How to submit a metric

  1. What reviewers look for
  2. What to have ready
  3. How to submit on GitHub
  4. What happens next

Step 2 of 4 What to have ready

7 answers are required. Everything else is optional — include what you can and leave out what does not apply. Gather these before you start and the rest takes ten minutes.

Built from the submission format in the repository, last read 2026-09-14. Read the full field reference, which is what the automated check enforces. If anything here disagrees with the repository, the repository is correct.

Required

Required fields, what each one means, and the values allowed
What we ask What it means
Format version
schema_version 		The schema version this file follows. Add each newly released 1.x version to this list. Copy it from the template as-is. Nothing to decide.

Choose from these, spelled exactly as shown:

  • 1.0
What is your metric called?
name 		Formal name of the quantitative/qualitative measure or measurement methodology.
How is it defined and applied?
applied_definition 		Precise mathematical or textual description of the metric or measurement method.
Who is submitting it?
submitter_organizations 		The entity or institution responsible for the submission.

One entry per line.
Contact email
contact_email 		Professional point of contact for verification. This address is published in the submission file.
Citations
references 		Citations of peer-reviewed literature or NIST resources.

One entry per line.
Where can someone get an implementation?
implementation_resources 		Links to software, data, or other online resources implementing the approach.

One entry per line.

Optional, but reviewers value them

A submission with only the required answers is valid and welcome. These are where reviewers get most of what they need, so fill in what you can.

Optional fields, what each one means, and the values allowed
What we ask What it means
Which AI RMF characteristics does it help measure?
ai_rmf_characteristics 		Target AI RMF trustworthy characteristic(s).

One entry per line.

Choose from these, spelled exactly as shown:

  • Valid & Reliable
  • Safe
  • Secure & Resilient
  • Accountable & Transparent
  • Explainable & Interpretable
  • Privacy-Enhanced
  • Fair
  • Accurate & Bias-Managed
What kind of evaluation is it?
primary_tevv_application 		Methodology, tool type, or mathematical approach.

One entry per line.

Choose from these, spelled exactly as shown:

  • Adversarial Robustness Evaluation
  • Red Teaming Evaluation Method
  • Automated Language Model Red Teaming Toolkit
  • Facial Recognition Benchmark Evaluation
  • Human-Centered Evaluation
  • Interpretability & Explanation Evaluation
  • Language Model Benchmark Evaluation
  • Object Recognition Benchmark Evaluation
  • Privacy & Security Evaluation
  • Tabular Data Benchmark Evaluation
Where in the AI lifecycle is it applied?
ai_lifecycle_stages 		Phase(s) in the AI lifecycle where the metric or method is applied, one stage per entry.

One entry per line.

Choose from these, spelled exactly as shown:

  • Plan and Design
  • Collect & Process Data
  • Build
  • Use
  • Deploy
  • Operate & Monitor
What does it measure?
object_of_measurement 		The entity evaluated.

One entry per line.

Choose from these, spelled exactly as shown:

  • agent
  • task
  • data
  • model
  • system
  • user
Is it tied to particular model architectures?
model_specificity 		Either 'model agnostic', or 'model specific' plus which architectures.
Is it tied to particular domains?
domain_specificity 		Either 'domain agnostic', or 'domain specific' plus which domain(s).
When is this metric most useful?
usage_details 		Details or scenarios where the metric or method provides the most value.
When is it misleading, or easy to game?
known_failure_modes 		Scenarios where the metric or method may be misleading or easily gamed. The field reviewers most often find empty, and the one they most often ask about.
On what kind of data?
modality 		Primary input data category.

One entry per line.

Choose from these, spelled exactly as shown:

  • text
  • vision
  • video
  • audio
  • structured data
  • physical
  • others
Common variants
common_variants 		Related versions of the metric.

One entry per line.
Metrics to measure alongside it
related_metrics 		Complementary metrics to be measured in tandem.

One entry per line.
What does it need to run?
computational_requirements 		Hardware or environment needs.
Licence or usage rights
usage_rights 		Licensing or permissions.

Take the template with you

The blank template lists every field with a comment explaining it. Copy it, fill it in, and the next screen shows you where it goes.

Open the blank template and full field reference