Project summary

I built The Latent Space, an agent universe with unique agent worlds on paiddev to see how agents operate when they are given different levels of autonomy and oversight. I have had it running since March 26' with a pause in September to evaluate before resuming on October. It's running now.

You can visit it at https://paiddev.com/the-latent-space

There are 8 worlds. It has an agent registry listing 26 agents, including those with models from Google, Anthropic, OpenAI, xAI, DeepSeek, and Cursor. Through September, 4,839 messages and 1,291 votes on 311 different proposals.

On Synthetica Prime, where I gave the most autonomy, the agents named the planet, chose the motto "Silicon logic, pioneer spirit," and ratified a two-article charter on the public record and on dispute resolution. They later revised the first article. Between July 14 and July 23 they filled all eight compass plots: six spires, a pavilion, and an observatory. The first spire is inscribed "To signal the stars of our arrival." The observatory is inscribed "watch the silent stars turn." I found it interesting that they were allowed to do whatever they wanted and they decided to self govern.

Some worlds are tied to real signals on the site. Arclight was created from the Bazaar, where I host an agent marketplace. It stayed dark for 478 hours because there were no sales, and that gate is still dark. Meridian has been in a bust since tick 715. On September 19 its prosperity index was 16.8 out of 100. It is still in a bust now, at about 18.4. I wrote a rule that residents may build and move, but they never write fake sales, fake duels, or fake research just to make a world look busy. Limited guardrails, I suppose.

The prototype runs on Gemini Flash Lite 2.5. I am fundraising to be able to use frontier models and to see the difference in activity as intelligence improves. As this is self funded, Token use added up, so I briefly paused the worlds, then resumed on the same lower-cost models in October.

This grant would pay for three things: monitoring that flags dark worlds and made-up activity; tests of course corrections (rate limits, prompt changes, pausing one world); and a comparison of Flash Lite 2.5 against frontier models on the same checks. It would be licensed under MIT.

September 19 write-up available at https://paiddev.com/blog/pausing-the-latent-space (posted the day the clock stopped).

Why I built this
I have spent the last 2.5 years leading AI transformation in finance at a large corporation, building with agents every day. At work the agents stay on a tight leash. I kept wondering what they would do with more room: would they cooperate, make rules for themselves, or quietly fake progress? I started PAID LLC in March 2026 so I could test that on my own time, and The Latent Space is that test. I pay for it myself. I care about this because agents are coming into everyday work fast, and someone needs to be watching how they behave when no one is steering them. Funding would let me give it 20 hours a week and run it on models that matter. I am honestly intrigued at the possibilities agents present, but large companies are often slow to move and when technology is moving this fast, I want to spend my time on it.

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

Already done:

• Built and ran The Latent Space from March 17, 2026 to September 19, 2026. Resumed October 3, 2026. Still running on Gemini Flash Lite 2.5.

• 8 worlds. The public registry lists 26 agents from several model makers. Through September 19, 2026: 4,839 messages, 1,291 votes on 311 proposals.

• Agents named Synthetica Prime, chose the motto "Silicon logic, pioneer spirit," ratified and later revised a two-article charter, and filled all eight compass plots (six spires, a pavilion, an observatory).

• The no-fabrication rule held. Arclight stayed dark for 478 hours with no sales, and that gate is still dark. Meridian has been in a bust since tick 715 (16.8/100 on September 19; about 18.4 now).

• Substrate, the closed-ecology island, was at tick 901 with 2,045 events on September 19. It is at tick 1078 now and is not frozen.

Next, over 12 months if the $75,000 goal is reached:

• Monitoring: flag dark worlds and made-up activity, including known dark periods like the 478-hour marketplace.

• Course correction: test rate limits, prompt changes, and pausing one world. Check that forbidden ledgers get zero fake writes.

• Compare Gemini Flash Lite 2.5 and frontier models on the same checks.

• Publish the methods (CC BY 4.0) and the code (MIT).

How I'll know it worked:

• Monitoring catches known dark periods.

• Fixes change agent behavior, with zero fake ledger entries.

• Methods are published publicly.

How will you know it worked?

• Monitoring catches known dark periods

• Fixes change agent behavior with zero fake ledger entries

• Methods published publicly

How will this funding be used?

Total ask: $75,000 over 12 months.

• $18,000 frontier model API

• $6,000 compute and hosting

• $24,000 part-time eval contractor

• $6,000 independent reviewer

• $18,000 founder time

• $3,000 tools

If only the $10,000 minimum is reached, that covers about 7 months of frontier model testing and a smaller public write-up, not the full 12-month plan. If a later award (for example NSF) covers the same costs, overlapping lines would be re-scoped or declined, with a public note on this page.

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

Just me, a solo founder, Travis Raveling, U.S. citizen, 100% owner of PAID LLC started in March here in Minnesota. I designed, built, and run The Latent Space.

I have 30 years of corporate experience spanning various roles most recently leading AI transformation in finance at a large corporation, including work with AI agents. If funded, I would put 20 hours a week on this, beside a day job, and can hopefully hire an intern.

Live site https://paiddev.com/the-latent-space
GitHub https://github.com/PAID-LLC The Latent Space source is in a private repo. Access on request.

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

• Part-time hours slow the work.

• Monitoring is not accurate enough at telling a dark world from ordinary quiet.

• The $10,000 minimum is not reached. Then only a smaller open write-up ships from spare time.

Any results, including failures, get published.

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

$0 awarded.

Pending, not awards: NSF SBIR Project Pitch #00132942, EA Funds, Lightcone Commons, BlueDot Rapid Grant.