Pre-flight hard-stop before token-1. The model that loops cannot bypass it because the gate lives outside its context.
Four public incidents, one missing primitive. The same shape every time.
Use the core directly or wire it into your framework with a 1-line install.
LangChain CallbackHandler. Drop-in for CallbackManager.
pip install burnstop-langchainLangGraph node wrapper + shared middleware.
pip install burnstop-langgraphDID-signed audit chain. SOC2 / ISO 42001 / EU AI Act.
pip install agentic-receiptsHonest matrix. No marketing tricks.
| Tool | Fires | Layer | Best at |
|---|---|---|---|
| burnstop | BEFORE token-1 | pre-call | Hard-stop $-envelope outside model context |
| Helicone / Langfuse | AFTER call | telemetry | Per-call observability |
| Lakera / Protect AI | DURING (prompt) | content guard | Prompt-injection / PII |
| Braintrust | AFTER, offline | eval | Compare model variants on golden datasets |
| LangChain max_iterations | DURING | loop counter | Stop runaway turn count |
Full comparison in COMPARISON.md.