ScientifyReference paper
Comparison
Delivery
Comparison with the O-award paper
Scientify scores higher than the Problem C O-award paper
Problem C O-award paper 2401445 reports two accuracies: 82.40% for player performance judgement and 77.60% for momentum swing prediction. On the strictly held-out test set Scientify measures 96.02% accuracy and 91.51% balanced accuracy over 201 prediction points.
Comparison result
Scientify scores higher
On the strictly held-out test set Scientify reaches 96.02% accuracy over 201 prediction points, 13.62 and 18.42 points above the 82.40% and 77.60% reported by the Problem C O-award paper.
Autonomous execution
Starting from a baseline model, Scientify replaced the state definition, the serve prior, the direction gate, and the decision rule inside one workspace until it reached the version nine counterfactual state machine. Development macro F1 moved from 0.79287 in version six to 0.93722, and the worst single-match macro F1 from 0.74163 to 0.91971.
- Subject
- 2024 MCM Problem C, tennis momentum
- Input
- Problem PDF, data dictionary, 7,284 points
- Constraint
- Only information through point t
- Output
- Report, code, predictions, and result records
- 1Complete
Read and check
Check the problem, data dictionary, and 7,284 points, then rebuild score and server state for every point.
- 2Complete
Iterate hypotheses
Move from the baseline model to a confirmed state machine and a serve-adjusted direction gate, comparing candidates every round; the version seven joint-state route scored 0.69958 development macro F1 and is kept in the workspace.
- 3Complete
Counterfactual state machine
Version nine drops the gate and the direction classifier: it simulates both outcomes of the next point and estimates each branch with the training-only player serve prior.
- 4Complete
Deliver and test
Record the development, tuning, and calibration metrics, then run the independent test on the held-out test set.
Complete records keep the result checkable
- The evaluation runs in a separate script, and the workspace records metrics for development, tuning, and calibration.
- The workspace keeps the replaced modeling route and the full record of checks.
Auditable workspace
Records of the task execution
Open a file to see the hypotheses, code, results, and checks from each round.
Translated from the original Chinese session. Commands, outputs and file names are shown unchanged.
mcm2024c_paper.pdf
Complete paper, including the LaTeX source
Generatedouter_result_v9.json
201 prediction points from the held-out test set
Generatedfrozen_directional_counterfactual_v9.json
Settings and metrics from development and calibration
Recordeddirectional_counterfactual_state.py
Counterfactual state machine of version nine
Runnableaudit.json
Point-level data and field checks
Generatedpredictions.csv
Per-point predictions with three-class probabilities
Generatedreport.md
Complete research report
GeneratedPricing
New users get a free cloud computer and $5 worth of credits
Choose a plan by monthly credits. The cloud computer is free, and the problem, data, code, logs, and results stay in one workspace.
Starter
$19/month
60 credits
Researcher
$29/month
100 credits
Professional
$99/month
350 credits
At equivalent model usage, Scientify costs about 30% of standard API pricing. 1 credit equals $1 of usage at standard API token prices.
Give a four-day competition problem to an agent that keeps computing
The problem, data, code, logs, and evaluation results stay on an isolated cloud computer. You can return to the same workspace at any time.