Research · Representation to output
Stable representation does not guarantee structurally admissible output.
By this stage, the research had established reproducible ordering effects, selectable operating regimes and a robust first-order representation hierarchy.
The unresolved question was whether those internal structures reliably survived generation.
The transfer test
Decoder outputs were compared against the representation-level ordering and regime structure under the evaluated configurations.
The expected monotonic transfer did not appear. Representation-level properties — ordering stability, regime differentiation and geometric coherence — did not reliably propagate into decoder output as structural admissibility.
Finding: representation-level improvement was necessary but not sufficient for output-level invariant control.
Representation
Does the model encode a reproducible structural distinction?
Output
Does the generated result preserve the critical invariant required by the application?
Why this is a central result
This result separated two questions that are often implicitly merged:
The answer to the first can be positive while the answer to the second remains negative or unresolved.
From Corpus D to Corpus D2
Corpus D therefore could not answer the next practical question. It described representation behaviour.
Corpus D2 was created on the other side of the bottleneck: its unit of analysis was observable output failure.
Corpus D asks what happens inside representation. Corpus D2 asks what actually fails in generated output.
Architectural consequence
The transfer boundary forced SIF to stop treating model-internal structure as sufficient evidence of output safety.
That decision leads directly to post-generation structural inspection, output failure classification and bounded repair.