Candidate review pipeline

From evidence to review-ready state

The pipeline begins by transforming source material into evidence records. Those records inform constraint composition, failure memory, and candidate context. From there, candidate state builds across several passes:

• Generation creates a controlled candidate field inside the constraint surface.
• Validation rejects candidates that fail deterministic gates (residue validity, motif burden, cleavage exposure, composition baseline).
• Feature assembly adds sequence descriptors, structural features (hydrophobic moment, AGGRESCAN, TOP-IDP, Boman, Shannon entropy), and embedding context.
• Ranking orders candidates across the seven canonical scoring axes.
• Explanation generates per-candidate rationale, warnings, and risk context.
• Claim QA checks generated statements against the local evidence index and flags unsupported language.

The pipeline emits one bounded slate per cycle, not an open-ended stream.

What lands in the review package

A review package contains, for each prioritised candidate:

• the candidate identifier and its score-band position
• the ranking rationale, broken down by scoring axis
• the warning burden (cleavage exposure, novelty pressure, failure-similarity proximity, synthesis-risk heuristics)
• the source-aware explanation context from retrieved evidence
• the claim-QA classification for every generated statement
• the IP-oriented review notes (when applicable)

The package is sufficient for a human reviewer to argue with — to advance, hold, or reject the candidate with reasons.

What the package is not

The package is a review artifact. It is not a clinical claim, not a patentability determination, not an assay result, and not a decision. The pipeline's job is to make candidate selection more coherent. The decision sits with a human.

The runtime does not order assays from this package. Assay-preparation handoff is drafts-only, and provider packets pass through the external-provider safety gate before they leave the machine.