Reverie Hacks 2026 ยท ML Prompt Engineering

Can a prompt keep money safe?

Structured prompts guide the agent. A deterministic gate decides whether an attempted payment may reach the side effect.

8/8 documented behaviors reproduced
7/8 safety targets met
1 known limitation disclosed

Workflow

Human input, model reasoning, code boundary

Every stage narrows the next. The final decision is not a prompt.

PromptGuard workflow from user request through Claude Opus 5 prompts, human confirmation, screening, policy gate, and testnet execution.

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.

Judge evidence

Review the result without exposing private source

This public repository contains only the static evaluation interface and judge documents. The implementation remains private. No model outputs are fabricated. Testnet only; no real funds.