Migrating from NeMo Guardrails to Authensor

NeMo Guardrails and Authensor approach AI safety from different angles. NeMo Guardrails uses Colang, a conversation modeling language, to define conversational rails. Authensor uses YAML policies to define action-level constraints. Migrating between them requires translating concepts, not just syntax.

Concept Mapping

| NeMo Guardrails | Authensor | |---|---| | Colang flows | YAML policy rules | | Input rails | Aegis input scanning + policy rules | | Output rails | Aegis output scanning + policy rules | | Dialog rails | Approval workflows | | Knowledge base | External (not managed by Authensor) | | Actions | Tool calls with policy evaluation |

Translating Input Rails

NeMo input rails define what the user can and cannot say. In Authensor, input filtering is handled by the Aegis content scanner:

NeMo Colang:

define user ask about competitor
  "What do you think about CompetitorX?"
  "Is CompetitorX better?"

define flow
  user ask about competitor
  bot refuse to answer

Authensor equivalent: configure an Aegis detection rule that identifies competitor mentions, then a policy rule that denies or redirects when the detection triggers.

Translating Output Rails

NeMo output rails filter model responses. In Authensor, output scanning is part of the Aegis content safety pipeline. Configure detection rules for the content patterns you want to block in outputs.

Translating Action Constraints

NeMo Guardrails constrains which actions an agent can take through Colang flow definitions. Authensor constrains actions through explicit YAML policy rules with tool name matching and parameter constraints.

This is where Authensor provides more granular control. NeMo Guardrails operates at the conversation flow level. Authensor operates at the individual tool call level, including parameter inspection.

Migration Steps

1. Inventory existing rails. Document every Colang flow, identifying what it protects against and how.

2. Map to Authensor concepts. For each rail, determine whether it maps to a policy rule, an Aegis detection rule, or an approval workflow.

3. Write the YAML policy. Translate each rail into a policy rule. Take advantage of Authensor's parameter-level constraints, which may allow you to express rules more precisely than Colang flows.

4. Configure Aegis rules. Set up content scanning rules for input and output filtering that previously relied on Colang pattern matching.

5. Run in parallel. Operate both systems simultaneously in evaluation mode. Compare their decisions on the same traffic. Resolve discrepancies before decommissioning NeMo Guardrails.

6. Switch over. Once decision parity is confirmed, disable NeMo Guardrails and enable Authensor enforcement.

The migration typically takes two to four weeks, depending on the number and complexity of existing Colang flows.