spb/autoscholar

1---23mode: computational4discipline: computer-science5title: "<Working title: a computational study>"6---78# Research brief — <working title> (computational)910## 1. Research objective and scope1112- **Question.** <e.g., Does method M improve metric K over baseline B on task T under budget C?>13- **Hypotheses.** <H1…; include the null.>14- **Compute scope.** CPU-only Apple Silicon node (16 cores, ≤ 32 GB RAM), ≤ <N> hours; no GPUs. Size experiments accordingly.1516## 2. Domain, mode and methodology1718- Discipline: <AI/ML/systems/algorithms/optimization>. Primary mode: computational.19- Methodology: <benchmark design, baselines, metrics, statistical tests across seeds>.2021## 3. Contribution and novelty2223- <Expected finding; the closest prior methods/benchmarks to compare with.>2425## 4. Literature and source rules2627- Indexes: arXiv + Semantic Scholar + Crossref. Target ≥ 12 verified sources; must-read: <ids>.2829## 5. Benchmark and data3031- Datasets/tasks: <public, licensed; exact versions/URLs; splits>. Leakage controls: <…>.32- Baselines: <at least two, including a strong simple one>. Implementation sources (cite).3334## 6. Procedure35361. Environment lock in `code/` (uv), deterministic seeds (≥ 5 seeds), fixed splits.372. Implement method and baselines; unit-test core functions.383. Run the pre-registered grid; log every run with commands, seeds, wall time, resources in `results/experiment-log.jsonl`.394. Aggregate deterministically into `results/metrics.json`; produce tables/figures with CIs.405. `code/reproduce.sh` regenerates everything.4142## 7. Baselines, robustness and falsification4344- Ablations: <components removed one at a time>. Sensitivity: <hyper-parameters, data size>. Report when the method loses.4546## 8. Reproducibility and artifacts4748- Pinned environment, seeds, logs, metrics, figure scripts, resource usage table.4950## 9. Manuscript5152- AutoScholar class; algorithm pseudo-code; results tables with mean ± CI over seeds; 8–16 pages.5354## 10. Review and revision5556- Reviewer focus: unfair baselines, seed variance, leakage, cherry-picked metrics, missing ablations.5758## 11. Limitations, ethics, legal5960- <Compute limits, dataset licenses, generalization.>6162## 12. Acceptance criteria6364- All runs logged and reproducible; ≥ 5 seeds; ≥ 2 baselines; ≥ 2 ablations; reported numbers traceable; ≥ 12 verified citations.65