One of nine builds treating alignment, binding and governance as one problem. A fix counts only when an auditor proves it.
Libra Prime is the current build of Alignment, an AI system I am building alone and in public. It is one of nine builds in a program that treats alignment, binding and governance as the same problem. Libra Prime is Alignment building itself: 15 software roles do real work on my own server with write access. Every action is published as a signed, hash-chained record anyone can check. A close counts as a real fix only when a separate Auditor role proves the role's change caused it. As of Nov 20, 2026: 736 closes, 32 proven real fixes, 704 reopened. The site is the proof and updates as the build changes: https://operationalsystemsgroupllc.com/
Goal: a public, checkable way to tell what an AI agent with write access actually did and whether it worked. Libra Prime is done when its plain-English test passes 8 of 8 with an Auditor pass and the backlog reaches 0 counting only Auditor-verified closes; today 1 of 2 finish-line checks pass. Over 12 months: finish Libra Prime; add signed records for memory reads and writes; record the first linked failure-to-recovery chains; fix and independently check the Auditor; publish the method and data so others can run the same test on their own agents; carry the result into the next builds.
The finish line is checked by the build's own fixed check and shown on the site. Claims marked "not proven yet" get proven or disproven in the signed record. It also worked if at least one outside party checks the record or runs the method themselves.
Minimum $25,000: about two months of my living costs and the server. Full $300,000 for 12 months: $120,000 living costs ($10,000 per month), $100,000 one-time costs to get personally stable so the work is not interrupted, $60,000 server and compute, $20,000 buffer.
Just me, Nick Evans. I built all of it: the roles, server, signing, audit rules and site. It runs over 200,000 tasks a day, mostly fixed scripts with no model, and all role sessions on local models. No prior AI safety grants or publications. LinkedIn: https://www.linkedin.com/in/nicholas-evans-ops/ GitHub: efiadm
Most likely: it stays a single-server case study nobody else uses. Second: the Auditor is not trustworthy enough; its own close audit does not pass today. Third: working alone, I drift into features instead of the open questions. If it fails, the record of what failed and why is still public.
$0. Applied Nov 20, 2026, all pending: BlueDot Rapid Grant ($20,000), BlueDot Seed Grant ($200,000), Transformative AI Fund ($300,000). All for this same work.
Comments
No comments yet. Sign in to create one!