What a decoder comparison can—and cannot—show
An early comparison was labeled as ML benchmarking before the experimental decoder actually consumed trained-model logits.
The attractive interpretation
Two codon-selection paths produced synonymous CDS candidates with preserved translation. Their codon choices, nucleotide identity, CAI, and GC values differed. It was tempting to call this a Rule-versus-ML comparison and attribute the difference to learned sequence context.
The code-level audit did not support that interpretation. The experimental decoder was operating without trained-model logits. It selected among synonymous codons through deterministic scaffold logic and masks. The numerical comparison was real; the label attached to its cause was not.
What remains valid
The run can still demonstrate several integration properties:
- both paths can preserve amino-acid identity;
- two deterministic policies can choose different synonymous codons;
- the shared metrics can describe those candidates; and
- hard masks can prevent selected forbidden outputs in the exercised path.
It cannot demonstrate learned contextual generation, model superiority, expression potential, or biological ground truth.
Why the superseded result stays visible
Deleting the earlier interpretation would hide the most useful lesson: output differences do not reveal the mechanism that caused them. Mechanism claims require inspection of the active code path and provenance of the model artifact.
A future Rule/LM comparison must show that frozen model weights are actually loaded and used, then evaluate Rule-only, LM-only, and hybrid paths under the same validator, metric authority, inputs, and held-out protocol. Until then, this note remains a corrected scaffold comparison.