# PDF-Forensics Genuine: changelog

Each version, newest first, with what changed. Counts are values of this version's `summary.json`.

## 1.3.1 (2026-11-03)

- Removed the link list: the dataset now holds only files our detector scored and confirmed CLEAN; the link records were never scored by release 0.18.0.
- `links.jsonl`, `links.csv` and `fetch_links.py` are no longer part of the release; `files.jsonl` with its CSV mirror `files.csv`, the datasheet card, `CITATION.cff` and this changelog remain.
- The same 82,955 files in the same PDF parts as 1.3.0 and 1.2.0: the parts are byte-identical, so an earlier download of the parts stays valid.

## 1.3.0 (2026-11-02)

- The same 82,955 files as 1.2.0, in the same PDF parts: the parts are byte-identical to 1.2.0's, so a 1.2.0 download stays valid.
- Adds link records to genuine originals whose terms do not let Scam.ai redistribute their bytes, listed for completeness: not redistributed by us and not screened by our detector (removed in 1.3.1).
- Adds `files.csv` and `links.csv`, CSV mirrors of the JSON Lines lists, and `fetch_links.py`, which downloads linked documents politely, checks each sha256 and falls back to the Internet Archive.
- Adds `CITATION.cff` and this changelog. The dataset card now follows the datasheet sections (motivation, composition, collection, preprocessing and selection, uses and out-of-scope uses, distribution and licences, maintenance and takedown) and describes every field.

## 1.2.0 (2026-11-01)

- Keeps only the 82,955 files of the genuine-only cohort that Trustable's detection engine 0.18.0 read CLEAN; 12,290 files it flagged, could not conclude on, or that Scam.ai holds out for its own evaluation were withheld.
- No link records (none was scanned by the engine). Adds DOWNLOADS.md with integrity checks.

## 1.1.0 (2026-10-24)

- Genuine-only: keeps real-source, unmodified originals (95,245 files) and their link records; edited, forged, synthetic, specimen, blank-form and reference documents are left out.

## 1.0.0 (2026-10-22)

- First release, as PDF-Forensics-1M: genuine documents together with Scam.ai's edited, forged and synthetic documents, as files and as link records.