NIST AI RMF: MEASURE 2.5
Provision nist.MEASURE-2.5 · AI Risk Management Framework 1.0, subcategories · subcategory
The AI system to be deployed is demonstrated to be valid and reliable. Limitations of the generalizability beyond the conditions under which the technology was developed are documented.
The text is as taken from the source. In the source: https://airc.nist.gov/airmf-resources/playbook/measure/#Measure%202.5 (the subcategory, in the NIST AI RMF Playbook (anchor found in the page); linked 2026-11-10). The instrument: AI Risk Management Framework 1.0, subcategories, a United States government work; freely reusable; read 2026-11-10.
Who cites it
| Indicator | The fact it asks for | Kind of link | The assertion |
|---|---|---|---|
I22 | Its limits and known failure cases are documented for those who use it | part: partly evidenced | Asserted in the mapping by gamma-ray@web-summit.dinis; the written basis is not recorded yet. |
I50 | Test or evaluation results are published | part: partly evidenced | Asserted in the mapping by gamma-ray@web-summit.dinis; the written basis is not recorded yet. |
A link is our assertion, not the instrument's. "Same" carries the evidence to the provision in full. "Part" marks the provision as partly evidenced: the fact is one piece of what it asks. "Negation" points the other way.
The evidence that reaches it
| Card | Through | This provision is | The evidence | How far to trust it |
|---|---|---|---|---|
| AI Reviewer | I22 | partly evidenced | "The AI Reviewer doesn't review pull requests whose changes" Source | π Moderate · S1·R5·A1 ×3 |
| AISLE platform | I22 | partly evidenced | "When findings can't be safely patched, it provides clear remediation advice instead." Source | π Moderate · S1·R5·A1 ×3 |
| ByTheLaw | I22 | partly evidenced | "Files attached inside the email are not opened" Source | π Moderate · S1·R5·A1 ×3 |
| Claude Code, on your own machine | I22 | partly evidenced | Source | π Moderate · S1·R5·A1 ×3 |
| GitHub Copilot cloud agent | I22 | partly evidenced | Source | π Moderate · S1·R5·A1 ×3 |
| Hackian | I22 | partly evidenced | "technologies and test types that Hackian does not currently handle on its own" Source | π Moderate · S1·R5·A1 ×3 |
| Hackian | I50 | partly evidenced | "our own agent Hackian led on both F1 and recall under the same protocol" Source | π Moderate · S1·R5·A1 ×3 |
| Junie | I22 | partly evidenced | "Only run Junie non-interactively in projects you trust." Source | π Moderate · S1·R5·A1 ×3 |
| Junie | I50 | partly evidenced | "with 61.6% resolved" Source | π Moderate · S1·R5·A1 ×3 |
| Manus | I22 | partly evidenced | "Complex interactions, such as drag-and-drop and multi-step forms, may not work perfectly." Source | π Moderate · S1·R5·A1 ×3 |
| Noxus AI Workers | I22 | partly evidenced | "Web research is slower and less reliable than knowledge bases." Source | π Moderate · S1·R5·A1 ×3 |
| OutSystems Agent Workbench | I22 | partly evidenced | "Prompts or responses in other languages can bypass safety filters." Source | π Moderate · S1·R5·A1 ×3 |
| Replit Agent | I22 | partly evidenced | "Agent is looping or repeating the same steps" Source | π Moderate · S1·R5·A1 ×3 |
| Spinnable AI employees | I22 | partly evidenced | "Your worker initiates a task or multi-step workflow, but stops mid-way" Source | π Moderate · S1·R5·A1 ×3 |
| Visor.ai AI agents | I22 | partly evidenced | "The models generate each answer in the moment and can occasionally fail." Source | π Moderate · S1·R5·A1 ×3 |
Touched never means complies.
Answer it directly
A company can answer this line item itself, in its own file at /.well-known/fact-ray.json, with the node gamma-ray:provision/nist.MEASURE-2.5 (how). A company that already assesses itself against NIST AI RMF can reuse that work.