From a single optimum to governed hypothesis slates

The limitation of one best-looking sequence

A single optimum is useful when the objective is complete and the constraints are known. Research often needs a different artifact: several feasible hypotheses that differ for traceable reasons. FactorForge v3.6.0 introduces a discovery-slate path while retaining the stable deterministic optimizer as the default.

Four responsibilities, kept separate

The new path divides work into four layers:

1. Several versioned design contracts create a candidate pool.
2. A feature-gated learned prior can propose synonymous codon-context hypotheses.
3. Deterministic translation, motif, composition, and homopolymer checks retain veto authority.
4. When generation reaches a dead end, exact dynamic programming can repair a suffix conditioned on a preserved, already checked prefix.

A suffix optimum conditioned on a fixed prefix is not the same claim as a globally unconstrained DP optimum. The exported lineage must say which path was requested, which path actually produced the candidate, and whether rescue occurred.

Why partial rescue is scientifically interesting

Earlier fallback behavior could discard an entire learned prefix when one later choice made the sequence infeasible. Partial rescue asks a narrower question: how much verified context can be preserved while deterministic search restores feasibility? A multi-arm benchmark can compare unconstrained generation, veto-only generation, global fallback, and partial rescue without attributing every difference to “AI.”

The evidence boundary

These artifacts measure computational feasibility, provenance, and sequence diversity. They do not establish expression, yield, synthesis success, or clinical utility. Prospective experiments require separate protocols, controls, observations, and promotion gates.