Why rule authority still matters in a language-model system
A code-level audit changed the question from whether a language model could replace deterministic design rules to how learned proposals could remain bounded by them.
The expectation
The original direction was easy to describe: train a sequence model, let it learn host-specific patterns, and use it to generate better coding sequences. That idea was attractive because it appeared to move beyond hand-authored heuristics.
The implementation audit forced a more precise question. The experimental decoder path did not yet contain trained-model inference. Without model weights providing contextual probabilities, it was a deterministic constrained decoder: useful as an integration scaffold, but not evidence of learned generation.
What the scaffold still taught us
The gap was not wasted work. It clarified which responsibilities should remain deterministic even after a trained model exists:
- preserving amino-acid identity;
- rejecting invalid or internal-stop outputs;
- applying explicit assembly constraints;
- recording the reference, configuration, and validator contract; and
- abstaining or falling back when a proposal cannot satisfy the contract.
These are not properties that should depend on a model being confident. They are properties that can be checked directly.
The revised position
The production direction is therefore not “rules or a language model.” A future model may propose or rank candidates, while deterministic rules remain the feasibility authority and fallback. An LM-only path can remain useful as a research comparator, but it should not be confused with the reviewed production path.
This distinction also changes how results are reported. A decoder scaffold is labeled as a scaffold. A trained model requires frozen weights, tokenizer, dataset, split, configuration, and held-out evaluation before it is described as learned generation. A computational pass remains a computational result; it does not imply expression, yield, or biological validation.
What remains open
The next meaningful milestone is not another architecture label. It is a frozen comparison in which Rule-only, LM-only, and Rule/LM paths use the same inputs, validator, metric authority, and provenance record. Until then, the honest result is the architecture lesson itself: learned proposals can be valuable, but deterministic feasibility remains independently testable.