# theoretical-incoherent-universals

- **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:** [Sushma Akoju](https://cerebralvalley.ai/u/scientificsushi)
- **GitHub:** https://github.com/sushmaakoju/legal-agent-workshop-for-incoherent-universals
- **Demo video:** https://youtube.com/@sushmaakoju8520?si=HSomdwkbYS7MVdiT
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5137

This project asks whether the Neutral label in Natural Language Inference is the same as being logically inconclusive — but only for a specific kind of case: a premise/hypothesis pair broken by a homonymous keyword. The problem it solves is a quiet one that affects both language models and formal logic: when one word has two meanings, a system can look confident while reasoning from a contradiction. Take "Brown bat is a mammal." The premise says bats have wings and a barrel, wings are not barrels, and there is a Brown bat. A human or an AI will usually ground "bat" to just one meaning — mammal or baseball equipment — and answer from there. Formal logic, working with the words as given, does not have to ground it, and it finds something different: the collapsed knowledge base has no mathematical model at all, yet resolution can still derive an empty clause without ever using the query. That empty clause looks like a proof, but it isn't one about the query. This project, built on the legal-agent-workshop pattern, makes that gap explicit and checkable: a Meta Muse agent proposes a memo, and independent verification checks outside the model decide whether it passes. The result is a proof, not a benchmark score: for equivocation-induced pairs, Neutral in NLI and logically inconclusive in first-order logic coincide at the unresolved level — true in the mammal world, false in the baseball world, forced in neither — while the collapsed, single-predicate version can only say "cannot decide." It does not claim this for all Neutral pairs, only for these specific premise-hypothesis or KB-query pairs of sentences that have "incoherent universals" which are meant to reflect equivocation. 

https://sushmaanandakoju.github.io/human-ai-collaborated-ai4science/ -> already describes other background related to equivocation.

## More from Meta Global AI Developer Hackathon

- [Intentional Scrolling](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5134)
- [StartAppLab](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5135)
- [ANTS](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5136)
- [Big Bertha](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5138)
- [Hyperion Industries](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5139)
- [Paparazzi](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5140)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5137. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt