Egeneered Atifitial Systems with layerd intelegence, 2 live systems and Currently engeneering BWFP system.
Pillar 1 — The contained agent
We are building an agent with its safety built inside it from the start, not added after. The structure is in place before the agent acts, receives, or processes anything. It checks itself before it acts. No single check authorizes it alone. A check that disagrees stops it. Every attempt is logged, so nothing happens in secret. That is the difference between a fence and a brake. A fence stands outside and hopes nothing climbs over. Our brake lives inside the engine and decides before anything moves. That is why it cannot be reasoned past or removed from outside.
Literacy and practice
We are building an agent whose work is open to be checked and understood, so people can judge what it does and when to trust it. That is AI literacy. And the agent is built to work in real situations, doing real work step by step. That is AI in practice. Safety people cannot see is safety people cannot trust. People do not have to take our word for it. They can read the test record, see the misses, and judge for themselves. That is how trust is earned: not by asking for it, but by showing the work.
The containment audit org
We are building a containment audit organization. Independent audits with written feedback, before deployment. An agent does not ship until someone outside the build has checked it and said so in writing. That is the review that turns a working design into a proven one. The audit is not a rubber stamp. The auditor reads the design, runs the tests, and writes down what holds and what does not. If the design fails, the failure is published and the work stops honestly. That is what makes the word proven mean something.
Areas of Investigation
My research integrates principles from engineering psychology and cognitive science with contemporary machine learning to address questions of human performance, system safety, and the ethical dimensions of AI. Core threads include ergonomics and biomechanics, sensation and perception, human-computer interaction, computer vision and natural language processing, situation awareness, and the statistical and experimental methods that underpin rigorous evaluation.
I'm interested in the gap between how systems are designed and how people actually use them. That gap shows up in ergonomics, in the way attention and memory fail under load, in the decisions an AI makes on someone's behalf. My work pulls from cognitive psychology, human factors engineering, and machine learning to close that gap — with an emphasis on safety, usability, and the policies that should govern intelligent systems.
The only way to know it works is to put it up against things we know it can work against. We know it works here, in our local build: 102 tests ran, 102 passed. But our local build is not the proof. Proof takes a platform to test on in that manner and a server to test on in that manner, and we do not have either yet. That is what we are pitching for. The funding buys the platform and the server. Then we run the same tests live, against the same attacks, and we publish every result, misses included. That is how we will know.
The brake is the agent's conscience. The tether is the human's authority. A system with only a brake can be slowly reasoned into compliance. A system with only a tether is stopped re actively, after the damage is done. You need both, and the audit exists to prove both are real.
https://muse.ai/s/bwfp-side-by-side-lm6nxe2xhxixp5xj/g5h6UF4xfPwAq2uMhdElpA1vacz-QGgTDjuExV8BORg
Founder / Technical Lead: $85,000/year × 2 years = $170,000. Fringe at 25% = $42,500. Personnel total = $212,500. The founder is the sole person who holds the full containment architecture in their head.
Contract engineer, three months: $50,000 (~$95/hr). Turns the private scoring script into a public tool. Deliverable: a GitHub repo that says point this at your agent, run it against Agent-Safety Bench, get a number, publish it.
External safety researcher, advisory: $85/hr × 3 hrs/week × 52 weeks × 2 years = $26,520. Reviews the architecture, validates the evaluation design, and is the person who says "no, you cannot ship that yet." Output: a written review, 2–5 pages, part of the public record.
Dedicated server: $150/month × 24 months = $3,600. 4-core, 16 GB RAM, 500 GB NVMe. No GPU — inference runs through the API.
Compute and API credits: $6,000. Agent-Safety Bench (~$2,000–3,000), SWE-bench (~$1,500–2,500), Terminal Bench (~$1,000–1,500). The output is a published table. That table is the product.
Cohort operations: $3,900. Leaderboard hosting, builder on-boarding, and the public infrastructure for the first cohort of 3–5 solo builders, ~10 weeks.
Itemized total: $212,500 + $50,000 + $26,520 + $3,600 + $6,000 + $3,900 = $302,520.
Contingency and reserve: $54,480 (about 18% of the itemized total), for price changes in API credits, server costs, and the engineer contract, and for the unexpected costs of running the first cohort. Held in reserve. Anything unspent is returned to the funder.
Total: $302,520 + $54,480 = $357,000.
The goal on the form is $375,000. Manifund takes a 5% fee, so $375,000 raised delivers $356,250 to the project — essentially the $357,000 above, with the $750 difference absorbed in the contingency.
A detailed budget is available on request.
Anthony DeCicio, solo founder. Mechanical engineer by experience — CNC, fabrication, fleet systems — with a background in psychology and social-behavioral science. Self-taught in the ML and agent stack. The control layer is built from first principles, not inherited from an ML background.
The track record starts in 1991. From 1991 through the early 2000s, I developed simple systems to fix software glitches and computer glitches, rewriting and resolving fatal design flaws and crashes on Windows systems. In other words: computer repair, computer software repair tech.
The control layer is designed the way a mechanical safety system is designed: deterministic gates, append-only records, no single point of failure. An ML background would have built a probabilistic filter. The mechanical background built a gate.
The check sits inside the agent, not as a fence outside it. No single check authorizes an action alone. A check that disagrees stops it. Every attempt is logged.
Agent 1, the live public agent (https://101-contained-agent.vercel.app): live-tested November 8. Two tasks, both failed. The gate did not halt. Logged and open — the cause is not yet diagnosed.
Agent 2, private: never tested live. No live failures, and no live passes yet either.
Agent 3, local test build: 102 tests run, 102 passed, November 13. The test files are not yet published. Publishing them is the first thing this funding fixes.
The six-layer design documents are public at https://github.com/101AGILabs/six-layer. The agent wraps an open-weight model run through Groq. The model is not ours. The control layer around it is.
Built and tested alone, with no money, no outside review, and no external funding, on a low-end computer. No team yet.
If we finish early: if the infrastructure is complete within one year and the org is self-sustaining, the remaining budget is returned to the funder or directed to the open-source eval project. We expect to be self-sustaining within one to two years.
This purpose cannot stop. It will never stop. So the outcome of the project never ends. It still has to be done. The safety for humanity is needed.
$0 raised in the last 12 months. No grants, no donations, no other funding. This would be the first.
A $67,000 wage and tools claim is in process with BOLI against a former employer. That is wages owed, not funding, and it has not been paid.
There are no bids on this project.
Comments
No comments yet. Sign in to create one!