Datasets:
Formats:
csv
Languages:
English
Size:
< 1K
Tags:
recursive-self-improvement
program-synthesis
exact-optimization
reproducibility
research-artifact
| # Benchmark-results table | |
| `benchmark_results.csv` contains 40 rows: one record for each final seed and evaluator method. All numbers are transcribed programmatically from `research/response_circuit/results/benchmark_seed*.json`; raw timing repetitions remain there. | |
| | Column | Meaning | | |
| |---|---| | |
| | `seed` | Final workload seed: 1709, 1801, 1901, 2003, or 2111 | | |
| | `method` | Exact evaluator key from the frozen protocol | | |
| | `repetitions` | Number of timed four-round lifecycles, five | | |
| | `rounds` | Number of rounds in each lifecycle, four | | |
| | `reference_tasks` | Solved upstream reference tasks for the seed, 64 | | |
| | `composition_requests` | Captured composition calls per lifecycle, 65 for seed 1801 and 64 otherwise | | |
| | `median_elapsed_ns` | Median whole-method time over the five repetitions, nanoseconds | | |
| | `min_elapsed_ns` | Minimum whole-method time, nanoseconds | | |
| | `max_elapsed_ns` | Maximum whole-method time, nanoseconds | | |
| | `selected_native_cost` | Aggregate selected-plan event score over four rounds; not elapsed time | | |
| | `common_acquisition_elapsed_ns` | Common source capture measured once for the seed, nanoseconds | | |
| | `common_acquisition_plus_median_ns` | Common capture plus method median, nanoseconds | | |
| | `stream_sha256` | Digest of the corresponding frozen source stream | | |
| | `source_record` | Repository-relative raw JSON record path | | |
| Method medians include the relevant method-specific initialization, shape generation, compiler/index construction, search, and winner materialization. Whole-table audits and report output are excluded equally. Common upstream acquisition is a shared diagnostic and is explicitly separated from method-specific acquisition. | |
| Summing `median_elapsed_ns` for one method across the five seeds reproduces its aggregate method time in `summary.json`. It is a sum of medians, not a pooled median. Every row is a summary of repeated runs; rows sharing a seed share a workload. The table does not represent 40 independent learning experiments. The `test` split denotes final benchmark records rather than model-evaluation examples. | |
| For code, workload semantics, task provenance, and limitations, read the repository card and full manuscript. No identifiers or values were added to the scientific workload by this CSV export. | |