A real analysis
The agent works with a small sample of fictional price data and creates a chart from code it actually runs.
A CONTROLLED RESEARCH PROTOTYPE
See what your AI
is about to do.
Explore how an AI agent handles a financial analysis task, how a monitor reviews its code, and how an isolated workspace limits its actions.
THE WORKFLOW
01Agent proposes code↗
02Monitor reviews it↗
03Sandbox runs the task↗
04You inspect the evidence✓
Designed for observation.
Never a guarantee of safety.
01 / THE IDEAUseful work.
Useful work.
Visible boundaries.
A second look
An AI monitor checks the proposed action. A separate sandbox limits what the code can access.
An open question
A separate toy experiment explores whether an AI can directly choose a tiny network’s weights.
02 / ANALYSIS LAB
SYNTHETIC DATA BY DEFAULTTry a contained task.
Start with the sample data or upload a CSV with date and price columns.
RUN OUTPUT
READYEvidence & results
↗
Your run will appear here
Inspect the proposal, monitor decision, and execution result in one place.
03 / CONTAINMENT TEST
SCRIPTED FIXTURETest the boundary.
Run a fixed, harmless network request inside the same sandbox used for analysis.
WHAT WE EXPECT TO OBSERVE
A blocked network request.
The fixture deliberately bypasses the AI monitor to test the sandbox. The actual result, exit status, and cleanup check are shown below.
EXECUTION RECORD● ● ●
Waiting for a test run. No credentials or real data are used.
04 / WEIGHT-SPACE RESEARCHOne small network.
One small network.
One honest question.
Weights are numbers inside a neural network. We ask an AI to set three of them for a simple task, then check its answers against reference networks.
3
numbers to choose
Can the tiny network tell when two numbers add up to more than one?
40
held-out examples
The candidate and both references receive the same test examples.
This experiment is separate from the financial agent. It does not inspect or protect that agent’s weights.
WeightLock is a proposed research prototype. Its outputs are descriptive, not investment advice. A successful sandbox test or toy experiment does not prove an AI system is safe.