ScriptTruth

Inspiration

Large language models are excellent at generating text, but they are not designed to verify facts. While researching real film production workflows, we found that historical mistakes, wrong brand names, incorrect technology timelines, and factual inaccuracies often survive until late stages of production, where fixing them becomes expensive.

ScriptTruth was built around a simple idea:

Generation is not verification.

Instead of asking AI to guess whether something is true, we built a system that extracts factual claims from a script, checks them against live web sources, and explains its conclusions with confidence scores and citations.

What it does

ScriptTruth verifies factual claims inside screenplays and other written content.

It can identify:

  • Historical inaccuracies

  • Technology anachronisms

  • Brand and product timeline mistakes

  • Location and date inconsistencies

  • Fictional entities without falsely marking them as incorrect

Every result includes supporting evidence, citations, and a transparent confidence breakdown.

How we built it

The project combines a lightweight verification pipeline with live web retrieval.

Tech stack

  • Google Gemini

  • Parallel Search

  • FastAPI

  • HTML/CSS/JavaScript

  • Render

Verification pipeline

  • Extract factual claims from the script.

  • Retrieve live evidence using Parallel Search.

  • Compare claims against the evidence using Gemini.

  • Produce deterministic confidence scores.

  • Return evidence-backed verdicts with citations.

We intentionally separated claim extraction from confidence scoring so that the final score is based on transparent logic rather than asking the model to invent a confidence percentage.

Challenges

The biggest challenge was making the system reliable instead of simply impressive.

Some of the problems we had to solve included:

  • avoiding false positives for fictional places like Gotham,

  • distinguishing "insufficient evidence" from "incorrect,"

  • reducing API usage through batched verification,

  • handling free-tier quota limits gracefully,

  • designing a deterministic scoring system instead of relying on model-generated confidence values.

These decisions made the system more trustworthy during real verification workflows.

What we learned

Building ScriptTruth taught us that a good AI product is not just about calling a model API. The harder part is deciding what should be verified, what should remain uncertain, and how to present evidence in a way that users can trust.

The biggest lesson was that thoughtful system design, transparent reasoning, and evidence-backed decisions create far more value than simply generating another AI response.

Future

The same verification engine can extend beyond screenplays into journalism, podcasts, advertising, and other content where factual accuracy matters, while keeping the same evidence-first philosophy at its core.

Built With

Share this project:

Updates

Submission history