Inspiration
“Men had a higher admission rate overall” and “Men had a higher admission rate in every department” sound close. In the 1973 UC Berkeley admissions sample, the first claim holds and the second fails. A few words change which rows you need to examine. Second Look helps student reporters, editors, and anyone writing from a spreadsheet catch that gap before publishing.
What it does
Paste a headline in plain language and get a verdict with the figures behind it. Start with UC Berkeley's six-department admissions sample or bring your own CSV. In the Berkeley example, the pooled admission rates are 44.5% for men and 30.4% for women. Change “overall” to “in every department” and only two of six departments match. Department A is one counterexample, at 62.1% for men and 82.4% for women.
On a phone, the verdict and main numbers sit just below the headline. Readers can open the calculation details or inspect the source table to see what was counted.
How I built it
GPT-6 Luna interprets the headline through the OpenAI Responses API and returns a structured plan identifying the measure, scope, and comparison. The app validates the plan, calculates the figures from the data, and uses those results to determine the verdict. The React and TypeScript interface presents the result along with its scope and supporting counts. Two Vercel Functions handle the interpretation requests and call OpenAI from the server.
For uploaded CSVs, the browser parses and evaluates the rows locally. The model receives the headline, column names, and inferred column types. It does not receive the file's rows.
Challenges and limitations
The same words can describe very different checks. For example, “more admitted” can refer to a headcount, while “a higher admission rate” needs a numerator and a denominator. A descriptive table also cannot establish what caused a disparity. Second Look reports insufficient evidence when the available data cannot answer a question. The CSV prototype focuses on counts, numeric summaries, and filtered comparisons.
What's next
Compare several drafts of a headline side by side, and export the final wording with its calculation and source for editorial review.
The demo video uses AI-generated English narration.
Built With
- gpt-6
- luna
- openai
- react
- typescript
- vercel
