests basic execution ordering
      and resource management. Baseline score ~ 0.978.
    difficulty: easy
    num_tasks: 3
    grader_range: [0.0, 1.0]
    reward_bounds: {min: -10.0, max: 8.0}
    seed: 42

  - name: medium
    description: >
      Moderate risk and tighter deadlines. One task is intentionally designed
      to likely fail under canonical seed — models realistic partial-completion
      scenarios. Baseline score ~ 0.679.
    difficulty: medium
    num_tasks: 3
    grader_range: [0.0, 1.0]
    reward_bounds: {min: -6.0, max: 7.0}
    seed: 42

  - name: hard
    description: >
      High uncertainty, tight deadlines, and CASCADE-X dependency propagation.
      Tasks form a dependency chain (task 1 depends on task 0; task 2 depends
      on both). Failed upstream tasks amplify risk across all dependents via
      phase transitions. Requires dependency-aware planning under compounding
      pressure. Baseline score ~ 0.988.
    difficulty: hard
    num_tasks: 3
    grader_range: [0.0, 1.0]
    reward_bounds: {min: -10.0, max: 7.0}
    seed: 19

  - name: triage
    description: >
      Wave-based content moderation with no task dependencies. Five moderation
      items (abuse escalations, spam waves, graphic content, VIP appeals) arrive
      in three release waves. Tests pure prioritization under time pressure.
      Baseline score ~ 1.000.
    difficulty: triage
    num_tasks: 5
    grader_range: [0.0, 1.0]
    reward_bounds: {min: -8.0, max: 8.0}
    seed: 31

# ---------------------------------------------------------------------------
# API surface
# ---------------------------------------------------------------------------
api:
  base_url: "http://localhost:7860"
  endpoints:
    reset:
      method: POST
      path: /reset
      returns: Observation
    step:
      method: POST
      path: /step
      returns: StepResult
    state:
      method: GET
      path: /state
      returns: dict
    grade:
      method: POST
      path: /grade
      returns: GradeResult
    health:
      method: GET
      path: /health
    validate:
      method: GET
      path: /validate
    tasks:
      method: GET
      path: /tasks

# ---------------------------------------------------------------------------
# Inference script configuration
# ---------------------------------------------------------------------------
inference:
  script: inference.py
  tasks_evaluated: [easy, medium, hard, triage]
  env_vars:
    required:
      - API_BASE_URL
      - MODEL_NAME
    optional:
      - HF_TOKEN
      - API_KEY
      - LOCAL_IMAGE_NAME
      - SEED
  runtime_limit_minutes: 20
  compute: vcpu=2, memory=8gb
  fallback: baseline-fallback  # used when HF_TOKEN is absent

# ---------------------------------------------------------------------------
# Observation and action space summary (for openenv validate)
# ---------------------------------------------------------------------------
observation_space:
  tasks: list
  time: int
  resources: float
  episode_done: bool

action_space:
  action_type: string  # execute | delay | reallocate
  task_id: int