# ConStellaration leaderboard, scored again

Our evaluations of leaderboard submissions, and the leaderboard's own
recorded results without the submitted boundaries.

| File                    | What it holds                                                                                                                             |
| ----------------------- | ----------------------------------------------------------------------------------------------------------------------------------------- |
| `data/entries.parquet`  | One row per single-objective submission and per MHD-stable point: submitter, date, recorded objectives, feasibility, score                |
| `data/rescored.parquet` | One row per boundary we evaluated: environment, our metrics and feasibility beside the recorded ones, time, memory                        |
| `build_data.py`         | Writes both files, and the evaluate notebook's `data/leading-front.json`, from the results dataset and the output of `scripts/rescore.py` |
| `scripts/rescore.py`    | Evaluates leaderboard submissions again; reads them from Hugging Face when it runs                                                        |

- **Source:** https://huggingface.co/datasets/proxima-fusion/constellaration-bench-results,
  last modified 25 October 2026. The results dataset states no licence, so
  the submitted boundaries are not copied here.
- **Licence of the code:** `scripts/rescore.py` adapts
  `compute_flux_compression_in_regions_of_bad_curvature` from
  constellaration v0.2.6, MIT licence, Copyright (c) 2025-present Proxima
  Fusion GmbH; the notice is in `UPSTREAM-LICENSE`.
- **Computed with:** constellaration 0.2.6 on x86-64. Rows with `run`
  "leaderboard environment" or "flux compression only" ran on Python 3.10
  with booz-xform 0.0.9 built from source, as the leaderboard does; rows
  with `run` "kernel image" ran in ScienceHub's kernel image on Python 3.13
  with booz-xform 0.1.0. The `python` and `booz_xform` columns say which.