Back to Gap Map Product / UX / Safety & Guardrails

Roblox Sentinel

Roblox
open source / Overall score: 2.0

Roblox Sentinel is a Python library for detecting very rare kinds of harmful text, such as grooming, across a run of messages from one source rather than one message at a time. It compares each message with small sets of harmful and ordinary examples by embedding similarity, then combines the results for that source, by default with skewness, to surface candidates for human review. Roblox built it for its own chat moderation.

Sentinel installs from its repository and is not published to PyPI. The sentinel package there is an unrelated project.

Openness

5 high confidence
5.0
license
Apache-2.0(OSI, the repository LICENSE and the pyproject license field)
source
public(the whole library installs from Roblox/Sentinel with pip install .)
self-host
yes(a local Python library, with Dockerfiles for GPU and CPU-only images)
core features withheld
no — the tree at 66c6985 carries one root Apache-2.0 LICENSE and no ee/, enterprise/, commercial/ or proprietary/ path. Roblox released it as part of its open-source safety toolkit and sells no edition of it
service
none

Sentinel is released under Apache 2.0, and the published repository holds the whole library, which runs on your own machines. Roblox open-sourced it as part of its safety toolkit and offers no paid or hosted version.

Adoption

1 low confidence
1.0

GitHub stars are the only signal, and a star is not a use. Roblox's own production use of Sentinel on its chat service is not counted.

  • https://pypi.org/pypi/sentinel/json recorded 2026-10-19

    The PyPI project "sentinel" is version 1.0.0 by Eddie Antonio Santos, homepage github.com/eddieantonio/sentinel, summarized as "Create sentinel objects, akin to None, NotImplemented, Ellipsis". It is not Roblox's library.

  • https://ungh.cc/repos/Roblox/Sentinel recorded 2026-10-19

    Roblox/Sentinel: 118 stars and 27 forks, pushed 2026-08-18

Capability

3 low confidence
3.0

Like Hasher-Matcher-Actioner, Sentinel does one detection job and hands what it finds to human reviewers: HMA recognizes copies of media already flagged, and Sentinel finds patterns across a user's messages that no single message reveals. Sentinel is a library with no service or review screen, so the two sit level.

Verified 2026-10-19, except some scores, still awaiting re-confirmation