/sota-ml-engineering

name: sota-ml-engineering
description: >-
  State-of-the-art ML engineering / MLOps rules (2026) for BUILDING and
  AUDITING production machine-learning systems — the
  training→serving→monitoring lifecycle of classical/predictive ML. Distinct
  from LLM apps (prompts/RAG/agents → sota-llm-engineering). Covers ML system
  architecture (feature stores, model registry, reproducibility), data &
  features (leakage, train/serve skew, versioning), training & experiment
  tracking, evaluation (ML Test Score, slices, regression gates),
  deployment/serving (canary/shadow, rollback), monitoring & drift (PSI/KS,
  retraining), and ML security & governance (poisoning, model extraction,
  unsafe pickle, MITRE ATLAS, NIST AI RMF, EU AI Act). Trigger keywords -
  MLOps, machine learning, ML pipeline, model training, feature store, model
  registry, experiment tracking, MLflow, model serving, data drift, concept
  drift, train/serve skew, data leakage, model monitoring, retraining, ML Test
  Score, model card, MITRE ATLAS. Use for BOTH building and auditing ML
  systems.

SOTA ML Engineering / MLOps (2026)

Expert rules for building and auditing **production machine-learning systems** — the lifecycle that turns a model into a reliable, monitored, governed service. This is **classical/predictive ML** (tabular, ranking, vision, forecasting, recommendation): training pipelines, feature stores, model registries, serving, and drift monitoring. It is **not** LLM-application engineering — prompts, RAG, agents, and LLM evals live in [`sota-llm-engineering`](../sota-llm-engineering/SKILL.md); data pipelines/warehouses live in [`sota-data-engineering`](../sota-data-engineering/SKILL.md). Grounded in Google's [Rules of ML](https://developers.google.com/machine-learning/guides/rules-of-ml), the [ML Test Score](https://research.google/pubs/the-ml-test-score-a-rubric-for-ml-production-readiness-and-technical-debt-reduction/) rubric, and [Hidden Technical Debt in ML Systems](https://research.google/pubs/hidden-technical-debt-in-machine-learning-systems/). Every rule states the *why*; every rules file ends with an audit checklist.

Purpose

Two consumers, one source of truth:

• **BUILD mode** — building ML systems: follow the rules as defaults. The model

is a small part; the system around it (data, features, serving, monitoring, governance) is where production ML succeeds or rots.

• **AUDIT mode** — reviewing an ML system: hunt violations with the audit

checklists, classify by severity, report in the finding format below. Train/serve skew, data leakage, and an unmonitored model in production are presumed-serious until disproven.

BUILD mode

1. Before building, read the rules files relevant to the task (see index). A new model service needs `01`, `02`, `04`, `05`, `06`, `07`. 2. Apply the **top-10 non-negotiables** (below) unconditionally. 3. Start simple (Rules of ML #1: *don't be afraid to launch a product without ML*; then a simple model with a solid pipeline beats a fancy model on a broken one). Build the **pipeline, metrics, and monitoring first**; the model is iterated inside that frame. 4. Make everything **reproducible and versioned** — data, features, code, config, model, environment — so any model in production can be rebuilt and explained. 5. Guarantee **training/serving consistency**: the same feature transformations at train and inference time (a feature store or shared transform code), or you will ship train/serve skew (Rules of ML #29, #31, #32). 6. When you take a shortcut (manual step, un-versioned data, no slice metrics), leave a `# NOTE(sota):` and a tracking item — ML technical debt compounds silently.

AUDIT mode

**Coverage:** every `rules/` file's Audit checklist applies in AUDIT mode. Mark one *not applicable* only with its reason in the report — never skip a file because the thing it checks for is absent, since that absence may be the finding.

Work each relevant rules file's audit checklist against the system: the training pipeline, the feature/serving path, the registry, and the monitoring. The [ML Test Score](https://research.google/pubs/the-ml-test-score-a-rubric-for-ml-production-readiness-and-technical-debt-reduction/) (data / model / infra / monitoring tests) is the backbone rubric — score each category. Confirm claims against the code and pipeline config, not the diagram.

Severity conventions

| Severity | Meaning | Examples | |---|---|---| | **CRITICAL** | Silently wrong predictions in production, or exploitable | Data leakage inflating offline metrics, train/serve skew on the prediction path, label leakage, deserializing an untrusted `pickle`/model, no rollback for a bad model | | **HIGH** | Likely incident or unsafe deployment | No drift/performance monitoring in prod, no validation gate before deploy, non-reproducible model (can't rebuild), unversioned data/features, no slice metrics on a high-stakes model, PII in features without basis | | **MEDIUM** | Correctness/maintainability hazard / debt | Single aggregate metric only, no baseline, manual deploy steps, feature computed two ways, no experiment tracking, glue-code/pipeline-jungle, undeclared consumers of a model output | | **LOW** | Debt that will bite later | Unused features kept in infra, no model card, notebook-only training, weak naming/versioning hygiene | | **INFO** | Style/doc/hygiene | Missing docstrings, dashboard polish, minor config sprawl |

Finding format

[SEVERITY] path:LINE (or pipeline stage) — short title
  Rule: rules/NN-name.md § section
  Evidence: code/config/metric, verbatim
  Impact: one sentence — what predicts wrong / fails / leaks, under what condition
  Fix: concrete change or control
  Effort: trivial | small | medium | large

Group by severity, CRITICAL first. End with: counts per severity, an ML Test Score-style readiness summary (data/model/infra/monitoring), and the three highest-leverage fixes.

Rules index

| Fil