Idea-tree search

Context and Problem Statement

The 2026-07-23 jobs review (reviews/2026-07-23-jobs/) read the kit through the Arbor (RUC-NLPIR) lens and asked, in one breath, for two different things. The kit took the first: Arbor’s safety belts — resume, the convergence guard, measured chronology, contractual backpropagation — shipped in v1.8.0. The second was Arbor’s steering wheel: a multi-hypothesis idea tree, where the agent branches several candidate approaches, scores them against a metric, and keeps exploring the promising ones.

It keeps being asked because the pitch is genuinely appealing and the vocabulary is fluid: idea tree, hypothesis tree, tree-of-thought, branch-and-score, multi-hypothesis exploration, beam search over approaches. Under any of those names it sounds like a strict upgrade — the kit “only” walks one path, so surely walking several is better.

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

Considered Options

• Adopt the idea tree: branch on approach, score, prune, keep the best.
• Adopt it only in the assessment phase, where the destination is genuinely open.
• Decline it, and say once, durably, why.

Decision Outcome

Not pursued by decision (2026-07-23): Arbor explores an open space — a metric to maximise, the best solution unknown — while the kit executes a known path to a binary destination (build green, tests green, production verified). Grafting exploration onto that destroys the property that makes the kit worth running: deterministic, reproducible, in measured minutes. A migration that branches is a migration whose cost and outcome nobody can quote in advance, and the second option — “only in assessment” — is the same trade with a smaller blast radius rather than a different one, since the assessment’s output is the plan every later phase is checked against.

Consequences

Every phase stays single-path, and a phase that cannot reach its gate stops and says so rather than trying a sibling approach — the failure is loud instead of expensive.

The one condition that reopens this: the kit starts optimising an open metric — performance, bundle size, cost — where there is no binary gate to reach and exploration therefore pays for itself. Reopening is a fresh ADR plus set_status deprecated on this one, never an edit in place. “A migration would probably be better with it” is not that condition, and neither is a new name for the same mechanism.

Prior requests

• reviews/2026-07-23-jobs/ — “Arbre d’hypothèses / recherche multi-hypothèses (Idea Tree)”: multi-hypothesis tree search that branches candidate approaches, scores them and explores an open solution space; also asked for as tree-of-thought, branch-and-score, or beam search over approaches (2026-07-23).