Inputs for computing rewards.

Metrics extracted from trajectory for archetype-aware scoring.

Complete training pipeline orchestrator. Manages the full workflow from agent simulation to model training.

Builds training datasets from trajectories with multi-prompt handling. This creates separate training data for each prompt type (reasoning,

Data for a single turn in a multi-turn episode. Contains all information needed for training on this turn, including computed advantages

Canonical project pipeline orchestrator.

Converts Babylon trajectories to Atropos format. Args: dropout_rate: Random dropout rate for data augmentation (0.0-0.5) max

Configuration for HuggingFace dataset reader.

Configuration for managed services

Reads Babylon trajectories from a HuggingFace dataset. Provides the same interface as PostgresTrajectoryReader for compatibility with th

Complete scenario for agent rollout. Contains all information an agent needs to make decisions: - Market state (prediction markets, perp

Manages background services for local training. Provides automatic startup, health checking, and cleanup of: - Atropos API server (for G

Orchestrates the complete training pipeline. Manages: - Service lifecycle (Atropos API, vLLM) - Environment server - GRPO traine

Manages scenario sampling for online rollouts. Features: - Load production snapshots from database - Generate synthetic scenarios

Result of executing an action

Result of executing an action

Credit assignment for a decision with delayed outcome. When a trade is made, the actual P&L may not be known until later (e.g., position

Configuration for Atropos GRPO training

Mock tokenizer for testing

Configuration for scenario pool

State of a single market

Configuration for post-training actions.

Ground-truth metadata about who the agent is interacting with. Mirrors the TypeScript ``CounterpartyContext`` interface in ``packages/agents/

Environment state at a given point

Reads Babylon trajectories from a local directory of JSON files.

A news item in the scenario

Single LLM call record. Matches the TypeScript LLMCall interface in plugin-trajectory-logger/types.ts

Executes actions against simulated game state. For training, this simulates action outcomes based on scenario data. Does not interact wi

Complete trajectory from database

A managed subprocess with metadata

Complete quality score for a response. Combines format, reasoning, and execution quality into unified scores for reward calculation.

Configuration for Tinker client

Action taken by agent

HTTP client for the Babylon A2A methods wrapped by this example.

Configuration for a single continuous RL agent.

Unified Tinker client for training and inference. This replaces local vLLM + PyTorch training with Tinker's cloud API: - No local GPU re

Adaptive curriculum for scenario selection. Tracks: - Per-scenario attempt counts - Per-scenario score history - Solved/unsolved

Client for TypeScript simulation bridge. Provides async methods for interacting with the simulation: - initialize(): Start a new simulat

Configuration for one team.

Single step in a trajectory

Configuration for Babylon RLAIF environment

A single agent with its own model that learns continuously from a shared Babylon game. Wraps: model + optimizer + Kondo gate + TurboQuant cac

Utility for collecting episodes with consistent structure. Provides helper methods for episode management during rollouts.

Raised when the Hermes benchmark bridge returns an error.

A ground truth event that agents should react to. Attributes: tick: When the event occurred event_type: Type of causal event

State of a perpetual market

Dataset of prompt samples grouped by purpose

Agent's starting portfolio

Social context for agent

Single message in a conversation.

Perpetual futures market data

Save rollouts for debugging and dataset generation. Saves: - Successful rollouts (score > threshold) for SFT - Paired rollouts (high

Comprehensive evaluation for tracking training progress. Components: 1. Held-out test scenarios 2. Baseline comparison 3. Archet

Current market state

Complete scenario for agent decision-making

Buffer for collecting turns during an episode. Tracks episode-level metrics and provides utilities for episode analysis.

Market condition during a trajectory's time window. Attributes: overall: Aggregate market direction across all tracked assets

Prediction market data

Track running reward statistics for advantage computation.

Singleton for reward weight configuration loaded from YAML.

Result of a rollout generation run

Configuration for the shared-model continuous RL trainer.

GRPO Trainer for Babylon using Atropos This trainer: 1. Registers with Atropos API server 2. Pulls batches of scored trajectories

Complete evaluation result

Rich game state for realistic simulation. Provides detailed market data, orderbooks, news, and social context that mirrors production en

Configuration for KL controller

Base KL controller that works without GPU/torch. Uses log-probability approximations when actual model isn't available.

Configuration for the multi-agent orchestrator.

Manages held-out test scenarios

Result of tokenization with masks

Input validation error

Manages baseline results for comparison

Configuration for GAE computation

Learning rate scheduler types

Manages N agents playing in a shared Babylon game.

Reads Babylon trajectories from a PostgreSQL database.

Single prompt sample for training. Each LLM call within a tick becomes a separate sample. Rewards are attributed based on: - Purpose

Track running reward statistics for advantage computation. Uses exact stats during warmup (first 20 samples), then EMA.

Configuration for the fast simulator

A social interaction scenario for DM/group channels.

A complete trajectory fixture for testing.

Result of schema validation

A2A protocol error

Metrics for a specific archetype

Configuration for archetype-specific training