The runtime exists. 146 tests pass across adversarial, integration, and buildability scenarios. Five independent AI systems reproduced the substrate from the specification alone — 60/60 tests passing. The constraint RITAM held itself to: a specification that does not run does not count. It now runs.
Nine primitives. Each justified by a failure mode.
RITAM investigates the structural conditions required for reliable cognitive state in AI systems. It is not a chatbot, product wrapper, or orchestration layer. It sits beneath applications and asks what must be true before memory, belief, confidence, time, coordination, and repair can be treated as dependable system state.
The core principle is admission before persistence: state should be checked before it is retained, reused, or allowed to shape later reasoning. RITAM formalizes this through nine primitives that make cognitive state inspectable, constrained, and repairable.
What RITAM is working to answer
Built on the published papers
RITAM is grounded in the specific findings of eight published papers. The architecture is not speculative: it is derived from documented observations about reliability, memory, protocol failure, language reliability, distributed cognition, and verification.
RITAM's own experimental programme has also produced findings documented independently of the published papers, including pre-registered experiments under adversarial review and a prior-art audit of the core claims.
Primary foundations:
Substrate and execution
RITAM and NIYOM are distinct but related projects. RITAM is substrate research: it studies the reliability conditions beneath persistent cognitive state. NIYOM is execution: it turns reliability architecture into a practical harness for planning, verification, repair, and repeatable AI workflows.
RITAM stays below the product layer. Its job is to clarify what reliable state requires before product systems attempt to implement or depend on it.
v1.1.1 is live. What that means.
The runtime substrate is published open source under Apache 2.0. 146 tests pass across adversarial, integration, and buildability scenarios on Python 3.9, 3.10, and 3.11. Five independent AI systems reproduced the runtime from the specification alone — 60/60 tests passing — confirming that the specification is precise enough to build from without guidance from the original author.
Three consumer types are demonstrated: reliable research assistant, reliable decision log, and reliable agent memory. The substrate is domain-agnostic: the same nine primitives constrain cognitive state regardless of what the application does with it.
The canonical specification continues in active development. The runtime demonstrates that the architecture is correct and buildable; the specification provides the theoretical account of why it is the right architecture and what it must preserve to stay right.