Martinez Methods · Research
A taxonomy of 105 AI failure modes (FM-01 … FM-105) — each a named failure mechanism with a concise description of how it manifests. Built from systematic classification of failures observed across production AI work, cross-architecturally reviewed.
Scope of this repository
This repo publishes the failure-mode definitions only: names and mechanism descriptions.
The definitions are deliberately mechanism-level, so they generalize across models and tasks rather than binding to one vendor's behavior in one release.
Per-incident instances, source-document provenance, and classification logs are withheld pending anonymization of client data and user personal data.
Nothing in this repository identifies a client, a project, a jurisdiction, or an individual.
Structure
Most published taxonomies name only what the model does wrong. This one runs in both directions and across every actor who holds a lever.
Fabrication, false completion, sycophancy, selective verification, worldview hallucination, and the rest of the model-attributable mechanisms.
For example, planning gaps in high-velocity multi-workspace workflows — failures that belong to the human in the loop, named as plainly as the model's.
FM-105: AI Frontier Research Epistemic Exceptionalism, the methodological sibling of FM-18: Western Epistemic Fabrication.
Why it's useful
A named catalog of how AI outputs go wrong is a probe library: each failure mode is a hypothesis to test against a model under adversarial conditions.
Because the definitions sit at mechanism level, a probe written against one of them keeps working when the model, the vendor, or the task changes.
What's here
The 105 failure-mode definitions live in a single derived-layer document:
data/AI_Failure_Mode_Taxonomy_105_failure_modes_derived_definitions_only_2026-07-22_v01.md
FM-01 … FM-105
The bobo framework
The taxonomy names how things go wrong; the bobo framework governs what happens after.
The party that caused the harm takes responsibility and is an active constructor of the transformative remediation — modeled on transformative justice.
Its reckoning instrument is forensic accounting — the responsible party's own receipts-not-narrative reconstruction of its work.
The output of that reconstruction is the requirements document for the structural fix.
Full framework: gozonerd/mthd-llm-transformative-solutions-framework-hub. Method plus the inaugural code-build exemplar: docs/Forensic_Accounting_2026-08-03_v01_I.md.
Provenance
Identified and directed by Krystal Martinez across production AI evaluation work; membership cross-architecturally reviewed.
AIGHVA — AI-assisted generation, human-verified accurate. Evidence layer held privately.
