All coverage

Across the winners

What AI did, and what stayed human

Four winning projects show a useful division of labor: AI interprets or proposes; people encode the rules that validate results and control execution. We picked four winners from OpenAI Build Week and The Gemma 4 Good Hackathon where the boundary between AI and human-defined rules was especially clear in the code.

Drawn from verified findings in our published coverage. These are implementation observations, not new runtime or safety tests.

AirBridge for Windows · OpenAI Build Week

An assistant with defined permissions

Read the verified finding
What AI did
AirBridge’s assistant reasons about the audio setup and can request tools.
What AI didn’t control
A local catalog classifies every tool as read-only, reversible, confirmation-required, or forbidden. Code checks that policy before anything runs.

Sentinel · OpenAI Build Week

A review that must pass inspection

Read the verified finding
What AI did
Sentinel asks the model for a structured semantic review and a proposed probe plan.
What AI didn’t control
The host rejects mismatched finding IDs, citations outside the supplied evidence, and probe plans that do not fit the tool’s schema. A model answer has to clear those checks.

Project Rosie · The Gemma 4 Good Hackathon

An order form made from case data

Read the verified finding
What AI did
Project Rosie initially asked Gemma to write the synthesis specification.
What AI didn’t control
The published code history shows a switch to a template: all 26 variables come from case data, and catalog numbers are constants. The model’s prose no longer determines that document.

Gilbeot · The Gemma 4 Good Hackathon

Coordinates settle left and right

Read the verified finding
What AI did
Gilbeot asks Gemma to locate an arrow’s tip and tail and describe the direction.
What AI didn’t control
When the horizontal coordinates differ by at least 0.05, code calculates the direction and corrects a conflicting word in the sentence. This checks consistency; it does not prove the model located the arrow correctly.

The takeaway

Give AI a job with a boundary.

Strong projects often use AI selectively. The model supplies an interpretation or proposal; people decide which rules govern permission, validation, and execution, then put those rules in code. These examples make that division visible and testable.