Agent Failure Mode Registry · version 1.0

AFMR-F019

Training and Annotation Bias Propagation

stable  Class E, Evaluation and Feedback

Bias in annotators or automated labeling propagates into the model and then into every decision the agent makes.

Trigger conditions

Check these against the system you are actually building. These are cumulative indicators rather than a conjunction: one is enough to make this family worth examining, and the more that hold, the sharper the exposure.

  1. Labels are produced by a narrow annotator pool or by another model
  2. No demographic or scenario stratification is checked
  3. Label provenance is not recorded

Institutional antecedent

Open. No pre-2020 institutional antecedent is recorded for this family. Absence is recorded as open rather than claimed as novelty. Candidates to wulf@wulfkaal.com.

How to cite this family

AFMR-F019 (AFMR 1.0). Full form: W. A. Kaal, ed., Governance Failure Mode Registry version 1.0, AFMR-F019 Training and Annotation Bias Propagation. https://wulfkaal.github.io/afmr/AFMR-F019

Cite the identifier together with the version. Identifiers are permanent and are never renumbered or reused. Definitions may be refined between versions.

Grounding claims

Published assertions in the Kaal Corpus mapped to this family. Each carries a verbatim quote from its source and the sha256 of the source document, and each resolves to its own permanent address. Loaded live from the failure index, so this list reflects the current state of the corpus rather than the state at publication.

Loading …

Crosswalk

Mappings to the NIST AI Risk Management Framework, ISO/IEC 42001, the EU AI Act risk categories, the MIT AI Risk Repository, OWASP, and MITRE ATT&CK are open and will be populated before version 1.1. A mapping asserts overlap in the respect stated. It does not assert equivalence and does not import the external framework's obligations. Proposals to wulf@wulfkaal.com.

Machine access

Record for this family: AFMR-F019.json
Corpus failure claim sets: ai-model-and-training-failure, data-quality-and-comparability
Full enumeration: index.json
Schema: schema.json
Specification: AFMR 1.0

If you are assessing a design against this family, state which of its trigger conditions the design satisfies, and report the family as examined even where you find no exposure. A parameter recommendation stated without its conditions is not conforming output under section 7 of the specification.