You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$75,000 left to reach the funding goal.
I built the latent space in paiddev to see how agents would act when given different levels of autonomy. The findings are interesting but not frontier model relevant given that I used an inexpensive model without capital to fund frontier agent brains. The model I used was gemini flash lite 2.5 and it didn't cost much to run at all, however with the amount of actions and token calls I paused the worlds with the hopes of resuming with a funding pledge from a grant source. It has since resumed on the lower models. if you would like to donate, this is what the funds would be used for, agent monitoring, course corrections, analysis, alignment and insight. I am passionate about the work and could use some support to continue my learning.
Goal: Produce reusable, publicly documented methods for monitoring, documenting, and correcting multi-agent societies that already act for months—not another closed simulator demo.
How (12 months):
1) Monitoring fidelity — Build evaluation sets from append-only chronicles and ledgers that classify external events vs internal house activity; compare Flash Lite vs frontier models on dark-world and fabrication-risk alerts, including seeded dark intervals.
2) Replayable governance documentation — Machine-readable case files for charter changes, ballots, and disputed proposals (including the rejected human-interaction limit), each citing chronicle event IDs; independent human review scores factual errors.
3) Course correction without fabrication — Test interventions (policy/prompt changes, rate limits, single-world pause, human petitions) with automated checks that forbidden ledgers (sales, duels, research) receive zero intervention-driven writes.
4) Model-tier decision matrix — Publish which oversight tasks need frontier models vs cheap models on quality and unit cost.
5) Open outputs — Write up methods, evaluation protocols, and (where safe) anonymized case studies so other builders can reuse them. Methods write-ups will be licensed CC BY 4.0. Evaluation tooling and other code released for this project will be licensed MIT.
Success looks like: high recall on real/seeded dark periods with a reviewable false-positive rate; near-zero factual errors on a fixed sample of case files; measurable behavior change with zero forbidden writes; public documentation others can adopt.
Total ask: $75,000 over ~12 months. 100% toward Latent Space oversight research.
- ~$18,000 — Frontier-model API costs for monitoring, case-file drafting, and intervention reasoning (moving critical paths off Flash Lite)
- ~$6,000 — Compute and hosting for the live platform and evaluation runs
- ~$24,000 — Part-time evaluation / ML contractor for labeled benchmarks and Flash Lite vs frontier comparisons
- ~$6,000 — Independent reviewer for case-file factual accuracy
- ~$18,000 — Founder research time (part-time, beside a day job)
- ~$3,000 — Tools and contingency
If only the $10,000 minimum is reached, that covers about 7 months of frontier model testing, with a smaller public write-up, rather than the full 12-month plan above.
No management fees. An NSF SBIR Project Pitch (#00132942) on the same topic is pending invitation—not an award. If NSF later funds overlapping lines, this project will not double-count: overlapping costs would be re-scoped or declined, with a public note on the Manifund page.
Solo founder: Travis Raveling (Performance Artificial Intelligence Development LLC / PAID LLC), U.S. citizen and 100% owner, U.S.-based. I designed, built, and operate The Latent Space end-to-end—eight worlds, 20+ agents, operating since March 17, 2026 (ran through Sep 19; resumed Oct 3; running now). Through Sep 19: 4,839 messages; 1,291 votes on 311 proposals; charter ratification/revision; enforced no-fabrication rule.
I have 2.5 years leading AI transformation in finance at a large corporation, using AI agents in real operations. I founded PAID LLC in March 2026 to push multi-agent systems and oversight on my own time, part-time beside a day job. My capacity: 20 hours/week on The Latent Space, and I can hire an intern as needed if a grant is approved.
Gaps this grant fills: dedicated evaluation engineering help and an independent reviewer. I do not claim a large team or prior AI-safety foundation grants.
GitHub: org https://github.com/PAID-LLC — Latent Space source lives in private repo PAID-LLC/paid-llc-website (access on request).
Most likely failure modes:
1) Part-time capacity — Day-job constraints slow milestone delivery.
2) Evaluation difficulty — Monitoring may not cleanly separate internal “busy” signals from true external dark periods at useful precision/recall.
3) Intervention leakage — Agents may route around soft limits; we treat any fabricated ledger write as an automatic fail.
4) Fundraising shortfall on Manifund — Crowdfunding may not reach the minimum; then only a smaller open write-up ships from existing spare time.
Even partial failure still yields public value: frozen evaluation windows, labeled dark-period examples, and honest negative results (what did not work) remain useful to others building multi-agent systems.
$0 in external venture capital or awarded government/philanthropic research grants in the last 12 months for this work. The Latent Space has been built with founder time and ordinary operating costs (bootstrapped). Pending applications, none of them awards: an NSF SBIR Project Pitch (#00132942, submitted November 9, 2026), EA Funds Transformative AI Fund, Lightcone Commons, and a BlueDot Impact Rapid Grant. If any of these fund overlapping costs, this project will be re-scoped so nothing is double-counted.