docs/license_audit.md

# License audit

This audit records what was found **on disk**, in the project's own benchmark
preparation directories, before any redistribution decision was made. Sub-agent
A performs the formal, authoritative audit in parallel; this file records the
independent check that drove this staging tree's contents.

## Decision rule applied

Default to caution. Do **not** stage upstream problem text, gold answers,
reference solutions, or answer options for any benchmark unless on-disk evidence
shows a license that clearly permits redistribution *and* the release plan needs
the text. Where in doubt, omit the text and keep stable identifiers.

**Outcome: no problem text is staged for any benchmark, including the two whose
upstream licenses (MIT) would have permitted it.** The decision was made
release-wide rather than per-benchmark, because the flagship table
`data/matched_labels.csv` interleaves all four benchmarks in one file and a
mixed-license text column would be hard for a downstream user to respect
correctly. Stable IDs plus `docs/reconstruction.md` give the same scientific
value without the licensing hazard.

## Evidence found on disk

| Benchmark | Evidence file | License stated | Redistribution permitted by that license? | Decision here |
|---|---|---|---|---|
| Omni-MATH-2 | `dataset_release/hf_release_v2/LICENSES/README.md`, `registry/licenses.parquet` | Apache-2.0 | yes, with attribution | **text withheld** (release-wide rule) |
| JEEBench | `data/source_benchmarks/jeebench/LICENSE`, `data/science-benchmarks/manifest.json`, per-row `license: MIT` field | MIT, Copyright (c) 2023 DAIR Group, IIT Delhi | yes, with notice | **text withheld** (release-wide rule) |
| SciBench | `data/source_benchmarks/scibench/LICENSE`, `data/science-benchmarks/manifest.json`, per-row `license: MIT` field | MIT, Copyright (c) 2023 Xiaoxuan Wang | yes, with notice | **text withheld** (release-wide rule) |
| LAB-Bench | `data/lab-bench-public/README.md`, `data/lab-bench-public/manifest.json`, `registry/licenses.parquet` | CC-BY-SA-4.0 with an upstream do-not-train request and a canary string | share-alike; redistribution would force CC-BY-SA on the redistributed content and carries the canary obligation | **text withheld** — the most restrictive of the four, and the binding constraint |
| MaScQA | `data/source_benchmarks/mascqa/LICENSE`, `data/science-benchmarks/manifest.json` | CC-BY-NC-SA-4.0 | NonCommercial; the project's own manifest marks it `internal_only_not_compatible_with_cc_by_sa_4_0_release` | **entire benchmark excluded** |

### Notes per benchmark

**LAB-Bench.** `data/lab-bench-public/README.md` records upstream revision
`5c77cec648430f30611808808861eb86f81d5eaa` of `futurehouse/lab-bench` and states
that the upstream canary field is deliberately kept out of the agent-facing
prompt. The project's own release code (`scripts/dataset_release/
prepare_hf_labbench_protocol_dataset.py`) carries the note "Check upstream
LAB-Bench terms before public redistribution." That unresolved check is itself a
reason to withhold. LAB-Bench was treated as the binding constraint for the
release-wide rule.

**MaScQA.** Excluded for two independent reasons. First, it is not in the paper:
there is no matched Gemma run, so the setting
`mascqa__text_only__gpt_oss_120b` (642 rows) has no paired counterpart and is not
one of the ten camera-ready settings. Second, CC-BY-NC-SA-4.0 is a NonCommercial
license, which the project's own `data/science-benchmarks/manifest.json` marks as
`internal_only`. The 642-row file exists in the source tree and was deliberately
not staged.

**Omni-MATH-2.** The Apache-2.0 finding comes from the project's own registry
(`registry/licenses.parquet`, source
`https://huggingface.co/datasets/martheballon/Omni-MATH-2`), not from a LICENSE
file sitting next to the data. No `LICENSE` file was found under
`data/omni-math-2-filtered/`. That is a weaker form of evidence than the MIT
LICENSE files present for JEEBench and SciBench, and is a second reason the
release-wide withhold rule is the safe call.

## Model output terms

`openai/gpt-oss-120b` has Apache-2.0 weights. `google/gemma-4-31B-it` is governed
by the Gemma Terms of Use. This release contains **no raw generated text** from
either model — only derived correctness flags, integer confidence values, and
token/call accounting — so the exposure is limited to derived measurements. The
Gemma Terms are cited in `NOTICE` regardless.

## What this audit does not establish

- It does not re-verify the upstream licenses at their sources. Every license
  above was read from a file in the project tree, not fetched from the upstream
  repository or dataset card at audit time. Confirm before public upload.
- It says nothing about whether the derived measurements themselves are
  copyrightable, which is a question for counsel, not for this file.