Reversible representations
before the model.
What changes when representation is treated as an explicit reversible machine instead of an invisible preprocessing footnote?
Representation is part of the machine.
OKC explores reversible transformations between raw data, structured token or ID streams, explicit sidebands and model-facing representations. Exact recovery is the hard constraint. The open question is whether exposing structure before prediction or entropy coding can improve inspectability, vocabulary pressure, latency, memory or downstream modeling in bounded regimes.
1. Construction premise
A representation should be measurable like any other machine: declared input, state, output, inverse, delay, memory and failure conditions. Compression ratio alone cannot describe deployment value.
2. Public proof trail
DU3, Rare1, ID-stream tooling and related experiments are partial evidence, not one completion claim. Useful records include exact reconstruction, dataset identity, benchmark boundaries and adverse results.
3. Current questions
Which reversible transforms admit bounded execution? When does an explicit sideband reduce active-model pressure enough to justify itself? When should a model operate globally, and when should global identity be separated from local execution? What changes when the downstream consumer is another machine rather than a human parser?
4. Module direction
The working architecture separates container/scope, structure, optional filters, optional models and entropy. Composition is encouraged; hidden coupling is not. A successful experiment should be extractable as a pipe with a named inverse rather than becoming another permanent integrated codec.
5. Boundary
This public note intentionally stops before private scheduling, reconciliation and implementation details. Deeper prototypes and scoped technical transfer are separate research or commercial engagements.