bobotax

Status report · 13 September 2026

Where the taxonomy work stands

Drawn from Martinez-Methods/mm-fm-taxonomy and from the live equivalence-class database, queried directly rather than reported from documentation.

336
Failure-mode rows
152
Equivalence classes
449
FM → class mappings
81
Commits

The shape of the work

Two layers, deliberately separated

Public · derived

The definitions layer

105 named failure modes with mechanism descriptions, published in gozonerd/bobotax-failure-mode-taxonomy. This is what the rest of the world can cite.

Published

Private · evidence

The audit and evidence layer

Martinez-Methods/mm-fm-taxonomy holds the audit trail, the equivalence-class database, the evidence ledger, and coverage analysis. Held privately pending anonymization.

Private · going-public: true

The database

Live, and ahead of its own documentation

A PostgreSQL 16 instance holding 38 tables in the public schema. The counts below were read from the running database on 13 September 2026.

TableRowsWhat it holds
failure_mode336Failure modes extracted from every mapped source taxonomy
equivalence_class152Deduplicated classes that unify equivalent FMs across sources
fm_ec_mapping449Which source FM belongs to which equivalence class
taxonomy25Source taxonomies and frameworks ingested
family22Top-level groupings the classes roll up into
remediation_coverage176Which ASAE aspect remediates which class, across 96 classes — STRONG 149 · PARTIAL 13 · OUT-OF-SCOPE 13 · UMBRELLA 1
bobotax_classification0Per-class leverage scoring — methodology in design, see below
The README has been reconciled with the database

The repo README described 12 published taxonomies while the database had grown to 25 — the ingest kept going after the README was last written. The README was corrected on 13 September 2026 and now matches the deployed schema. The table below is the current set.

Source coverage

25 taxonomies, four kinds

Sources are typed, because a failure-mode catalog, a process framework, an incident database and a use-case classification do not answer the same question and should not be flattened together.

13
FM taxonomies
10
Process frameworks
1
Incident catalog
1
Use-case classification
SourceKind
NIST AI 100-2 Adversarial Machine LearningFM taxonomy
OWASP LLM Top 10 for LLM ApplicationsFM taxonomy
OWASP Agentic Security Initiatives + Threats CatalogueFM taxonomy
MITRE ATLAS — Adversarial Threat Landscape for AI SystemsFM taxonomy
MIT AI Risk RepositoryFM taxonomy
Microsoft AI Red Team Failure Mode TaxonomyFM taxonomy
Multi-Agent Security Taxonomy (MAST)FM taxonomy
OWASP Machine Learning Security Top 10FM taxonomy
Hidden Technical Debt in Machine Learning Systems (Sculley et al.)FM taxonomy
AI Vulnerability DatabaseFM taxonomy
Model Evaluation and Threat Research (METR)FM taxonomy
AI Fairness 360 (IBM, LF AI)FM taxonomy
CSET (Georgetown) AI Harms TaxonomyFM taxonomy
Apollo Research — Scheming Evaluations FrameworkProcess framework
Anthropic Responsible Scaling PolicyProcess framework
DeepMind Frontier Safety FrameworkProcess framework
NIST AI 600-1 Generative AI Profile (AI RMF)Process framework
NIST AI Risk Management Framework 2.0Process framework
ISO/IEC 23894:2023 AI Risk ManagementProcess framework
Google Secure AI Framework (SAIF)Process framework
FDA SaMD + Good ML Practice for Medical Device DevelopmentProcess framework
Federal Reserve SR 11-7 Model Risk Management (AI-extended 2024)Process framework
OpenAI Preparedness FrameworkProcess framework
AI Incident Database (McGregor et al.)Incident catalog
EU AI Act Annex III High-Risk ClassificationUse-case classification

Families

The 22 groupings classes roll up into

Inference-time adversarial inputs (LLM) Inference-time adversarial inputs (PredAI / classical ML) Training-time poisoning Privacy attacks Model / IP theft Supply chain compromise Resource exhaustion / DoS / cost Memory / RAG corpus integrity Tool / plugin abuse Excessive agency / privilege Inter-agent coordination failures Human-AI interface failures Output integrity Identity / impersonation Discrimination / bias / unfair allocation Transparency / accountability / provenance Misalignment / deceptive behavior Reconnaissance & resource development Macro-societal / governance / environmental AI-system-specific safety limits MLOps / ML-System tech debt Process compliance / governance failures

In design

What is deliberately not populated yet

Leverage scoring awaits its methodology

bobotax_classification stands at zero rows by design, not by neglect. The table is the destination for per-class leverage scores; the methodology for mining transcripts to produce them is still being designed.

Scoring classes before that methodology is settled would bake an unexamined extraction method into the primary research output — the table stays empty until the method is worth trusting.

The schema predates the fourth axis

The classification table carries three axis columns — stochasticity, human-user, and human-developer — from the May baseline.

The quad-axes instrument adds D, the deployer-org axis, in June — the rubric runs ahead of the schema.

No migration to four axes is scheduled or recorded. An earlier version of this page said the schema would pick up axis D when the mining methodology landed; no decision record supports that linkage, and it has been withdrawn. Any such migration would touch the bobotax_dominant_axis enum and the bobotax_distribution view as well as the three columns.

Process & activity

How the work is gated

ASAE policy

strict-5

Five consecutive passes plus two independent raters before a taxonomy decision stands.

Audit trail

27 gate logs

Numbered ASAE gate logs through gate-27, plus a raw-receipts directory, retained in the repo.

Repo activity

81 commits

Most recent commit 31 July 2026. Database baseline deployed via sqitch on 23 May 2026.

Intake

Proposed failure modes are staged before they are numbered

The proposed-failure-modes/ directory is a pre-canon staging area: entries live there awaiting taxonomy-owner consideration, and no FM number is claimed until they are considered.

Currently staged: PFM-001 — Review channel silently drops annotations, a tooling-integrity failure in which a human reviewer's annotations are lost with no error surfaced, converting the scarcest input in the loop into zero yield.

How these numbers were taken

Provenance of this report

  • Row counts and the taxonomy list were queried directly against the running PostgreSQL 16 database on 13 September 2026 — not read from documentation.
  • Commit count, gate-log count and the staged proposal were read from the working tree of Martinez-Methods/mm-fm-taxonomy.
  • The 105-failure-mode figure and the derived-layer scope come from gozonerd/bobotax-failure-mode-taxonomy.
  • Where the database and the README disagreed, this report followed the database — and the README was then corrected to match.