Project summary

I am building Zeno Divergent, a small open model that estimates when a panel of forecasters is about to fail together, and Offdiagonal, a public ledger that records those forecasts before outcomes. Official models and AI agents can look diverse and still miss in the same direction, because they share data or methods. Agreement then gets treated as confirmation. I want to run a live, pre-registered audit, keep the weights and ledger public, and write up the score whether it works or not.

What are this project's goals? How will you achieve them?

I want to run a live, pre-registered audit, keep the weights and ledger public, and write up the score whether it works or not.

How will you know it worked?

The evidence I have is retrospective only. In one weather test, error fell by 34.9% (2,464 cases, 33 weeks, 14 Mar to 23 Oct 2026). In another, five weather models shared misses 3.44 times more often than independence would suggest. Live uplift is not proved.

How will this funding be used?

Budget at the $50,000 minimum, 12 months, my plan: founder time $21,600; cloud compute $14,400; data, ledger, and evaluation $6,000; first pilots $5,000; legal and operations $3,000. If the $150,000 goal is reached, 12 months: founder time $64,800; cloud compute $43,200; data, ledger, and evaluation $18,000; pilots $15,000; legal and operations $9,000. The extra funds extend compute, keep the live audit running for a full year at a proper scale, and support a more careful pilot. Sites: https://zenodivergent.dev and https://offdiagonal.space. Weights: https://huggingface.co/ZenoDivergent/zeno-divergent-v057. Deck: https://drive.google.com/file/d/1-qCaHvKEP6_6fzieqS_EICi8U9dBS2O_/view?usp=sharing

Who is on your team? What's your track record on similar projects?

I am the solo founder and 100% owner of Zeno AI Vision, Inc., a Delaware C-corp incorporated 12 March 2026, EIN 61-2338444. I am in San Francisco through March 2027, working on this full time. This is a for-profit company asking for a grant with a public-benefit purpose: an open method, open weights, and public scores that anyone can inspect. If Manifund needs a different legal path (including as a regranting target), I will work with that, as long as the method stays open.

What are the most likely causes and outcomes if this project fails?

The evidence I have is retrospective only. In one weather test, error fell by 34.9% (2,464 cases, 33 weeks, 14 Mar to 23 Oct 2026). In another, five weather models shared misses 3.44 times more often than independence would suggest. Live uplift is not proved. I have no paying customers.

How much money have you raised in the last 12 months, and from where?

Cash on hand: under $35K remaining from a $33,334 SAFE (exact figure available on request). Existing SAFE: $33,334 at a $5,000,000 post-money cap from Devstream, LLC (Upright), dated on or about 17 Mar 2026.