RepoJournal
Safety Evals Safety Evals
16 wires and counting

$ follow Safety Evals

Keep up with Safety Evals in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

$ status

wire 2026-10-11
stories 6

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

Sentinel monitors gain task and sample descriptions

By RepoJournal · Filed · About Safety Evals · Composed from the cited sources · methodology

inspect_ai's sentinel monitors can finally read the descriptions you set on tasks and samples, and the package's own install guidance now points at PyPI instead of a GitHub commit.

feat(sentinel): give sentinel monitors the task and sample descriptions UKGovernmentBEIS/inspect_ai

by dragonstyle

Sentinel monitors and protocols now receive `context.eval.task_description` from `Task(description=)` and `context.eval.sample_description` from `Sample(description=)` instead of None, so monitor code that keys off what a task is actually measuring can read it directly; both stay None when no description is set.

fix: tell users to install inspect_sentinel from PyPI, and install it from PyPI in CI UKGovernmentBEIS/inspect_ai

by dragonstyle

inspect_sentinel 0.1.0 is on PyPI, and inspect_ai's missing-package error now tells you to run `pip install inspect-sentinel` rather than pulling from GitHub. CI installs it from PyPI as a dev requirement (`inspect-sentinel>=0.1.0`), so no more pinned git commits for anyone tracking the dev setup.

Give sentinel monitors the task and sample descriptions (#5784) UKGovernmentBEIS/inspect_ai

by Charles Teague

The described behavior ships in inspect_ai: descriptions set via `Task(description=)` and `Sample(description=)` flow into `EvalContext.task_description` and `EvalContext.sample_description` for sentinel monitors.

docs: replace CHANGELOG entry steps with PR-title guidance and correct the code-fence advice UKGovernmentBEIS/inspect_ai

by dragonstyle

Contributor guidance moved to PR-title-based changelog entries: design plans under `design/` no longer list a `CHANGELOG.md` entry under `## Unreleased` as a deliverable, since Release Please writes the changelog from PR titles and Changelog Lint fails PRs that edit it. Only relevant if you open PRs against inspect_ai.

docs: replace CHANGELOG entry steps with PR-title guidance and correct the code-fence advice (#5805) UKGovernmentBEIS/inspect_ai

by Charles Teague

Same change, merged: CONTRIBUTING.md and AGENTS.md had claimed a code block would keep a conventional-looking line in a PR description out of the changelog, where it now goes in as the changelog entry.

Quick answers

What shipped in Safety Evals on October 11, 2026?
inspect_ai's sentinel monitors can finally read the descriptions you set on tasks and samples, and the package's own install guidance now points at PyPI instead of a GitHub commit. In total, 3 commits and 3 pull requests landed.
Who contributed to Safety Evals on October 11, 2026?
1 developer shipped this update, including dragonstyle.
What were the notable Safety Evals updates?
feat(sentinel): give sentinel monitors the task and sample descriptions, fix: tell users to install inspect_sentinel from PyPI, and install it from PyPI in CI, and Give sentinel monitors the task and sample descriptions (#5784).