You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
When I started using AI, I noticed something: the AI's dream. Not the dream we know, the hallucination. In mid-2025 the field was mostly trying to stop it. But a dream looks like an uncontrolled error more often than it is one, and my own dreams are usually where my work starts. So instead of blocking it, I thought about controlling it. I went to the deepest point, where hallucination is most dangerous: medicine. I built Wendy, a clinical decision-support demo, and developed the HALT levels. The level is never HALT 5; the strongest is HALT 4. A robot cannot be a doctor. I learned that I could never license Wendy, or even use it. But it had one usable part: the security system. HALT does not stop an agent, it redirects it, and it decides where to redirect by judging what the agent is doing. That judgement sits inside the conversation, and anything inside the conversation can be argued with. That is why the floor is not built on judgement.
The Floor That Cannot Be Lowered is a containment floor for AI agents that run with guardrails off. Every action costs a single-use key issued from outside the agent. Keys are consumed, never held. Every step is written to an append-only, hash-chained log kept by a witness. To take authority back you stop handing out keys. Anyone can verify a run from the frozen files and the published hashes, without the lab's network. Reference code: https://github.com/bioman35/containment-floor (Apache-2.0). Report submitted to the Apart Research AI Incident Response Sprint, Track 1, 13 September 2026.
If this is funded, the first thing I do is test everything I have built so far, and where something is missing or wrong, I close it. The gaps I already know about get closed the same way, by testing them.
Goal: make "did the agent stay inside its box" a question an outsider can answer from published artifacts, not from the lab's word. Six months, four deliverables. 1. Egress attestation: a signed network policy plus a flow-log digest, with an open verifier. 2. Comparison set: the 18 clauses of the standard against the 9 phases of the July 2026 incident, cost per clause measured, plus a null-model control. 3. A pilot at a small operator, with an unrelated third party verifying the run from the published hashes only. 4. A short paper and a public verification guide. Everything is released under Apache-2.0 in the same repository.
Total 18,000 USD: 520 hours of my time at 25 USD (13,000), API and cloud (2,000), three external verifier runs (900), pilot operator setup and support (1,100), publication and presentation (1,000). Minimum 6,000 USD funds deliverables 1 and 2 only. Payment goes to the registered business of Bilal Yüksekdağ, which operates Triamind's products. Disclosure: the same six-month plan is pending with the Apart sprint fund, Lightcone Commons, Mercor's AI Safety Fund and the AI Alignment Foundation; a three-month pilot-only subset is pending with BlueDot Rapid Grants; a 12-month follow-on application is being prepared for the Foresight Institute; and a 12-month career transition enquiry is with Halcyon Futures, covering the same deliverables over a longer period. If any of them funds a deliverable, the amount here drops by that deliverable. No deliverable is funded twice.
I work alone on the build. I am a biology teacher by training (Dicle University, 1998 to 2003), founder and systems architect at Triamind in İzmir. I build and operate a multi-agent production line whose rule is that no agent's report is accepted until an independent measurement confirms it. Track record on this project: floor.py and group_head.py frozen with SHA-256 hashes; an 18-clause standard with per-clause evidence; six adversarial scenarios; a 246-call comparison set with blind judges; and identical decision lines on 3 operating systems and 6 Python versions. Those runs were done by four people: me, two of my children stated openly, and one runner I have never met, a university classmate of my son who agreed to run the frozen files during the holiday.
My two children study engineering. I am training them in this work so that they can support me. This lightens my own load, and it opens the road for them to become engineers who build better security systems than I do. Their relationship to me is stated openly in the report and in the repository. The pilot's verifier will be someone unrelated to me.
The most likely failure is not finding a small operator willing to host the pilot; then deliverable 3 slips and the money for it is returned or not drawn. The second is a result I do not want: if the comparison set shows the floor's cost per clause is too high for real workloads, or an unrelated verifier cannot reproduce a pilot run from the published artifacts alone, the claim fails and I will publish that as the result. Known limits already stated: the floor protects a supervised run, not a self-hosted open-weight model; a captured witness, a zero-day below the floor, or observations fabricated before they enter the chain are outside what it can promise.
One more thing about failure. I do not stop working. Suppose the time runs out and I did not finish. That is not like me, but suppose it happens. I cannot give up until I solve it. I would find a way to finish it.
Nothing. No grant funding to date. The work so far was done in the evenings alongside a public-sector job, with API costs paid out of my own pocket. Applications submitted in the last week are listed above; none has been decided yet. I write and think in Turkish. The substance, the decisions and the corrections in this proposal are mine, sentence by sentence; where the text was wrong, I was the one who caught it. The English was drafted and edited with AI assistance and reviewed by me. The code, the runs and the published hashes are checkable by anyone.