Granite Guardian vs Guardrails AI

Granite Guardian

One ungated Apache 2.0 model judges harm, jailbreaks, RAG groundedness, function-call errors and custom criteria, with signed weights and published evaluation code. It is trained and tested on English only, each call checks one criterion, the 4.1 prompt format differs from 3.x, and IBM's watsonx.ai lists only the deprecated 3.0 model.

Guardrails AI

Validators have configurable actions for failed checks. Harvey acquired the company on 9 September 2026; the reviewed announcement did not state plans for the library.

Granite Guardian

1. Append the <guardian> block as the final user message, with the mode line, ### Criteria: and ### Scoring Schema:. Copy the strings from the model card, because no package builds them
2. Use the no-think instruction for gating and parse <score>. Think mode writes a reasoning trace first, and the card's examples allow up to 2,048 output tokens
3. Treat yes as the criterion being met, which for built-in criteria means the risk is present. Treat a missing <score> tag as a failed check
4. Pass retrieved text through documents= and tool schemas through available_tools= in apply_chat_template, not inside the message text
5. Under Ollama, set num_ctx in the request options. IBM's docs say the default context is short and long requests are truncated

Guardrails AI

1. Pin guardrails-ai==0.11.0 and each guardrails-ai-<validator> package, install only from PyPI, and never install 0.10.1
2. Import validators from guardrails_ai.<name>, not guardrails.hub, and don't run guardrails hub install
3. Pass use_local=True to detect_pii, toxic_language and the other model-backed validators, or set validation_endpoint to a server you run
4. Set enable_metrics to false in ~/.guardrailsrc if you don't want usage metrics sent
5. Avoid building on reask and RAIL. The open 1.0.0 issues plan to remove both