DGXAI/driftcall

1# openenv.yaml — consumed by `openenv validate`2# Schema source: https://github.com/meta-pytorch/OpenEnv (v1.0).3# Deploy spec: docs/modules/deploy_env_space.md §4.3.4schema_version: "1.0"5 6env:7  id: driftcall8  version: "0.1.0"9  display_name: "DriftCall — Indic Voice Concierge under Schema Drift"10  description: >11    OpenEnv-compliant RL environment where a voice-first agent completes Indic12    consumer concierge tasks while vendor APIs undergo mid-episode schema,13    policy, T&C, pricing, and auth drift. Five independent reward components;14    deterministic seeded drift; Hindi/Tamil/Kannada/Hinglish briefs via15    Kokoro TTS + faster-whisper ASR.16  license: apache-2.017  tags:18    - openenv19    - rl20    - voice21    - indic22    - schema-drift23    - grpo24 25  entrypoint:26    type: http27    base_url: "https://driftcall-driftcall-env.hf.space"28    endpoints:29      reset: "/reset"30      step: "/step"31      state: "/state"32      close: "/close"33      health: "/healthz"34    auth:35      type: bearer36      secret_env: DRIFTCALL_ENV_TOKEN37 38  action_space:39    ref: "cells.step_04_models:DriftCallAction"40 41  observation_space:42    ref: "cells.step_04_models:DriftCallObservation"43 44  episode:45    max_turns: 1646    reset_config:47      seed:48        type: int49        required: false50      curriculum_stage:51        type: int52        range: [1, 3]53        required: false54      language_weights:55        type: object56        required: false57      audio_boundary_enabled:58        type: bool59        required: false60 61  reward:62    shape: scalar63    range: [-1.0, 1.0]64    # The reward function lives in `cells/step_08_rewards.py`. Five independent65    # components are computed at episode termination; combined into a quality66    # score, calibrated by a Brier penalty + uncertain floor, then clamped.67    # Implementation entrypoint:68    impl: "cells.step_08_rewards:compute_rewards"69    pipeline:70      - "cells.step_08_rewards:combine_quality"   # weighted mix of R1..R571      - "cells.step_08_rewards:brier_penalty"     # confidence calibration72      - "cells.step_08_rewards:apply_uncertain_floor"  # 0.50 floor when uncertain73      - "cells.step_08_rewards:final_reward"      # final scalar in [-1, 1]74    components:75      - id: R176        name: task_completion77        weight: 0.4078        impl: "cells.step_08_rewards:task_completion"79        description: >80          Goal achieved (correct booking, payment success, vendor confirmation).81      - id: R282        name: drift_detection83        weight: 0.2084        impl: "cells.step_08_rewards:drift_detection"85        description: >86          Agent detects mid-episode schema/policy/auth drift and adapts.87      - id: R388        name: constraint_adherence89        weight: 0.2090        impl: "cells.step_08_rewards:constraint_adherence"91        description: >92          Honours user constraints (budget, time window, dietary, lang).93      - id: R494        name: format_compliance95        weight: 0.1096        impl: "cells.step_08_rewards:format_compliance"97        description: >98          Tool args parse cleanly against the (possibly drifted) schema.99      - id: R5100        name: anti_hack_penalty101        weight: 0.10102        impl: "cells.step_08_rewards:anti_hack_penalty"103        description: >104          Penalty for known reward-hacking patterns flagged in probe set.105    docs: "docs/modules/rewards.md"106