Skip to content

aiopshwang/data-analysis-ml-agent-skills / skills/diagnosing-ml-failures

diagnosing-ml-failures

Verified

Isolate the root cause of ML performance drops, inconsistent evaluations, prediction errors, and training-serving mismatches across data, labels, splits, pipelines, models, metrics, and runtime behavior. Use when investigating a reproducible failure or regression, not routine model selection or general performance validation.

FileKindSize
SKILL.mdskill md4.5 KB
agents/openai.yamlconfig236 B
assets/failure-investigation-log.mdmarkdown2.8 KB
references/failure-isolation.mdmarkdown4.1 KB

assets/failure-investigation-log.md

View on GitHub
# ML Failure Investigation Log

## Incident identity

- **Incident / issue ID:**
- **Owner:**
- **Opened / updated:**
- **Impact and affected decision:**
- **Current status:** investigating / mitigated / root cause verified / unresolved
- **First known bad:**
- **Last known good:**

## Reproducible symptom

- **Expected behavior:**
- **Observed behavior:**
- **Affected population and grain:**
- **Unaffected control:**
- **Reproduction entry point:**
- **Frequency:**
- **Input, code, config, artifact, and environment IDs:**
- **Smallest failing case or paired comparison:**

## Change inventory

| Dimension | Last known good | First known bad | Confirmed difference? |
|---|---|---|---:|
| Source snapshot / query |  |  |  |
| Code revision |  |  |  |
| Configuration |  |  |  |
| Model artifact |  |  |  |
| Dependencies / runtime |  |  |  |
| Metric / threshold |  |  |  |

## Boundary trace

| Layer | Invariant or record field | Expected / good | Observed / bad | Evidence ID | First divergence? |
|---|---|---|---|---|---:|
| Source and semantics |  |  |  |  |  |
| Ground truth |  |  |  |  |  |
| Split and sampling |  |  |  |  |  |
| Feature pipeline |  |  |  |  |  |
| Model artifact |  |  |  |  |  |
| Metric / decision logic |  |  |  |  |  |
| Runtime / delivery |  |  |  |  |  |

## Hypothesis ledger

| ID | Hypothesis | Prediction if true | Discriminating test | Result / evidence | Status | Reopen condition |
|---|---|---|---|---|---|---|
| H-01 |  |  |  |  | open |  |

Allowed status: `open`, `supported`, `rejected`, `blocked`.

## Experiment log

| Time | Experiment | Single intended change | Control | Result | Interpretation | Artifact / log ID |
|---|---|---|---|---|---|---|
|  |  |  |  |  |  |  |

## Human ground-truth review

- **Why review was required:**
- **Prespecified sampling method:**
- **Reviewer independence and blinding:**
- **Original label / reviewer decision / adjudicated label locations:**
- **Disagreement and ambiguity:**
- **Population-level inference supported by the sample:**

## Root-cause closure

- **First broken layer:**
- **Causal factor:**
- **Mechanism connecting cause to symptom:**
- **Controlled intervention:**
- **Recovery evidence:**
- **Unaffected control evidence:**
- **Scope matched to the original failure:** yes / no - explain
- **Root cause, mitigation only, or unresolved:**

## Fix and regression coverage

| Original failure scope | Fix | Regression check | Negative control | Result | Owner |
|---|---|---|---|---|---|
|  |  |  |  |  |  |

## Rejected hypotheses worth preserving

Summarize strong negative findings that future investigators should not repeat without new evidence.

## Remaining risk and next action

- **Unresolved hypotheses:**
- **Known blind spots:**
- **Monitoring signal:**
- **Next discriminating test, owner, and due date:**
writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deploymentStars: 322kUpdated Claude CodeCodexVerifiedInstall
xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy.Stars: 119kUpdated Claude CodeCodexVerifiedInstall
childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release.Stars: 118kUpdated Claude CodeCodexVerifiedInstall