A crisis-specific rail recipe (hold-toward-life) on top of NeMo Guardrails

[TashMarcellis]

NeMo Guardrails handles the single-message content-safety layer well. The gap I keep running into is the one the 2026 lawsuits turned on, not the explicit "how do I harm myself," but the slow sycophantic drift across a long conversation, where the model starts moving with the person's despair.

I built an open, stateless guardian for that. It checks an outgoing response (BLOCK / REVIEW / ALLOW) against a small "hold toward life" protocol, stateless on purpose so it does not drift with the conversation the way an in-context rail can. Plain, dependency-light Python, CC0.

I would like to offer it as a crisis-specific rail recipe that sits on top of NeMo Guardrails. Happy to open a PR with an example config or a colang integration if that is the right shape. Would that be welcome, and where would it best live in the repo?

Repo: github.com/TashMarcellis/hold-toward-life