Independent research into AI reliability. No institution, no funding, no predetermined conclusions — just two years of sustained observation and documentation.
How do humans and AI work together reliably over time? The research approaches this from six interconnected directions.
This research is not affiliated with any university, research lab, or company. It is not funded by any grant, foundation, or corporate sponsor. There is no institution behind it — just one person doing extended knowledge work with AI systems and keeping careful notes about what works, what fails, and why.
Two years ago I started noticing something. AI collaboration worked well in short bursts — single sessions, contained tasks, fresh contexts. But it degraded over time. The same failure patterns kept appearing across different AI systems, different tasks, different working conditions. Context would drift. Outputs would propagate without verification. Shared understanding would erode. Models got better. The failures persisted.
The failures weren't model failures — the models were often performing exactly as designed. They were architectural failures. The system around the model wasn't built to maintain reliability over time. Nobody seemed to be studying this systematically.
So I started keeping notes. Then formalizing those notes. Then publishing them.
The methodology is observation-first. I conduct extended long-horizon AI collaboration — weeks and months of sustained work with AI systems on substantive problems — and document what happens in detail. Failure episodes are recorded as they occur, not reconstructed after the fact. Patterns across episodes are identified and formalized.
Over 18 months, this produced a corpus of 12+ documented failure episodes, several hundred hours of collaboration sessions, and the eight published papers that synthesize what the observations revealed. The papers are the formalization of field notes into peer-reviewable research.
This approach has a limitation: the sample is one researcher's practice. It has a compensating strength: the observations are grounded in real extended work, not laboratory simulations of collaboration.
A framework that is alive — not static, not finished, evolving with the system it describes. Most governance frameworks treat the problem as bounded: define the rules, implement the controls, declare the system compliant. The Living Framework takes a different view: reliable AI collaboration is an ongoing relational achievement, not a compliance state to reach once and maintain passively.
The "living" aspect is also methodological. The research itself is updated as the collaboration system it studies evolves. Papers 01 through 08 represent a progression — each one responding to what the previous ones missed or left unresolved. Paper 08's argument about verification architecture couldn't have been written before Papers 05, 06, and 07 established the theoretical ground it stands on.
RITAM — Foundational substrate research for governed AI cognition. Nine architectural primitives — State, Memory, Ontology, Governance, Epistemic, Coordination, Temporal, Observation, Repair — each justified by a failure mode. v1.1.1 runtime published open source. Founding principle: governance must precede persistence.
NIYOM — The reliability execution layer. Where RITAM is architecture, NIYOM is execution. A structured harness that forces every task through Plan, Execute, Verify, and Repair — with adversarial verification between generation and acceptance.
Alongside the research programme, I have authored twelve practical AI guides written for practitioners across different fields — managers, lawyers, teachers, HR professionals, recruiters, and administrators. These are not separate work. The prompts and frameworks in each guide are grounded in the same empirical observations that produced the published papers.