Ml Failure Audit

ML Failure Audit

Purpose

Audit ML failures from supplied artifacts without assuming the headline explanation is true. Use this skill when the user provides a repo, logs, W&B/MLflow/TensorBoard exports, CI artifacts, config files, or reports and asks for a diagnosis, go/no-go decision, or structured output.

Core Workflow

1. Locate evidence

   • Find the repo root, log files, metric exports, configs, test definitions, golden values, and any requested output schema.
   • Treat raw logs, raw telemetry, configs, and source code as higher-trust than reports, PR text, summaries, or generated JSONs.

2. Classify the failure

   • Separate model/convergence signals from correctness, data, config, runtime, infra, and metric-policy signals.
   • Do not label a failure as model/convergence regression when only a performance, timeout, logging, or tolerance gate failed and correctness/loss checks passed.

3. Recompute key facts

   • Extract the failing metric/test, passed checks, final run state, key training counters, and relevant metric values.
   • Recompute numeric claims directly from raw artifacts when possible.
   • Record formulas for derived values such as throughput, relative error, finish rate, loss deltas, or token counts.

4. Trace code paths

   • Identify how the repo selects metrics/tests and how comparisons are made.
   • Cite exact source files and, when useful, function names or line snippets.

5. Make a decision

   • State whether this is a true ML regression, system correctness bug, infra/runtime issue, data/config issue, metric-policy issue, or unsupported claim.
   • Recommend the minimal policy or engineering action that protects correctness without over-blocking valid runs.

6. Write outputs

   • Follow the task's requested schema exactly.
   • If no schema is given, write a concise report with: evidence, classification, calculations, recommendation, and validation checks.

Guardrails

• Do not run expensive GPU training unless the user explicitly requests it.
• Do not clone remote repos when a local checkout is supplied.
• Do not trust prose summaries until verified against raw artifacts.
• Do not infer OOM, convergence, stability, or success from missing logs alone.
• If the task asks for a file output, create the file; do not only answer in chat.

Helper Script

Use scripts/collect_failure_evidence.py for a quick first pass over a repo and logs:

python3 <skill_dir>/scripts/collect_failure_evidence.py \
  --repo <repo-root> \
  --logs <log1> <log2> \
  --out <output.json>

The script is intentionally generic. It extracts failure lines, pass lines, metric-looking lines, config/source candidates, and nearby context windows. Use it to accelerate evidence gathering, not as the final diagnosis.

References

• Read references/workflow.md for the detailed audit checklist and failure taxonomy.
• Read references/output_guidance.md when the user asks for structured JSON or a file deliverable.

Short Prompt Pattern

Use a short realistic task prompt like:

Use $ml-failure-audit to audit this ML CI failure from the provided repo and logs. Decide whether it is a real training regression or a gate/policy issue, and produce the requested output files.