Hypothesis Forge

Inspiration

Bold and unusual claims are often handled in one of two unhelpful ways: they are dismissed before being examined, or accepted because an interesting pattern feels convincing. I wanted a third option—a workspace that preserves curiosity while forcing every claim to face evidence, alternatives, controls, and possible failure.

What it does

Hypothesis Forge is a bilingual English–Arabic research workspace and reusable Codex skill that turns speculative claims into testable research.

It helps users:

• freeze the exact claim and its data boundaries;
• perform a blind analysis before revealing the proposed pattern;
• separate established facts, reasonable inferences, hypotheses, and symbolic interpretations;
• compare the claim against matched controls and rival explanations;
• define a risky prediction and an explicit falsification condition;
• export a portable Markdown research report;
• hand the structured claim to the included $forge-hypotheses Codex skill.

The tool does not pretend to prove whether a claim is true and never fabricates sources.

How I built it

I used a split architecture. The browser application handles deterministic intake, transparency, scoring, and report export. The reasoning-intensive analysis is handled separately through the reusable Codex skill.

The interface was built with React, TypeScript, Vinext, and Vite, with deployment through Cloudflare infrastructure. It supports Arabic RTL, responsive layouts, keyboard accessibility, reduced-motion preferences, and client-side privacy.

How Codex and GPT-5.6 accelerated the work

Codex was the primary implementation workspace for the project. Working with GPT-5.6 accelerated the architecture, bilingual interface, reusable skill design, documentation, validation, and testing.

Key decisions made during the Codex workflow included:

1. separating deterministic browser behavior from model-generated inference;
2. preventing the interface from inventing evidence or sources;
3. making matched controls and falsification first-class parts of the workflow;
4. keeping user claims client-side until the user deliberately copies them into Codex.

GPT-5.6 was used through Codex to pressure-test the methodology, challenge weak assumptions, refine the bilingual language, implement the working product, and define a structured research-report contract.

Challenges

The main challenge was preserving creative exploration without presenting symbolic resonance as scientific proof. Another challenge was making a complex research protocol understandable in both English and Arabic while maintaining full RTL support.

Accomplishments

I created a working public application, a reusable Codex skill, a stable report format, a bilingual interface, and a methodology that clearly shows what evidence would weaken or defeat a claim.

What I learned

A useful research tool should not only help a claim look stronger. It should define how that claim could lose.

What's next

Next steps include benchmark datasets, stronger matched-null controls, optional citation workflows, collaborative reports, and systematic evaluation of the methodology across scientific, historical, numerical, and textual claims.

Built With

• cloudflare
• codex
• github
• gpt-5.6
• javascript
• node.js
• openai
• react
• typescript
• vinext
• vite
• workers