MOVE 37 → MOVE 42
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?
A requirement is a board. The contract defines legal play. The referee decides what worked.
- 600 valid runs · 353 wins · 247 losses
- 58.8% against TabFM Ensemble
- Not a prospective universal success rate
01 / 9Same abstraction
01 · Same abstraction
Same abstraction. A larger action space.
Move 37 and Move42 share a formal pattern: a state receives a legal action, an external consequence determines what happened, and preserved experience changes later proposals.
A requirement is a board. The contract defines legal play. The referee decides what worked.
state → 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.
What this supports
The analogy is structural, not a claim of identical search mechanics or equivalent empirical validation. Go supplies fixed semantics; each technical requirement defines a new game.
02 · Output type
The breakthrough is a different output type.
The move is not merely an answer, prediction, benchmark score, or research-loop suggestion. It is an independently runnable method that must complete the declared task.
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.
- Not merely a better prediction.
- Not merely a benchmark score.
- Not merely a task-time research loop.
- An independently runnable method.
What this supports
The precise thesis is a change in output type and learning target: propose an executable algorithm, observe its consequence, and improve later executable proposals.
03 · Where search occurs
Search now. Learn for later. Or combine both.
AutoResearch improves through proposal, execution, inspection, and retry on the current requirement. Move42 aims to improve the proposal distribution across requirements.
Where search occurs
Search now. Learn for later. Or combine both.
AutoResearch improves through a proposal–execution–inspection–retry loop on the current requirement.
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.
Evidence boundary. The current record does not measure superiority over research agents, task-time speedup, or an admitted prospective strict first move.
What this supports
Hybrids remain possible. This distinction describes where search occurs and what persists between tasks; the current record does not measure superiority, task-time speedup, or a prospective first move.
04 · Five-part blueprint
Five parts turn a requirement into a game.
The board defines permitted state, the move is executable, rules close shortcuts, the referee measures external consequence, and memory preserves successes and failures.
- 01
Board
The requirement plus every permitted input, state transition, and resource boundary.
Encode the requirement and permitted state before any proposal is made. - 02
Move
One complete executable algorithm whose outputs can be independently reproduced.
Define the runnable artifact, interface, and terminal outputs that count as a move. - 03
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. - 04
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. - 05
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.
What this supports
A result is meaningful only when legality and consequence are outside model self-assessment and every terminal path leaves a canonical receipt.
05 · Current evidence
What if AI were a game?
Across 600 valid runs, jope-prime recorded 353 wins and 247 losses against TabFM Ensemble, a 58.8% result. Four-model Elo puts those comparisons on one common scale.
| Rank | Model | Final Elo |
|---|---|---|
| 1 | jope-prime | 1038.1 |
| 2 | TabPFN v3 | 1028.9 |
| 3 | TabFM Ensemble | 996.5 |
| 4 | TabFM | 936.6 |
- Valid runs
- 600
- Wins
- 353
- Losses
- 247
- Result vs TabFM Ensemble
- 58.8%
What this supports
All six pairwise lower-RMSE comparisons per run feed the final ranking. The 58.8% result describes these 600 runs and is not a prospective universal success rate.
06 · Algorithmic moves
Different requirements call for different moves.
CPU, greenhouse, and spectrometer each receive a different complete algorithm under the same consequence-based referee, with exact programs and clocks preserved in the atlas.
- Cpu6.66× vs TabFM · 5.13× vs TabFM Ensemble · 11.34× vs TabPFN v3
- Greenhouse3.40× vs TabFM · 4.66× vs TabFM Ensemble · 5.03× vs TabPFN v3
- Spectrometer5,434.54× vs TabFM · 4,373.89× vs TabFM Ensemble · 4,530.65× vs TabPFN v3
Inspect the named within-dataset comparisons and verbatim programs.
What this supports
These focal cases are post-hoc descriptive records, not causal explanations or evidence of a future success frequency. Comparator wall-clock was not receipted.
07 · Memory
Every outcome can teach the next proposal.
The external referee records executable successes, losses, and named failures. Those receipts become stored supervision for a later challenger across requirements.
A ten-state learning loop preserves success and failure receipts, trains a later challenger after execution, and conditionally promotes it after frozen validation.
Static overview · all ten states
- 01Freeze parent
Lock the retained policy and evaluator identities.
Output · parent + evaluator receipt - 02Select positions
Choose recorded dataset contracts for this generation.
Output · dataset contract set - 03Propose programs
Sample complete structured artifacts from the frozen parent.
Output · program candidates - 04Execute referee
Parse, compile, fit, predict, and score; failures leave at named gates.
Output · success + failure receipts - 05Bind receipts
Bind artifact, dataset, split, evaluator, resources, and outcome.
Output · hash-bound receipts - 06Store experience
Retain measured scores, ranks, runtimes, stability, and failures.
Output · immutable experience - 07Train challenger
Apply gradients only to the model from stored supervision.
Output · candidate checkpoint - 08Validate
Run the frozen validation contract against parent and challenger.
Output · comparison receipt - 09Choose outcome
Reject a regression or accept an eligible improvement.
Output · retain / promote - 10Start later
A promoted challenger may become the frozen parent of a later generation.
Output · new retained parent
Execution boundaryMeasurements cross as stored data. Gradients do not cross executed code.
Preserve receipts, train later proposals, and separately freeze a true one-proposal prospective evaluation.
What this supports
Gradients update the proposing model only after execution; they do not pass through arbitrary executed code. Frozen validation decides whether a later challenger is retained.
08 · What does this mean?
The game pattern can reach beyond tables.
Compiler, control, experimental, and scientific-design games illustrate how requirements could become executable moves under domain-specific rules and referees.
Vision · Compiler game
- Board
- A source program, 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
- Board
- 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
- Board
- 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
- Board
- 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.
What this supports
Every implication on this page is labeled vision. None is a finding of the current regression study, a safety claim, or evidence of performance in those domains.
09 · Generation scale and limits
Generation scale is not performance evidence.
317,156,096 generated algorithms describe the search-generation scale. The public performance result remains exactly 600 valid runs.
Generation scale
317,156,096 generated algorithmsThis generation-scale receipt is separate from the performance evidence.
Performance evidence
600 valid runsThe 58.8% result and final four-model Elo ranking use only these runs.
Boundary
Generation is not evaluationNo additional run, game, dataset, file, finalist, cohort, or campaign total is a public performance claim.
What this supports
Generated algorithms are proposals, not additional performance runs. The 600-run result remains historical and does not establish a prospective universal success rate.
Keep reading