balancelab

Detect runaway exchange loops in any token economy — game balance, AI spend, or reward-function auditing.

Quick Start · How It Works · CLI Reference · MCP / Claude · vs. Alternatives · Contributing

Why

Any system with exchange rules — tokens, compute budgets, reward points, or in-game currency — can develop runaway loops that the designers never intended. The loops are mathematically inevitable once the exchange graph contains a negative cycle; the only question is whether you find them first.

Game economies: A crafting loop that converts gold → silver → gems → gold at a net gain of 24x will be found by players within hours of launch — not by QA. Manual balance spreadsheets don't scale. Playtesting can't enumerate all cycles.

AI token budgets: An analytics pipeline discovers that re-expanding summaries and re-scoring them earns more "quality credit" than it costs in tokens — so it loops. By month-end the pipeline consumed 26x its monthly budget. 78% of teams have no per-workflow token alert; they see the overrun on the billing page, not in a dashboard.

Reward-function auditing: An RL agent finds a sequence of actions where each step earns more reward than it costs — a classic reward-hacking loop. balancelab encodes the reward structure as an exchange graph and detects the profitable cycle before the agent finds it.

balancelab treats any set of exchange rules as a directed graph, applies Bellman-Ford on log-weighted edges, and produces an exploit report with exact gain ratios — before launch or before your next billing cycle.

balancelab scan --format json   # CI-friendly, fails on exploits

Related MCP server: agent-cost-mcp

Use Cases

AI token spend monitoring

Model a multi-workflow AI pipeline as an economy where each model call costs tokens and each quality-score credit refills the budget. balancelab finds the loop before it inflates your bill 26×.

from balancelab import EconomyGraph, EconomyRule, ExploitFinder, recommend_fixes

graph = EconomyGraph()
# Analytics pipeline: each quality_score credit = 3500 new budget tokens
graph.add_rule(EconomyRule("budget_tokens", "quality_score", 200.0, 1.0, rule_id="score"))
graph.add_rule(EconomyRule("quality_score", "budget_tokens", 1.0, 3500.0, rule_id="redeem"))

report = ExploitFinder().find_exploits(graph)
# → ExploitReport(exploits=[ExploitPath(path=[budget_tokens → quality_score → budget_tokens], gain_ratio=17.5)])

for fix in recommend_fixes(report):
    print(fix.fix_type, fix.description)   # rate_cap: cap the redeem edge

See examples/ai_token_budget_monitor.py for a full walkthrough including budget projection and sensitivity analysis.

Game economy balance

A crafting loop gold → silver → gems → gold at 24x gain will be found by players within hours of launch. balancelab encodes your exchange rules and proves whether arbitrage is possible — before it ships.

See examples/demo.py for the classic game economy example.

How It Works

flowchart LR
    A[Define EconomyRule\nsource_item → target_item\nat source_qty:target_qty] --> B[Build EconomyGraph\ndirected exchange graph]
    B --> C[ExploitFinder\nBellman-Ford on\nlog-weight edges]
    C --> D{Negative cycle\ndetected?}
    D -->|Yes| E[ExploitPath\ngain_ratio > 1.0]
    D -->|No| F[Economy balanced\ntotal_found = 0]
    E --> G[ExploitReport\nall cycles ranked\nby gain_ratio]

Core primitives:

• EconomyRule — an immutable, content-addressed exchange: give source_qty of source_item, receive target_qty of target_item. ID = SHA-256[:16] of the rule parameters. Same rule always produces the same ID.

• EconomyGraph — a directed graph of EconomyRules. Supports neighbor traversal and serialization.

• ExploitFinder — converts exchange rates to log-weights (weight = -log(rate)). A negative cycle in the log-weight graph corresponds to a positive-gain cycle in the economy. Uses Bellman-Ford for O(V·E) detection.

• ExploitPath — a single circular trade path with its gain ratio (e.g., 24.0x).

• ExploitReport — the full scan result: item count, rule count, all exploit paths, timestamp.

Facts are stored in a local SQLite database. No server required.

Features

Quick Start

pip install balancelab

from balancelab.economy import EconomyRule, EconomyGraph, ExploitFinder
from balancelab.report import print_report

# Define your economy's exchange rules
graph = EconomyGraph()
graph.add_rule(EconomyRule("gold", "silver", 1.0, 3.0, rule_id="mint"))
graph.add_rule(EconomyRule("silver", "gems", 1.0, 2.0, rule_id="jeweler"))
graph.add_rule(EconomyRule("gems", "gold", 1.0, 4.0, rule_id="trader"))

# Find exploits
finder = ExploitFinder()
report = finder.find_exploits(graph)

# Display results
print_report(report)
# Exploit Report (id: a3f8b2c1d4e5f6a7)
#   Items: 3  Rules: 3
#   Exploits found: 1
# ┌──────────────────┬─────────────────────────────────────────┬────────────┐
# │ ID               │ Path                                    │ Gain Ratio │
# ├──────────────────┼─────────────────────────────────────────┼────────────┤
# │ 4d7e9c2a1b8f3e6a │ gold → silver → gems → gold             │ 24.00x     │
# └──────────────────┴─────────────────────────────────────────┴────────────┘

CLI Reference

balancelab [--db PATH] COMMAND [OPTIONS]

Global options:

Examples:

# Add exchange rules
balancelab add gold silver 1.0 3.0 --rule-id mint
balancelab add silver gems 1.0 2.0 --rule-id jeweler
balancelab add gems gold 1.0 4.0 --rule-id trader

# Scan for exploits
balancelab scan

# Machine-readable output (for CI)
balancelab scan --format json

# Review previous scans
balancelab log

MCP / Claude

balancelab ships a Model Context Protocol server. Add it to Claude Desktop:

{
  "mcpServers": {
    "balancelab": {
      "command": "balancelab-mcp"
    }
  }
}

Available MCP tools: add_rule, scan_economy, list_reports.

Install with MCP support: pip install "balancelab[mcp]"

You can also find balancelab on Smithery for one-click MCP installation.

REST Server

balancelab ships a FastAPI server exposing all core operations over HTTP.

pip install "balancelab[api]"
uvicorn balancelab.api:app --reload

Available endpoints:

The full OpenAPI spec is in openapi.yaml.

OpenAI

Use tools/openai-tools.json for OpenAI function calling:

import json, openai
tools = json.load(open("tools/openai-tools.json"))
response = openai.chat.completions.create(
    model="gpt-4o",
    tools=tools,
    messages=[{"role": "user", "content": "Scan my economy for exploits"}],
)

Advanced API

Beyond the core exploit detection primitives, balancelab exposes several higher-level APIs imported directly from the package:

from balancelab import (
    simulate,
    recommend_fixes,
    sensitivity_analysis,
    critical_path,
    BalanceFix,
    SensitivityResult,
    SimulationResult,
)

simulate(graph, initial_levels, n_steps=100, agent_strategy="greedy") -> SimulationResult

Run a forward simulation of the economy. Strategies: "greedy" (apply every rule each step), "balanced" (one rule per source item), "exploit" (agent actively exploits known cycles).

from balancelab import simulate
from balancelab.economy import EconomyGraph, EconomyRule

graph = EconomyGraph()
graph.add_rule(EconomyRule("gold", "silver", 1.0, 3.0))
result = simulate(graph, {"gold": 100.0, "silver": 0.0}, n_steps=50, agent_strategy="exploit")
print(result.inflation_detected)   # True/False
print(result.final_levels)         # {"gold": ..., "silver": ...}

SimulationResult fields: steps, final_levels, violated_rules, inflation_detected, inflation_resource, summary.

recommend_fixes(report) -> list[BalanceFix]

For each exploit in an ExploitReport, suggest the minimum intervention to neutralize it.

from balancelab import recommend_fixes
fixes = recommend_fixes(report)
for fix in fixes:
    print(fix.fix_type, fix.target_edge, fix.suggested_value, fix.description)

BalanceFix fields: exploit_path, fix_type ("rate_cap", "cooldown", "daily_limit", "require_prerequisite"), target_edge, suggested_value, description, estimated_reduction_pct.

sensitivity_analysis(graph, report) -> list[SensitivityResult]

Rank all economy nodes by how much they affect overall balance, descending by impact_score.

from balancelab import sensitivity_analysis
results = sensitivity_analysis(graph, report)
for r in results:
    print(r.node_id, r.impact_score, r.recommendation)

SensitivityResult fields: node_id, node_type ("hub", "source_only", "target_only"), impact_score (0–1), connected_rules, exploit_involvement, recommendation ("monitor", "rate-limit", "gate").

critical_path(graph) -> list[str]

Find the sequence of nodes with the highest economic throughput (most exchange-rate flow). Useful for identifying which items act as economic bottlenecks.

from balancelab import critical_path
path = critical_path(graph)
print(" -> ".join(path))   # e.g. "gems -> gold -> silver"

vs. Alternatives

Repository Tree

balancelab/
├── src/balancelab/
│   ├── __init__.py          # Public API
│   ├── economy.py           # EconomyRule, EconomyGraph, ExploitFinder
│   ├── store.py             # SQLite persistence
│   ├── report.py            # Rich/JSON/Markdown formatters
│   ├── cli.py               # Click CLI
│   ├── api.py               # FastAPI server
│   └── mcp_server.py        # MCP server
├── tests/
│   ├── test_economy.py      # Data model tests
│   ├── test_exploit.py      # Bellman-Ford exploit detection
│   ├── test_store.py        # SQLite CRUD
│   ├── test_report.py       # Formatters
│   ├── test_cli_runner.py   # CLI integration
│   └── test_api.py          # FastAPI endpoints
├── tools/openai-tools.json  # OpenAI function calling spec
├── openapi.yaml             # Full OpenAPI 3.1 spec
├── examples/demo.py         # Standalone demo
└── smoke_test.py            # End-to-end smoke test

GitHub Topics

Suggested topics for discoverability: #game-economy #arbitrage #balance-testing #agents #mcp #llmops #python #cli #exploit-detection #graph-algorithms

Case Studies

See how teams are using balancelab in production:

• Eliminating Economy Exploits Before Launch — Stellar Forge finds 5 exploits (including a 1,250x gain ratio) before shipping to 2M DAU

• Catching Reward Hacking in an AI Agent Token Economy — Orbital Systems catches synthetic-task reward hacking in simulation before it reaches production

Stay Updated

Subscribe to The Silence Layer — weekly dispatches on production AI infrastructure, new releases, and the failure modes that production AI systems don't surface until it's too late.

Star History

Closed loop / Non-Ornament

See docs/CLOSED_LOOP.md for when this library is load-bearing vs ornamental, and when not to use it.