What changes
A new dataset still defines a new position. The timing distinction is the claim. Online AutoResearch begins a new propose–execute–score loop after the task appears. InnovationZero uses earlier scored algorithm games to change the first executable artifact. A bounded probe may then search, but its result is reported separately.
The algorithm game
- Encode the visible table and evaluation contract as state.
- Emit a legal executable pipeline.
- Run it under fixed resource and failure rules.
- Bind the outcome to the artifact and evaluator.
- Retain validated experience for a later opening policy.
A beautiful program that does not execute receives no score.
Policy and value
The policy proposes algorithm structure. The value signal comes from reproducible executions—not a model grading its own prose. The campaign ledger records 36 generations, 7 promotions, and 375,522 processed rows. Those are company-reported campaign totals, not a claim of exhaustive search.
Three evaluation modes
Strict first move freezes one global artifact before task-time selection. Bounded probe permits a limited task-time choice. Frontier diagnostic asks what the retained frontier contains. Each answers a different question and remains visibly separate.
Implementation boundary
“Alien code” is vision language. Any unfamiliar artifact still has to parse, execute, obey the contract, and survive an honest referee. The evidence here does not establish safety or generality outside the reported evaluation record.