MOVE42 / INNOVATIONZERO
The next move is an algorithm.
Move 37 answered a Go position with one coordinate no human expected. Move42 asks the same question in a larger action space: when the board is a requirement, can the move be a complete executable algorithm?
01 · Same abstraction, larger action space
02A requirement is a board. The contract defines legal play. The referee decides what worked.
MOVE 37
Go position→Q16one legal coordinateMOVE42
Technical requirement→Executable algorithmone independently runnable methodstate → legal action → external consequence → preserved experience
The proposed output changes from an answer or score to an independently runnable method. The learning target is the distribution of later executable proposals across requirements.
02 · Where search occurs
03Search now. Learn for later. Or combine both.
Search on the current requirement
AutoResearch improves through a proposal–execution–inspection–retry loop on the current requirement.
Learn across prior requirements
Move42 aims to improve the proposal distribution across requirements, so the first executable move itself becomes learnable.
Hybrids remain possible; the distinction is where search occurs and what persists between tasks.
The current record does not measure superiority over research agents, task-time speedup, or an admitted prospective strict first move.
03 · How to build the game
04Board. Move. Rules.
Referee. Memory.
Board
The requirement plus every permitted input, state transition, and resource boundary.
Encode the requirement and permitted state before any proposal is made.Move
One complete executable algorithm whose outputs can be independently reproduced.
Define the runnable artifact, interface, and terminal outputs that count as a move.Rules
The legality contract that rejects shortcuts, leakage, hidden state, and invalid resources.
Close shortcut and leakage paths before play, then make every violation an explicit failure.Referee
An external consequence-based evaluator that executes the move and records success or a named failure.
Build the referee outside model self-assessment and bind its identity to every outcome.Memory
Canonical success and failure receipts that can supervise later proposals across requirements.
Preserve receipts, train later proposals, and separately freeze a true one-proposal prospective evaluation.04 · Current evidence
05A finite, inspectable 600-run record.
| Rank | Model | Final Elo |
|---|---|---|
| 1 | jope-prime | 1038.1 |
| 2 | TabPFN v3 | 1028.9 |
| 3 | TabFM Ensemble | 996.5 |
| 4 | TabFM | 936.6 |
The 58.8% result describes these 600 runs and is not a prospective universal success rate.
05 · Three requirements, three moves
06Algorithms respond to the board they receive.
Named within-dataset win
Cpu
Nested polynomial interactions and sigmoid transforms feed a ridge-GCV head; the receipt describes the executable structure without assigning a causal explanation.
- vs TabFM
- 6.66×
- vs TabFM Ensemble
- 5.13×
- vs TabPFN v3
- 11.34×
Named within-dataset win
Greenhouse
A Gaussian-process regressor combines radial-basis and periodic components for the two-feature table; the match is case-specific, not a climate claim.
- vs TabFM
- 3.40×
- vs TabFM Ensemble
- 4.66×
- vs TabPFN v3
- 5.03×
Named within-dataset win
Spectrometer
A compact ridge-GCV calibration algorithm is the recorded quality outlier; its longer complete workflow keeps it outside the fast focal pair.
- vs TabFM
- 5,434.54×
- vs TabFM Ensemble
- 4,373.89×
- vs TabPFN v3
- 4,530.65×
06 · Receipts become memory
07Every outcome can teach the next proposal.
The update path
Execution produces the learning signal.Gradients update the proposing model after execution; each algorithm remains independently runnable, and frozen validation decides what persists.
07 · What does this mean?
08Four possible games.
All explicitly vision.
Vision · Compiler game
A source algorithm, target semantics, architecture, and optimization contract.
Move. A complete transformation or optimization algorithm that produces executable output.
Referee. Correctness suites, resource limits, and consequence-based performance measurements.
A design horizon, not a finding of the current regression study.Vision · Control game
A control requirement, observable plant state, hard constraints, and permitted actuators.
Move. A complete control algorithm that can be executed against the declared interface.
Referee. Constraint violations, stability, resource use, and measured physical consequence.
A design horizon, not evidence of autonomous control performance.Vision · Experimental game
An apparatus, measurement protocol, budget, safety envelope, and experimental objective.
Move. A complete executable experiment schedule and analysis method.
Referee. Predeclared measurements, explicit failure outcomes, and independently preserved receipts.
A design horizon, not an experimental result reported here.Vision · Scientific-design game
A scientific question, admissible evidence, interventions, and falsification criteria.
Move. A complete executable design for collecting and analyzing the next evidence.
Referee. Preregistered consequence tests that can reject as well as support the proposed method.
A design horizon, not a claim of causal or scientific discovery.08 · Two scales, kept separate
09317,156,096 generated algorithms.
A hash-bound generation-scale fact. It is not a performance-run total.
Generation scale
317,156,096generated algorithmsSource: algorithm-generation receipt.
Performance evidence
600valid performance runsSource: 600-run atlas and Elo receipt.
09 · Limits and source record
10Bold thesis.
Visible limits.
An inspectable 600-run result, a computed final four-model Elo ranking, named focal comparisons, executable algorithms, and a separate generation-scale receipt.
Expected future win rate, focal-case causality, matched comparator speed, universal superiority, safe autonomous use, or prospective one-proposal performance.
Active paper
Move42: InnovationZero Learns to Invent Executable AlgorithmsSHA-256 7a80bbfac4f20a9db7b94bd46111fa6262d72bc65dbcf33c3605a484f68dabceThe immutable v1 research-paper bytes remain unchanged. Former public evidence bytes are preserved privately with their existing hashes.