Novelty search

Context and Problem Statement

Arbor queries alphaXiv to ask whether an idea is academically novel, and uses that verdict to steer its exploration. The 2026-07-23 jobs review (reviews/2026-07-23-jobs/) proposed wiring the same lookup into the kit: before committing to an approach, check the literature and prefer what has not been published.

It returns under several phrasings — novelty search, a prior-art check, an alphaXiv or arXiv lookup, “has anyone solved this already”, a research-literature gate — and all of them share one unstated premise: that the kit is choosing between ideas at all.

Recorded as French prose in docs/backlog.md §Non-adoptions until this ADR replaced it.

Considered Options

• Wire an alphaXiv novelty verdict into the assessment phase.
• Keep a weaker version: a literature link in the report, informational only.
• Decline: novelty is not an input to any decision the kit makes.

Decision Outcome

Not pursued by decision (2026-07-23): an academic novelty verdict improves no migration. The kit’s decisions are “does this build”, “do these tests pass”, “is this API still supported”, “what does the analyzer say” — every one of them answered by running something against this codebase, none of them by what a paper does or does not already describe. A novel upgrade path and a well-trodden one are worth exactly the same here, and the well-trodden one is usually worth more. The weaker “informational link” option is the same cost — a network dependency, a rate limit, a failure mode — bought for output nobody acts on.

Consequences

The kit reads the codebase, the toolchain and its own gates, and never the literature; an assessment never blocks or slows on an external research API.

The one condition that reopens this: the kit starts optimising an open metric where “has this been tried” is genuinely decision-relevant — the same condition that governs ADR-0008, because it is the same missing premise. Reopening is a fresh ADR plus set_status deprecated here, never an edit in place.

Prior requests

• reviews/2026-07-23-jobs/ — “Recherche de nouveauté (novelty search alphaXiv)”: an academic novelty verdict or prior-art check, read from alphaXiv papers and the arXiv research literature, used to rank candidate approaches — “has anyone solved this already” as a gate (2026-07-23).