PunditPit

Inspiration

Prediction markets are fascinating because they turn uncertainty into a visible crowd probability. But most people see an odds number without understanding what it reflects: resolution rules, liquidity, market movement, or incomplete information. I wanted to build a companion that makes prediction markets more understandable, useful, and fun without asking users to risk real money.

What it does

The Pundit Pit is a play-money, AI-powered companion for prediction markets. Mood shows active market questions with Yes/No probabilities, resolution context, and crowd-signal explanations. Voices creates a running, evidence-bound commentary feed from Quant Bro, Contrarian, and Trash Talker. Radar gives users a quant-style view of public order-book depth, pricing math, and play-money fill simulations. Arena lets users replay resolved markets, write a thesis before seeing the outcome, challenge it with GPT-5.6, and review their confidence after the reveal. The app does not execute trades or provide investment advice.

How we built it

I built Pundit Pit with Codex using a vanilla HTML, CSS, and JavaScript client plus a Node.js server. The app uses public Polymarket market and price-history data, saved replay fixtures for reliable demonstrations, and the OpenAI Responses API with GPT-5.6. GPT-5.6 is constrained through narrow server-side evidence tools. It receives only selected public market evidence when generating commentary, Radar explanations, and Conviction Lab reviews. Before an Arena outcome is revealed, the outcome is deliberately excluded from the AI’s evidence payload. I deployed the project as a live web app with secure server-side environment variables and persistent storage for the Voices commentary tape.

Challenges we ran into

The hardest challenge was keeping the experience genuinely useful without making unsupported market claims. Public price movement does not reveal why a market moved or what individual traders intended. We addressed that by separating observed facts from interpretation, showing caveats around liquidity and execution, using play money only, and forcing GPT-5.6 to work from supplied evidence. We also built deterministic fallbacks and saved fixtures so the app remains usable when public data or AI services are unavailable.

Accomplishments that we're proud of

We are proud that Pundit Pit became a complete consumer experience rather than a single AI demo. The strongest feature is Conviction Lab: users commit to a thesis before the result is visible, receive a structured challenge to their assumptions, and then review their calibration after the reveal. This makes AI a reasoning partner instead of a prediction machine. We are also proud of the balance between fun and responsibility: Trash Talker adds personality while remaining bounded away from mocking people, sensitive events, or giving trading advice.

What we learned

We learned that trustworthy AI experiences need clear boundaries. The most useful role for GPT-5.6 was not predicting the future—it was helping users ask better questions about evidence, uncertainty, and confidence. We also learned that replayable data, transparent assumptions, and graceful fallbacks are essential for making an AI product both credible and reliable.

What's next for PunditPit

Next, we want to expand the resolved-market replay archive, add richer calibration and prediction-journal insights, and let users compare their reasoning over time. We also plan to add shareable learning summaries, more market categories, and opt-in strategy experiments—all while keeping the product play-money only, evidence-first, and focused on helping people develop better judgment around crowd probabilities.

Built With

• codex
• gpt5.6
• openai