Eijex editorial archive
What we learned while building reproducible, evidence-aware bio-design software—including investigations that failed, assumptions that changed, and work that remains incomplete.
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.
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.
An early architecture note assigned models to future roles before the runtime and evidence needed to support those roles existed.
A reference table shaped the outputs, and those same outputs had been used to justify the range that judged them.
Two CAI values for the same sequence were both mathematically correct—and still misleading when shown without their references.
Codon choice, protein architecture, experimental context, and measured yield belong to connected—but different—evidence layers.
A tracer workflow completed successfully, but its computational values were fixtures rather than live outputs.
Moving JSON records into tables can silently discard identity, append-only history, approval, and outcome-linkage semantics.
Table presence showed that future provenance could be represented; it did not prove migration replay, leakage control, access policy, or end-to-end lineage.
Why host preferences, deterministic findings, and laboratory SOP decisions need separate versioned records.
The external model never produced a candidate in this trial, but the failure exposed practical reproducibility costs that benchmark tables often omit.
An early comparison was labeled as ML benchmarking before the experimental decoder actually consumed trained-model logits.
Conflicting runs showed why a partner-derived case cannot become a publication result until seed, reference, commands, hashes, and validator outputs are frozen together.
Why FactorForge now separates learned exploration, deterministic vetoes, exact suffix rescue, and evidence promotion.
Why sequence-optimization metrics belong in the input feature space while closed-loop surrogate models must train on normalized experimental outcomes.
Two plausible debugging theories were wrong. A traceback showed that an unbounded cubic-time folding call was the real cause.
Throughput counters said every combination had run, while a required artifact was absent and a profile label overstated what its output achieved.
A branch trigger, committed-byte mismatch, cross-repository version drift, and missing downstream CI appeared outside the task's acceptance criteria.
Why a deterministic sequence check, a user policy and a measured biological outcome must remain different kinds of evidence.
What the v3.5.4 maintenance work taught us about standalone installs, recorded UI explanations and deployment evidence.
Why FactorForge now separates researcher decisions, sequence-bearing reports, and sequence-free evidence—and what must exist before that history becomes a controlled audit trail.