A framework you program, versus an engine you run.
NVIDIA's NeMo Guardrails is an open-source framework: you write guardrail and dialog logic directly in its Colang language. Both run on your own infrastructure, so the difference is what happens after the logic: versioning, backtesting, evidence, state, and who operates all of it.
What NeMo is
A free, code-first library with a five-stage rail model: input, dialog, retrieval, execution, output. You write the logic and you run it.
What Swiftward does differently
Here policy is a versioned artifact, not code you maintain: draft, candidate, frozen, archived. You test a change in shadow against live traffic, backtest it against your own recorded traffic to see which past decisions would flip, then roll a deployment back by naming the previous version.
Every decision lands in a tamper-evident, hash-chained record. State is kept per entity — counters, windowed counters, labels, metadata — across requests, not just inside one conversation.
With NeMo you write the logic in Colang and build and run all of that yourself.
Running both
Swiftward is the policy and evidence layer around a guardrail step, the same way it orchestrates any other detector. NeMo decides one conversation turn; Swiftward versions the policy that called it, holds the state across requests, and proves the decision to an auditor later, on infrastructure you still own. See the full field.