Move 37
Search begins after the new board is revealed.
Verifier-driven research / active paper v1
The first move becomes an algorithm.
InnovationZero is reported to learn from earlier scored algorithm games—changing the first executable artifact before a new task-time search begins.
Paper SHA-256
7a80bbfac4f20a9db7b94bd46111fa6262d72bc65dbcf33c3605a484f68dabce
One referee.
Two timings.
Move 37 discovers inside a game. Move 42 asks what experience can change before the next game opens. Scroll forward, or use J and K to move chapter by chapter.
01 / The game
A model submits one executable answer
The board is not a metaphor for taste. It is a contract: data enters, code executes, an honest referee scores what survives.
Both
Both systems must produce legal artifacts under the same verifier.
Move 42
Earlier scored games alter what can be played before new search begins.
02 / State
Data becomes a position, not a prompt
A table defines constraints: shape, types, missingness, target geometry, and the evaluator that will reject invalid play.
Move 37
Read the new position and form a task-specific plan.
Both
Observe the same visible state and hidden evaluation boundary.
Move 42
Begin with a learned policy over executable algorithm structure.
03 / Move
The move is code
A proposal counts only when it parses, trains, predicts, and returns the required shape within the referee's limits.
Move 37
Generate or edit code during task-time search.
Both
Pay the real cost of execution and receive measured outcomes.
Move 42
Emit a first artifact shaped by prior algorithm games.
04 / Policy
Experience changes the prior
InnovationZero is reported as learning from histories of executable candidates and their scores, not from prose descriptions alone.
Move 37
A general code model supplies the initial proposal distribution.
Both
Can still use a task-time evaluator loop.
Move 42
Algorithm-game outcomes train the policy that supplies the first artifact.
05 / Value
The score must mean the same thing every time
A useful value signal is bound to the artifact, dataset split, evaluator, and failure rules. Unverified elegance receives no credit.
Move 37
Use fresh evaluations to estimate which branch is promising.
Both
Rely on the referee rather than self-assessment.
Move 42
Carry learned value from earlier games into the opening choice.
06 / Search
Beyond AutoResearch is a timing claim
Online AutoResearch starts a new task-time search. InnovationZero uses earlier scored algorithm games to change the first executable artifact; bounded search remains a separate mode.
Move 37
Start the loop now: propose, run, inspect, revise.
Both
May improve when granted additional task-time evaluations.
Move 42
Change the first legal program before the new loop starts.
07 / Score
Three modes, three honest labels
Strict, bounded, and frontier numbers answer different questions. Combining them would turn a useful comparison into a false one.
Move 37
Often reports the best artifact found within an online budget.
Both
Must publish denominator, field, and selection policy.
Move 42
Reports strict first move separately from bounded probe and diagnostic frontier.
08 / Experience
A campaign leaves a ledger
The training record is finite and inspectable: generations, promotions, processed rows, and summed generation-hours.
Move 37
Task-time trajectories disappear when only the winning artifact is shown.
Both
Need receipts to distinguish a measured run from a story about one.
Move 42
Publishes campaign totals with their limits and source hashes.
09 / Self-play
Algorithms become the games
Candidate programs play against a verifier. The retained frontier supplies experience for later openings without pretending every generated program is useful.
Move 37
Search traces optimize one newly presented task.
Both
Learn only from executions that the referee can reproduce.
Move 42
Aggregate scored games into training data for the next policy.
10 / Transition
From a surprising move to a changed opening
Move 37 symbolizes a discovery inside one game. Move 42 asks whether accumulated algorithm games can improve the move made before a new search begins.
Move 37
A singular task-time discovery.
Both
A result that matters only after the referee accepts it.
Move 42
A learned first move across future algorithm games.
11 / First move
The strict result is the center of the claim
The strict mode freezes one global executable artifact before task-time selection. It is less flattering than the frontier and more informative about the opening policy.
Move 37
Performance includes fresh search on the revealed task.
Both
Can be evaluated in a field of executable contenders.
Move 42
The frozen opening scores 1108.8 Elo and ranks 14 of 98 in the reported ledger.
12 / Evidence
The atlas is a lens, not the population
The 600 released cases are outcome-ranked to show mechanisms and failures. The complete matched cohort is the denominator for population statements.
Move 37
Showcase cases can illuminate how a search found a useful structure.
Both
Must disclose selection when examples are chosen after outcomes are known.
Move 42
Labels 600 atlas cases as non-representative beside the 1,649-case cohort.
13 / Limits
Reported does not mean universal
These are company-reported results on a historical evaluation record. They do not establish general intelligence, causal mechanism, or safety in untested domains.
Move 37
Online search may spend more compute and adapt more deeply to one task.
Both
Can fail on distribution shift, bad objectives, leakage, or weak referees.
Move 42
Depends on the coverage and integrity of earlier algorithm games.
14 / Every game
Alien code still has to survive an honest referee
‘Alien code’ is vision language: unfamiliar executable artifacts may become possible, but every artifact must parse, run, and earn its score under an honest referee.
Move 37
Search for a move no human would have proposed.
Both
Accept surprise only after executable verification.
Move 42
Learn openings that make new forms of algorithmic play more likely.