Runtime Architecture

Protean’s runtime is the operating layer that keeps discovery moving. It coordinates evidence intake, candidate generation, validation, ranking, interpretation, learning, and review as a continuous system.

The runtime is built for controlled autonomy. It is not an open-ended self-training loop and it is not a chatbot workflow. Each cycle has defined responsibilities, scoring contracts, and system boundaries.

Runtime Responsibilities

• Maintain candidate flow.
• Preserve constraint discipline.
• Route model capabilities by task.
• Keep failure memory active.
• Produce reviewable candidate state.
• Apply bounded learning only inside defined limits.

Cycle Contract

Every cycle follows a bounded contract:

ingest
-> extract entities
-> index evidence
-> normalize
-> generate
-> validate
-> score
-> explain
-> check claims
-> learn within caps
-> rerank once
-> write research cognition artifacts
-> snapshot cycle
-> package for review

The runtime does not self-modify code, recursively retrain itself, or remove failure motif penalties. It can improve prioritization only through constrained adjustments that remain visible to the review layer.

Runtime State

The platform tracks state so repeated cycles remain operational rather than opaque:

• Source and evidence availability.
• Candidate counts and rejection reasons.
• Model route availability.
• Embedding availability.
• Scoring weight normalization.
• Failure-memory influence.
• Learning mode and bounded adjustment status.
• Explanation coverage and warning burden.

This state allows the platform to remain operationally observable without becoming uncontrolled.

Model Routing

Model routes are selected by task. Local reasoning models can support proposal, extraction, failure reasoning, and explanation. Protein sequence models can support embeddings and similarity context. Deterministic fallback paths keep the runtime reproducible when models are unavailable.

Why It Matters

Peptide discovery is not improved by generating more sequences alone. The advantage comes from sustained orchestration: stronger constraints, faster rejection, better memory, and clearer handoff into scientific review.

The same runtime state also supports Protean’s broader network architecture. Reviewed collections, provenance commitments, lifecycle state, and public-safe lineage can become durable coordination surfaces without exposing private scientific payloads.