Rethink

Inspiration

Most AI systems try to answer the question they are given. Rethink starts one step earlier: is that actually the question that matters most?

I built Rethink because complex projects often fail long before execution. People optimize solutions before validating the underlying problem, build on untested assumptions, or gather evidence that confirms what they already believe.

Rethink is designed around a simple principle:

Do not optimize the branches before validating the trunk.

The system identifies the unanswered question whose answer would change the greatest number of downstream decisions, then selects the reasoning method best suited to reduce that uncertainty.

What it does

Rethink is a human-controlled AI reasoning system that maintains a living project state consisting of:

• the current problem definition
• assumptions
• evidence
• reasoning history
• project stage
• unresolved uncertainty
• recommended next action

Rather than immediately generating a solution, Rethink identifies the highest-leverage unanswered question and routes the problem to an appropriate reasoning method such as validation, stress testing, root-cause analysis, measurement, prioritization, simplification, testing, or decision-making.

The system distinguishes between evidence, assumptions, research questions, planned tests, test results, public-source findings, user assertions, and synthetic data.

When public research cannot resolve a question that requires interviews, private records, physical observation, authorization, or real-world testing, Rethink creates a Human / Real-World Gate instead of pretending the uncertainty has been resolved.

The human always retains final authority. Users can inspect evidence and assumptions, choose another reasoning method, override a recommendation, add evidence, control project progression, or make the final judgment themselves.

How I built it

I developed the underlying Rethink reasoning framework and used Codex as my engineering partner to turn it into a working application during OpenAI Build Week 2026.

The application uses GPT-5.6 Sol through the OpenAI API for structured reasoning, research synthesis, evidence evaluation, method routing, and analysis.

Codex helped implement and refine:

• the reasoning engine and method router
• evidence classification and traceability
• GPT-5.6 integration
• public research and citation handling
• background reasoning execution and recovery
• project persistence
• project backup and restoration
• human decision gates
• report generation
• automated testing and regression fixes

I defined the product logic, reasoning architecture, acceptance criteria, test scenarios, and required system behavior, then repeatedly stress-tested the implementation to determine whether it was behaving according to the Rethink framework.

Challenges I faced

One of the hardest challenges was preventing an AI reasoning system from becoming overly helpful in the wrong way.

A system designed to help move projects forward can easily start treating weak evidence as strong evidence, interpreting the absence of contradictory information as confirmation, or continuing to generate answers when the correct response is that more evidence is required.

I deliberately tested these failure modes.

For the demonstration project, I created synthetic data engineered to strongly support a business proposition. The numbers exceeded the project's predefined thresholds, but the data was simulated.

Rethink recognized that the evidence was synthetic and refused to allow it to validate a real-world proposition or satisfy the Human / Real-World Gate.

The result remained Insufficient Evidence.

That behavior represents an important part of the project: sometimes the best AI response is not another answer. It is identifying exactly what remains unknown and what evidence would actually change the decision.

What I learned

Building Rethink reinforced the difference between generating information and managing uncertainty.

AI can be extremely capable at analysis, research, and reasoning, but useful decision support also requires clear evidence boundaries, explicit assumptions, traceability, and human authority.

I also learned how effectively Codex can act as an engineering partner when the human provides a clear architecture, acceptance criteria, and continuous testing rather than simply asking it to generate an application from a single prompt.

What's next

Future development will focus on improving guided cycle management, expanding the reasoning-method library, strengthening collaboration capabilities, and creating more polished professional report exports.

The larger goal is for Rethink to become a reusable reasoning layer for complex projects: a system that helps people determine what should be investigated next before deciding what should be built next.

Built With

• api
• codex
• css
• gpt-5.6
• html
• javascript
• node.js
• openai
• responses
• web