Noema
An intermediate representation for moving meaning between models without pretending uncertainty disappeared.
The problem
Two models can understand the same task while using very different internal conventions. Passing raw prose between them wastes context and often drops constraints, uncertainty, and references.
Noema is a semantic intermediate representation. It is not meant to replace natural language. It is a compact transport format for intent, context, references, constraints, and confidence.
Objective
The working objective balances three costs:
L(y | C, R) + lambda * D_T(M, decode_R(y, C)) + mu * K_R(y)The first term pays for the encoded representation. The second measures how much task meaning is lost after decoding. The third pays for decoder complexity. A shorter message is not better if the receiver needs a giant private rulebook to understand it.
Current shape
The work includes a parser, compiler, validator, terminal translator, Prompt IR, anchors, reference tokens, and domain examples. Validation records what survived exactly, what was approximated, and which decoding context is required.
Boundary
Noema does not claim a universal language for every model. The useful question is narrower: can a representation preserve enough recoverable meaning to make model-to-model work cheaper and more reliable than raw prose?