RIGHTS AND ETHICS
- Source: 277 PDFs of The Journal of Finance (Wiley / American Finance Association), volumes 72–75, obtained on 2026-11-03 through the Université du Québec en Outaouais library subscription (proxybiblio.uqo.ca) for research use by a subscribing faculty member. Access is personal and institutional; the PDFs are not redistributed, not committed to git, and never modified.
- Derived artefacts kept private (gitignored): structured full texts (
data/derived/parsed/), fixtures containing excerpts, raw engine outputs. Committed artefacts contain only metadata, counts, structural annotations, paraphrased summaries, short excerpts (≤400 characters) inside the retrieval index units, and aggregate profiles. - External API exposure: annotation calls send bounded excerpts (abstract, introduction paragraphs, conclusion paragraphs, first paragraph of each section, table notes ≤350 chars, captions ≤300 chars; total ≤ 60,000 characters per document,
config/rights.yaml) to the Anthropic API under the account's standard data terms; no full article is ever sent in one call; every call is logged. No fine-tuning is performed. - Originality: the engine is instructed never to reuse sentences; every output is screened for 8-gram overlap with the corpus; results are reported (H5). The system learns conventions, not wording, and does not claim Journal of Finance endorsement.
- Evaluation ethics: tasks use verified material (published abstracts/structures, the author's own WP29); no fabricated finance findings are presented as real; outputs are proposals that a human must verify (FLAGS). Human raters were not available; LLM-judge results are labelled as such.
- Research integrity statement for WP45: all counts, tables and figures are produced by scripts from recorded data; API costs and model versions are logged; the preregistration was frozen before the runs; the full retrieval index and annotations can be regenerated from the PDFs with
make.