When an integrity framework stops its own best-looking result
A six-arm benchmark prototype produced attractive demonstration values, but provenance and evidence checks showed why they could not yet be treated as scientific results.
The attractive result
The benchmark architecture was meant to compare four design paths with two diagnostic controls. On paper, that is exactly the comparison a sequence-model study needs: a historical reference, statistical and dynamic-programming methods, a real language model, and controls for random synonymy and host codon frequency.
The prototype also displayed a striking advantage for the nominal model arm. That was the moment to slow down. Inspection showed that the displayed values were demonstration fixtures, not outputs computed by six executed methods. They had no replicate distribution and no statistical test. Publishing the numbers as performance would have converted interface scaffolding into evidence.
What the integrity checks found
The leakage-audit workflow had the right two-stage shape—a fast candidate filter followed by local alignment—but its current data were placeholders rather than the frozen corpus used to train a model. An exact-match result from that fixture could say nothing general about memorization.
The historical-reference check found a second boundary: the accession attached to the intended fluorescent-protein comparison did not identify the asserted construct. Without an unbroken publication-to-accession-to-feature mapping, any derived sequence-proximity comparison had to be withheld.
These were not merely missing details. They were failures of the evidence gates that determine whether a number may become a claim.
The architecture that remains useful
The work still produced a clear governance model:
- AI proposes candidate sequences.
- Deterministic algorithms optimize and validate what can be checked directly.
- Hard rules retain terminal authority over feasibility.
- Fuzzy sets describe soft suitability without creating a universal score.
- Humans approve dispositions, and evidence remains attached to the exact object that was evaluated.
The fuzzy-policy idea follows the multidimensional reasoning introduced by Zadeh's fuzzy-set formalism. It does not turn a computational profile into biological truth. Experimental variation in plant transient expression remains a separate empirical problem, as illustrated by recent controlled work in Nicotiana benthamiana.
What comes next
A result-bearing run needs a frozen trained model, the exact training corpus, a verified historical sequence, repeated diagnostic controls, fixed metrics, and uncertainty analysis. Wet-lab performance requires a prospective experiment after that computational gate.
For now, the most useful result is procedural: the framework stopped its own best-looking numbers from being promoted beyond their evidence.