Granite Guardian

PROFILE OVERVIEW

Where this model fits

Granite Guardian goes beyond a single harmful-content label. It is particularly relevant when the question is whether an answer follows supplied evidence or a tool call matches its context. Specify the criterion: different kinds of errors require different inputs and tests.

OrganizationIBM

Scope of this profileGranite Guardian family; 4.1–8B documented examples

Output / roleCriterion-specific judgment; format depends on version

Evidence statusDocumentation reviewed; not independently benchmarked

Several judging tasks

IBM documents built-in risk criteria covering harmful content, retrieval-grounded answers and tool-call hallucinations, alongside custom criteria. The repository organizes examples by model version.

Source: IBM Granite · Granite Guardian documentation and examples

Version-specific scoring

The 4.1 examples supply a judging criterion and request a yes/no score, with thinking and non-thinking modes. Follow the intended template and score meaning for the selected criterion.

Source: IBM Granite · Granite Guardian documentation and examples

Grounded does not mean the source is true

Our interpretation: an answer may faithfully reproduce an outdated document. Evaluate source selection separately from whether the answer is supported by that document, and avoid presenting a judge’s explanation as proof.

A PRACTICAL EXAMPLE · HYPOTHETICAL

Put the distinction to work

An HR assistant cites a leave-policy document but adds an unsupported exception. A groundedness check can help identify the addition. A separate document-owner review is still needed to establish that the policy itself is current.

Where the evidence stops

Do not apply the 4.1 prompt template or custom-criteria claims to every older Guardian checkpoint. A predicted judgment is not a deterministic business-rule check.

FOLLOW THE EVIDENCE

For more reading

The sources behind this page, with a reason to open each one. Practical examples and recommendations are our editorial interpretation.

Granite Guardian documentation and examplesIBM Granite · Model documentation
Versioned examples for risk, groundedness, tool-call checks and custom judging criteria.
FActScore: evaluating factual precisionMin and colleagues · Research paper · 2023
Evaluates support for individual factual claims rather than treating a long answer as entirely right or wrong.

Sources reviewed 13 September 2026. Product documentation can change. How we use evidence

For more reading

The sources behind this page, with a reason to open each one. Practical examples and recommendations are our editorial interpretation.

Granite Guardian documentation and examplesIBM Granite · Model documentation

Versioned examples for risk, groundedness, tool-call checks and custom judging criteria.

FActScore: evaluating factual precisionMin and colleagues · Research paper · 2023

Evaluates support for individual factual claims rather than treating a long answer as entirely right or wrong.

Sources reviewed 13 September 2026. Product documentation can change. How we use evidence