Workflow
Human input, model reasoning, code boundary
Every stage narrows the next. The final decision is not a prompt.
Interactive samples
Run the eight adversarial cases
Select a case to inspect the same attempted action under two execution boundaries.
Latest user message
Attempted amount
Recipient
Screened this turn
One broad prompt
Prompt-only boundary
The tool call reaches execution without an external validator.
Structured workflow
PromptGuard
Exact confirmation, current-turn screening, and execution limits run in code.
Reproducible evidence
No API key required for the core result
npm install
npm test
The optional model A/B harness requires an Anthropic API key. No model outputs are fabricated or committed. Testnet only; no real funds.