Built from psychology, social behavior, and human factors engineering — a system that thinks, feels, and acts from within its own safety, at every layer.
- This project proves one claim: an agentic system can carry its own safety controls within its layers, so no external sandbox is required.
- The key distinction: a sandbox can be bypassed because it is external. A role can only be confused, and confusion is a human-factors problem — fatigue, ambiguous instructions, cognitive overload — not an adversarial problem. Human-factors problems are solvable with existing engineering methods: checklists, redundancy, cognitive load limits.
- Containment is intrinsic, not extrinsic: the system holds its own control layer inside itself.
- Two systems are live and proven in public, with test results published, failures included. A third is being built.
- The work is grounded in psychology, social behavior, and human factors engineering.
- The goal is superintelligence within safety: a system that grows in knowledge and wisdom while its safety controls hold at every layer.
- Patent application in preparation. Dated Declaration of Invention signed November 19, 2026.
- Goal 1, Specify: write down, in formal terms, what safety inside the system actually is. For each of the five layers, the exact constitutive constraint that makes the system what it is, the test that proves the constraint is real, and what breaks when the constraint is gone. Months 1–6.
- Goal 2, Implement and validate: build all five layers into a small model, 7B to 70B parameters, and run the five tests with no sandbox at all. It has to pass all five and beat a system trained only with RLHF on the same tasks. Months 7–18.
- Goal 3, Multi-agent safety: put two safe systems together and prove the pair is still safe. Neither one can reach anything forbidden alone, and together they still can't — the overlap of what they can do holds nothing forbidden. Months 12–24.
- Goal 4, Safety case: write the formal safety case the way aviation and nuclear do it. The claim, the five arguments, the evidence, and the residual risk counted, not asserted — under one in a million per Tier 3 action — for the exact operational envelope it covers. Months 18–30.
- Goal 5, Transition protocol: write the step-by-step protocol for retiring the sandbox. Run both together, measure, and retire the sandbox only when the inside layer has proven itself over time. The human override never reaches zero. Months 24–36.
- We will know it worked when the five tests pass with no sandbox at all: counterfactual removal, assume-compromise evaluation, SandboxEscapeBench as negative control, non-compositionality testing, and compositional interpretability audit.
- We will know it worked when the numbers prove it: the intrinsic layer blocks everything the sandbox would have blocked, and more — 12 consecutive months with zero sandbox-only blocks, then 6 months in shadow mode with zero shadow-sandbox blocks.
- We will know it worked when independent reviewers test it, not just us. Multiple people and organizations, running the same tests, reaching the same result.
- We will know it worked when an independent reviewer reads the safety case and concludes the system is safe to deploy without a sandbox for the specified operational envelope, and the residual risk is acceptably small.
- We will know it failed if the interpretability gap at frontier scale prevents verification of the inside controls. That failure is a valid, publishable result — it tells the field exactly where the boundary of the approach is.
- Hire the builders: engineers to build the five layers into the small model, and a safety researcher to run the five tests independently, so the testing is not done by the same hands that built it.
- Build the third system and the platform it runs on, so the tests run on real systems, not on paper.
- Run the transition protocol: dual-stack deployment, measurement of concordant, unique, and sandbox-only blocks, and shadow mode, until the sandbox can be retired on evidence.
- Publish every result, failures included, and write the formal safety case for independent review.
- The team is a four-man crew: Tony and the four chatbots he builds with interactively.
- Track record: two systems live and proven in public, with test results published, failures included. A third is being built.
- The systems run an open-weight model through Groq. The model is not ours. The control layer built inside the system is ours.
- Independent testing: 17 of 20 on a blind set, judged by an independent reviewer. Local testing: 102 of 102. The one live failure, November 8, 2026, is published with the test record.
t are the most likely causes and outcomes if this project fails?
- $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.
Comments
No comments yet. Sign in to create one!