FactorForge technical whitepaper
Deterministic CPU-First Core.
Reproducible CDS design with explicit constraints, reviewable outputs, and deterministic optimization with explicit review boundaries.
Scientific benchmark
CAI optimization, reported with its corpus size.
Static benchmark figures are shown for technical context. They are not experimental validation or a claim of biological performance.
FactorForge
Mean CAI · n=49,257 CDS
Native reference (unoptimized)
Mean CAI · n=49,257 CDS
| Reference | Mean CAI |
|---|---|
| FactorForge | 0.94 |
| Native reference (unoptimized) | 0.72 |
Choose 1DP or experimental structure-conditioned 3DP.
Sequence-only design remains the baseline (1DP). Experimental 3DP uses reviewed structure positions to condition synonymous codon selection and compares its candidate with DP v2. Direct browser testing accepts an uploaded exact-match PDB/mmCIF chain and reviewed protein positions. It returns both candidates without automatically choosing a biological winner. Custom SOP management is paused; the built-in FactorForge default policy is applied.
The current direct test path uses the FactorForge API without ResearchOS approval or ValidationHub recording. The existing registered-input, approval and evidence workflow is retained for later operational use. Synthetic engineering tests do not establish experimental performance.
Watermark: disabled for current testing.
Version contract
One product release. Independently versioned engines.
FactorForge v3.5.11
Product versions describe the integrated FactorForge release. Engine versions describe each optimization implementation and do not advance automatically with the product version.
| Generation | Engine | Version | Status | Availability |
|---|---|---|---|---|
| Gen 1 | Rule / Profile | 1.0.0 | Stable | Public · default comparison engine |
| Gen 2 | DP v2 | 2.0.1 | Stable | Public · default feasibility path |
| Gen 2 | DP v2.1.1 | 2.1.1 | Research candidate | Public opt-in · calibration complete / holdout pending |
| Gen 2 | DP v2.2 | 2.2.0 | Research candidate | Local Python opt-in · initiation-aware reviewed position penalties |
| Gen 2 | DP v2.3 | 2.3.0 | Research candidate | Local Python opt-in · multi-class structural pause (3DP static & 4DP +35 aa window) |
| Gen 2 | DP v2.4 | 2.4.0 | Research candidate | Local Python opt-in · tropical semiring matrix recurrence & TASEP buffer smoothing |
| Gen 2 | ConstrainedLocalRewriter | 1.0.0 | Research candidate | Local Python opt-in · equivalence-preserving synonymous peephole & paired swap rewriter |
| Gen 3 | sLLM Hybrid | 0.2.0-preview.1 | Research preview | Feature gated · restricted deployment |
| Gen 3 | Partial-DP Rescue | 1.0.0 | Research preview | Feature gated · restricted deployment |
FactorForge v3.5.5 retains the versioned sequence-policy workflow and SOP Builder, improves standalone packaging and optional persistence configuration, and keeps stable DP v2 as the default. Product and engine versions remain independent; DP v2.1.1, discovery-slate, and sLLM results are computational evidence and do not establish biological performance.
Biological Guardrails
Findings and laboratory policy, kept separate.
FactorForge records deterministic sequence findings separately from the policy used to review them. A versioned SOP profile can classify supported findings as required, preferred, informational, or ignored while retaining profile provenance. The bundled conservative profile is an editable starting template, not a complete or approved SOP for any named laboratory, and advisory scans do not establish biological risk or wet-lab performance.
Architecture & transparency
Two lanes, one evidence boundary.
Production (Deterministic)
The CPU-first dynamic-programming path produces the same result for the same inputs and constraints. Provenance is part of every payload and rendered result.
Exploratory Lab (ML)
ML/AI comparisons are isolated as non-deterministic exploration. Any sequence from this lane must pass through Production validation before wet-lab synthesis.
Output provenance
{
"generator": "cpu_deterministic_dp",
"lane": "production",
"validation": "required"
}The generator flag makes the execution lane auditable in exported JSON and review packages.
DP v2.1.1 local-guard contract
Three axes, with local composition and evidence classes kept separate.
Assembly feasibility
Configured global GC, Type IIS motifs, and homopolymers of 6 nt or longer are enforced by deterministic search.
Codon adaptation
Host-relative codon weights contribute to the optimization objective.
5′ initiation
At layer 15, the active 45-nt GC count is pruned to 9–13 when reachable; the position-dependent proxy then ranks retained paths.
INDEPENDENTLY_EVALUATED covers post-generation measurements such as the ViennaRNA 5′ MFE value and 50-bp sliding-window GC. The anonymized two-chain calibration recovered 45-nt GC from 11.11% to 20.00% and removed ≥6-nt homopolymers while preserving amino-acid identity and configured BsaI/BsmBI/BpiI cleanliness. These are in-silico calibration observations; the 36-protein holdout and all biological outcomes remain pending.
Discovery slate · v3.5.0
Several research hypotheses, one deterministic evidence boundary.
Top-K slate
Versioned design contracts create a diverse pool; every emitted candidate passes the shared hard-constraint filter.
Constrained sLLM
A feature-gated learned prior proposes synonymous choices while automaton and translation masks retain veto authority.
Partial-DP rescue
A blocked path can preserve a verified prefix and solve an exact suffix conditioned on it; this is not a global DP optimum.
The discovery surface supports computational comparison and prospective experiment design. It does not guarantee expression, yield, synthesis success, or clinical performance.
Algorithm
Constrained CAI maximization, computed exactly.
For a fixed protein and a configured GC% target range, feasibility_best searches every synonymous CDS that encodes the protein and returns the one with the highest Codon Adaptation Index inside that range — an exact dynamic-programming solution, not a heuristic approximation.
Objective
Among all synonymous codon sequences whose GC% falls within the configured range, choose the one that maximizes CAI.
maximize CAI(C)
subject to GC_low <= GC%(C) <= GC_high
over C in synonymous(protein)Recurrence
Each DP v2 state is (codon position, cumulative GC-base count, configured-motif automaton state). It keeps only the highest-scoring path into every reachable state.
D[i][h] = max over codon c
of D[i-1][h - gc(c)] + log(weight(c))Fallback
If the configured GC window is unreachable, DP v2 returns the valid synonymous candidate closest to that band and reports the target as infeasible. Configured motif constraints remain active.
Tie-breaking
Equal-scoring paths keep the first codon encountered in the reference table's fixed order — deterministic by construction, not by chance.
Read the implementation notes
FactorForge documents its scoring contract, codon reference, and review boundaries so benchmark context stays attached to the software that produced it.