Novelty

novelty checks whether a proposed edit is sufficiently different from prior accepted candidate diffs. It is optional and runs before spending a canonical evaluation on a near-duplicate.

Contract

class NoveltyOperator:
    def assess(self, checkout, ctx) -> NoveltyResult: ...

The result contains a novelty score, where 1.0 is wholly novel and 0.0 is an exact duplicate, plus an acceptance boolean.

Variants

Variant	Rule
accept_all	accept every candidate edit and disable deduplication
diff_similarity	compare the candidate diff with recent accepted diffs and reject excessive similarity

Configuration

operators:
  novelty:
    variant: diff_similarity
    threshold: 0.98
    history_k: 8
    timeout_s: 600

• threshold is the similarity level at which an edit is considered too close to prior work.
• history_k bounds the accepted history examined.

Use accept_all when novelty is not part of the method contract. Adding or changing novelty filtering changes which candidates receive evaluator budget and therefore changes the experiment policy.