# Meaning Check

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Paul AGYEMANG](https://cerebralvalley.ai/u/pagyemang)
- **GitHub:** https://github.com/pagyemang-dotcom/meaning-check
- **Demo video:** https://youtu.be/5yTS6-LR2_c
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/64

Meaning Check is a pre-flight check for what you mean. The same words are read differently by people with different backgrounds and by AI agents with different settings: "Interesting approach in your draft, maybe worth another look at section 3" means "fix this before Thursday" to the sender, but most readers hear an optional suggestion. Every day we hit Send blindly.

Meaning Check has 24 AI readers on two Meta models read your message first: 16 Muse Glimmer readers (open weights, on our own GPU service) and 8 Muse Spark readers (Meta Model API), set in the UK, the US, Ghana, and India, with different jobs and quick or careful reading. It shows who misreads it and what they take it to mean, and you can click any reader to see its exact setup, what it read, and how likely it found every meaning.

Then it fixes the message with the fewest words. An integer programme finds the fewest facts that rule out every reading some reader prefers to yours and proves nothing shorter works (the audience theorem from my mathematical theory of being understood, AIMS Ghana, under review at JAIR). The writer drafts candidate wordings; both models re-read every one, and the shortest wording both understand is kept, shown as a word-level diff. Eight further readers, never used to build the fix, test it with a stated bound for new readers.

It works for workplace messages and for prompts to AI agents, with one-click scenarios, photo and voice input, a downloadable report, and a CI command (meaning check lint) for policies and prompts.

Pre-registered results: on readers never used to build the fixes, understanding rose from 53% to 100% across 14 messages (gain of 47 points, 95% interval of 21 to 71), with no message made worse. Muse Spark, which built none of the fixes, understood them 54 points more often than the originals. What didn't work is published too: a careful plain rewrite did as well on these tests, and the first Muse Spark run was void because of a scoring bug, fixed and rerun. Every number reproduces offline with one command (meaning check reproduce).

Live app: https://pagyemang-dotcom--meaning-check-web-web.us-east.modal.run/

## More from Meta Global AI Developer Hackathon

- [gradepack](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/61)
- [Nexa](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/62)
- [DarFlow Agent](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/63)
- [Sam Williams](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/65)
- [MoyaBiz AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/66)
- [JAS_AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/67)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/64. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt