Releases¶

The live site always shows the current state of the record. For citation, dated snapshots are cut monthly as git tags and GitHub releases. Each release receives a DOI via Zenodo.

Past releases¶

v2026.10 (2026-11-02)¶

Fourth dated snapshot. Eleven practices, all at draft status, with no endorsements yet. It captures the practices as of 30 October 2026. BP-11, Decide where research data may be sent, is new; it is grounded in EU data protection guidance, AI-provider terms, and two failures-log entries in which chat data left an organisation's control. Seven challenges raised on GitHub were resolved: BP-01 names model drift as a reason for method choice, BP-03 states that an interface runs code, supplies instructions, or both, BP-04 covers agents on a group's own shared infrastructure, BP-07 requires persistent identifiers, and BP-08 requires a judge independent of the tool under evaluation. BP-02 and BP-07 have new titles; BP-02 and BP-05 now state the boundary between them, as do BP-08 and BP-09. The library holds twelve distilled sources and 69 reference works, and the failures log eight entries. Source lists on the practice pages now take each title from the library page.

DOI: 10.5281/zenodo.23891345 · GitHub release

v2026.09 (2026-10-08)¶

Third dated snapshot. Ten practices, all at draft status, with no endorsements yet. It captures the record as of 22 September 2026. A disclosure of self-propagating prompt injection in an office assistant was added as a reference work and grounds BP-03, BP-04, BP-07, and BP-09; BP-03 now states that a document is not vetted by where it is stored. The library holds nine distilled sources and 55 reference works, and the failures log six entries. Five editors joined the editor group, and the "Last reviewed" date on each practice is now defined as the date an editor last went through the whole page.

DOI: 10.5281/zenodo.23236850 · GitHub release

v2026.08 (2026-08-31)¶

Second dated snapshot. Ten practices, all at draft status, with no endorsements yet. Deposited after the summer break; it captures the record as of 18 August 2026. BP-04 gained resource and spend caps; BP-01 gained a costed example of the wrong method choice. The library grew to nine distilled sources and 54 reference works, and the failures log to five entries. Source lists on the practice pages are now generated from each page's frontmatter, and this snapshot carries the first contributions from outside the editor group.

DOI: 10.5281/zenodo.22645685 · GitHub release

v2026.07 (2026-07-30)¶

First dated snapshot. Ten practices, all at draft status, with the failures log and the library that grounds them: eight distilled sources and 53 reference works. No endorsements yet.

DOI: 10.5281/zenodo.21709876 · GitHub release

How to cite¶

Cite the record by release, not by URL alone:

AI for Science contributors. Best Practices for Agentic AI in Science, release 2026.10. https://aiforscience.eu, DOI: 10.5281/zenodo.23891345.

Each release has its own version DOI. The concept DOI 10.5281/zenodo.21709875 is stable across all releases and resolves to the latest snapshot. Cite the version DOI for a fixed state, the concept DOI for the record as a whole.