BoundaryLens

• 

  Collaborating with GPT-5.6 Thinking + Codex on BoundaryLens to beat the Hackathon deadline.

Inspiration

BoundaryLens grew from a simple problem: AI-generated conclusions can sound confident even when the supporting evidence is incomplete, ambiguous, or missing.

My background in medical-malpractice underwriting and risk analysis taught me that conclusions should never be stronger than the evidence behind them. I built BoundaryLens to bring that same discipline to AI-assisted reasoning.

What it does

BoundaryLens reviews a test case and produces an evidence-integrity audit using the OpenAI API and GPT-5.6.

The audit provides:- 1) An overall integrity score 2) Claim-level findings 3) Supported and unsupported conclusions 4) Identified evidence gaps 5) Human-review requirements 6) A calibrated rewrite 7) A transparent score breakdown

The goal is not merely to label an answer “right” or “wrong.” BoundaryLens shows where the reasoning is supported, where certainty exceeds the available evidence, and where human judgment is still required.

How I built it

I developed BoundaryLens as a working web application during OpenAI Build Week. Codex assisted with implementation and repository work, while GPT-5.6 performs the live evidence-integrity analysis through the OpenAI API.

The application sends the selected test case to the model using a structured audit framework and displays the returned findings in a readable report.

Challenges

The largest challenges involved configuring the API key, validating the live model connection, resolving local startup issues, and making the audit output both detailed and understandable.

I also had to balance automation with an important governance principle: the system should assist human review, not replace it.

Accomplishments

I completed a functioning BoundaryLens prototype, connected it to the live OpenAI API, created a GitHub repository, and recorded a demonstration.

The final live test produced a 36/100 integrity score with three claim-level findings, evidence gaps, supported-versus-unsupported conclusions, a human-review requirement, a calibrated rewrite, and a transparent scoring explanation.

What I learned

I learned how to move from an evidence-governance concept to a working AI application. I also gained hands-on experience with Codex, GitHub, JavaScript, local development, API configuration, and structured model outputs.

Most importantly, I confirmed that AI evaluation can be made more useful when uncertainty and evidentiary limits are shown clearly rather than hidden behind a confident answer.

What's next

Next, I would expand BoundaryLens with additional test cases, exportable audit reports, configurable scoring criteria, side-by-side model comparisons, and a stronger human-review workflow for high-stakes use cases.

Built With

• codex
• github
• gpt-5.6
• javascript
• openai-api