DATA LICENCE — the SCROLLBACK derived dataset

The derived dataset published as `dump.csv` and `dump.sqlite` is licensed under
the Creative Commons Attribution 4.0 International Licence (CC-BY-4.0).

    https://creativecommons.org/licenses/by/4.0/

You may share and adapt it, including commercially, provided you give
attribution.

Attribute as: SCROLLBACK (mirc.tech), CC-BY-4.0.


UPSTREAM ATTRIBUTION

The dataset is derived from open-licensed sources. Attribution to them is a
condition of their licences and travels with any redistribution:

  - Epoch AI, Benchmarking Hub (CC-BY-4.0) — the sole source of every MEASUREMENT
    in `benchmark_result` (2343 of 2343 rows; `measurer` is "Epoch AI" on all of
    them).
    https://epoch.ai/data/ai-benchmarking-dashboard

    The `benchmark` and `benchmark_version` tables are NOT Epoch data. Their
    descriptive values — publisher, score_kind, human_baseline, cost_kind,
    homepage_url, saturation_state and the two licence columns — are this
    repository's own curated registry, written by hand and versioned with the
    code. Attribution here would be over-attribution, which is the safe
    direction but still not what happened.

    `benchmark.license_spdx` / `benchmark.redistribution` describe the terms
    EPOCH DISTRIBUTES ITS MEASUREMENTS UNDER. They say nothing about the
    benchmark itself: several of these benchmarks are published by third
    parties under their own, more restrictive terms.

  - LiteLLM model catalogue (MIT, excluding the repository's `enterprise/`
    directory, which this project does not read) —
    `model_prices_and_context_window.json` and its git history.
    https://github.com/BerriAI/litellm

  - models.dev catalogue (MIT) — `api.json`, and `providers/*/models/*.toml`
    with its git history.
    https://models.dev  ·  https://github.com/anomalyco/models.dev

  - Official vendor deprecation pages (developers.openai.com,
    platform.claude.com), read for the AVAILABILITY and EVENT layers: they back
    44 `availability` rows and the 44 `event` rows derived from them. Only dates
    and retirement kinds are reproduced; no page is republished. Those rows carry
    the page they came from in `availability.source_url` / `event.source_url`.

    NO `price_fact` ROW CITES A VENDOR PAGE. The price layer is the two
    catalogues above and nothing else (measured: 1749 models.dev and 1147
    LiteLLM observations behind 6702 price rows). `price_fact` has no
    `source_url` column at all — its evidence handle is `source_ids`, which
    names archive-ledger rows, not URLs.


WHERE THE URLs IN THIS DATASET COME FROM

  `benchmark_result.evidence_url` is UPSTREAM'S OWN CITATION STRING, republished
  verbatim from Epoch's CC-BY-4.0 corpus and never fetched by this archive. It is
  present on 1245 of 2343 rows and absent on the other 1098. It points wherever
  Epoch pointed — 336 rows to arcprize.org, 318 to htihle.github.io, and a
  handful carry a leaderboard's title rather than a URL.


WHAT IS NOT PART OF THIS DATASET

  - THE RAW ARCHIVE IS NOT PART OF THIS DATASET. The archived source bodies
    under `data/raw/` are stored for measurement and reproducibility and are not
    redistributed. `price_fact.source_ids` and
    `benchmark_result.source_snapshot_id` name rows of that archive; they are
    evidence handles, not resolvable joins.

  - The WeirdML CSV (htihle.github.io) carries no licence statement upstream. It
    is archived locally and is NEVER redistributed. It produces no derived row:
    every quality-axis row in this dataset comes from Epoch AI's CC-BY-4.0
    corpus, including the WeirdML measurements Epoch mirrors under that licence.
    The 318 htihle.github.io strings noted above are Epoch's citation of the
    leaderboard PAGE, not of that CSV, and no cell of it appears here.

  - Source code is not covered here. It is MIT, in the separate LICENSE file.


TWO THINGS TO KNOW BEFORE YOU USE THE NUMBERS

  - `price_fact.suspect = 1` marks rows this archive does NOT stand behind. They
    are published rather than deleted because the archive records what the
    catalogues actually said; the `suspect` table gives the rule that fired and
    the reason. Filter them before computing anything.

  - A NULL `valid_to` or `available_to` means the period was open AT
    `dump_meta.archive_as_of` — not that it is open today. A model with no
    `availability` row is UNKNOWN, not available.