AI Tools
Safety Shield is our in-house content-moderation checkpoint. It inspects prompts and model responses inside your network and returns a policy verdict for each one.
Six capabilities the checkpoint carries into a Private Vector delivery. Every one of them ships in the software today.
Your applications submit a prompt or a model response, and the checkpoint returns a clear policy verdict scored by Google's ShieldGemma 2B model. The same checkpoint screens prompts on their way in and responses on their way out.
Verdicts come back under the policy your team defines, and each verdict carries a verifiable attribution to the exact policy revision that produced it. Without a policy of your own, the checkpoint moderates under Google's published default safety policy.
Every verdict is recorded in an audit trail, and the choice to store the moderated text beside it belongs to your team. The record is written for every verdict, so the trail stays complete even when the content is left out.
When the moderated text is kept out of the record, its integrity still rides on salted one-way hashes your team can verify later. Every record carries the fingerprints and the salt it needs, and verifying them takes no model access.
Each audit record is sealed with AES-GCM encryption and organized by month for retention. Sealed bytes are all the audit folder ever receives, and a record reopens only under the right key.
The service runs on one pinned GPU on your own hardware, and prompts and responses stay inside your network during moderation. Every model parameter is verified resident on that GPU before serving begins.
Safety Shield writes its moderation record on infrastructure inside your network. The rows below record what every verdict writes down, how content exposure is controlled, how the records are sealed and organized, and how the build behind them stays secret-free.
The moderation record that stays inside your network.
Any change to the policy text produces a new verifiable attribution, and identical policy text always yields the identical attribution.
A team holding the recorded salt and the original text can recompute each recorded integrity hash without any model access.
Records land on disk as opaque sealed bytes that reopen only under the right key.
The build definition pins its base image by full digest and carries no directive that binds a token, a weight file, a salt, or a key.
A shipped scanner checks the built image, the committed files, and the full git history for tokens, model weights, and key material, and reports a clean verdict only when none is found.
The service pins itself to a single GPU and refuses to run on any other device.
Safety Shield is one part of a Private Vector delivery, not the delivery itself. The pages below carry the rest of the story: the stage where moderation outcomes become measured improvement, and the capability that deploys AI infrastructure inside your network.