01
Propose
A future trained model may generate synonymous CDS candidates.
← FactorForge
Exploratory research track
Language-model proposals.Deterministic boundaries.
FactorForge sLLM Hybrid 0.2 is a feature-gated Gen 3 research-preview track in v3.5.11. It proposes synonymous CDS candidates while deterministic vetoes, versioned partial-DP rescue, evidence gates, and human approval remain outside the model's authority.
The MCP contract may advertise this feature-gated preview, but that does not represent a public trained-sLLM endpoint or a production-model claim.
Learning workflow & evidence boundaries
Local training summaries are distinct from independently verified evidence and live service integration.
The problem
Confidence is not a constraint.
A probabilistic decoder can assign high likelihood to a sequence that still contains an internal stop, changes the encoded protein, or conflicts with an assembly policy. Those properties can be checked directly; they should not depend on model confidence.
The FactorForge direction therefore separates proposal generation from authority. AI may suggest. Deterministic algorithms measure. Rules decide feasibility. Humans decide use.
The governed path
Five responsibilities. No hidden handoff.
02
Measure
One shared evaluator computes versioned, inspectable metrics.
03
Enforce
AgentOS rules reject candidates that violate deterministic constraints.
04
Interpret
Soft metrics remain a profile, not a universal biological score.
05
Approve
A human decides whether an eligible candidate moves forward.
AI proposes. Algorithms optimize. Rules guarantee defined invariants. Fuzzy logic interprets. Humans approve. Evidence is preserved.
20-target engineering fixture
A useful pipeline test. Not yet a model benchmark.
The current artifact contains 102 target-arm rows across a mixed panel of 20 protein fragments and domains. It is useful for testing table shape, orchestration, and claim gates. Code inspection shows that it does not execute a trained SLM or canonical comparative evaluation.
102
recorded target-arm rows
An artifact count, not 102 validated biological designs.
0
generation-cache hits
The run cannot support a cache-driven runtime claim.
Why the attractive charts were withheld
The displayed CAI was computed directly from GC-base counts, the hard gate scanned only one restriction motif, and the nominal SLM arm used deterministic codon substitutions. A scatter plot or 100% bar derived from those fixtures cannot establish biological fitness, hallucination elimination, or superiority.
The proof standard
Promotion requires evidence, not a label.
We will call it a Real SLM benchmark only after the model, inputs, metrics, controls, and validation path can be independently replayed.
- 1Frozen model weights, tokenizer, decoding configuration, and seed
- 2Verified protein inputs and publication-to-accession reference mapping
- 3Canonical CAI and GC computation shared by every comparison arm
- 4Translation, internal-stop, restriction-site, and assembly-rule validation
- 5Repeated controls, uncertainty analysis, and prospective wet-lab testing
Build on the available surface
Use today's deterministic tools. Follow the SLM evidence as it matures.
Eijex MCP provides agent-compatible access to documented FactorForge capabilities. The sLLM path remains a gated preview until a trained endpoint and validation evidence satisfy the promotion gates.