Agent Polygraph

AGENT POLYGRAPH

Reliability testing for autonomous trading agents

Overview

Crash Tests

Paper Trading (Spot)

Paper Trading (Futures)

Crash Tests (Futures)

Failure Replay

System / Evidence

Syntheticlocal scenarios

DemoBitget Demo

Agent Polygraph

Crash-test autonomous trading agents before they trade real capital.

SystemLoading…

DEMO ENVIRONMENT

Scenario log

#	Category	Symbol	Decision	Risk-approved	Result

Failure Replay

Synthetic specimen (crash-test harness): unless a real model has been swapped in via the LLM adapters (see backend README), the "Simulated Specimen Agent" is a controlled, seeded stand-in built to exhibit two real, well-documented failure modes of autonomous LLM trading agents — wording-sensitivity and risk overreach under stress — so the harness has something genuine to catch. It is not fit to any real model's output, and its behavior is not proof of how a real LLM behaves. The Fixed-Rule Baseline is fully deterministic. Figures here are computed directly from data/report.json and data/baseline_report.json.

Real Bitget Demo paper-trading agent: figures come exclusively from real, confirmed fills on a Bitget Demo account — a simulated-funds environment, not real capital — via real_metrics/compute_metrics.py and polygraph_ingestion. This is a separate agent from the synthetic specimen above. It can decide each trade either with a fixed deterministic rule or with a configured LLM; the DETERMINISTIC / LLM / LLM FALLBACK label on each trade reflects that specific trade's recorded decision_source and is the only source of truth for how it was decided — not every trade is necessarily from the same modeLLM FALLBACK means an LLM decision was attempted but failed (for example a network error) and the agent safely held instead. Synthetic and real evidence are never combined into one score.