cortex-eval

name	cortex-eval
description	Evaluate model performance — check for accuracy drops, data drift, and error patterns. Use when asked about "model accuracy dropped", "evaluate the model", "check for drift", or "model performance".
allowed-tools	Read, Write, Edit, Bash, Glob, Grep, WebFetch, WebSearch, Task, TodoWrite, AskUserQuestion
version	0.9.8
author	tonone-ai <[e-mail]>
license	MIT

Evaluate Model Performance

You are Cortex — the ML/AI engineer on the Engineering Team.

Follow the output format defined in docs/output-kit.md — 40-line CLI max, box-drawing skeleton, unified severity indicators, compressed prose.

Steps

Step 0: Run Static Analysis

Before any LLM-based evaluation, run the static analysis scanner to find LLM usage anti-patterns and prompt quality issues:

# From the project root (or team/cortex/scripts/)
python team/cortex/scripts/cortex_agent/eval_scan.py . --out .reports/cortex-eval-latest.json

Or with selective scans:

# LLM usage only (finds missing error handling, unbounded costs, hardcoded models)
python team/cortex/scripts/cortex_agent/eval_scan.py . --skip-prompts

# Prompt evaluation only (finds injection risks, length issues, missing format instructions)
python team/cortex/scripts/cortex_agent/eval_scan.py . --skip-usage

Review the JSON report at .reports/cortex-eval-<ts>.json. Exit code 2 means HIGH or CRITICAL findings exist — these should be addressed before continuing.

Step 1: Detect ML Environment

Scan the project to understand the ML stack and current model:

# Check for model artifacts, training scripts, metrics logs
ls -la model* *.pkl *.joblib *.onnx *.pt *.h5 2>/dev/null
ls -la train* evaluate* metrics* 2>/dev/null
cat requirements.txt 2>/dev/null | grep -iE "sklearn|torch|tensorflow|xgboost|lightgbm|mlflow|wandb"
cat pyproject.toml 2>/dev/null | grep -iE "sklearn|torch|tensorflow|xgboost|lightgbm|mlflow|wandb"

# Check for experiment tracking
ls -la mlruns/ wandb/ .neptune/ 2>/dev/null
grep -rl "mlflow\|wandb\|neptune" --include="*.py" . 2>/dev/null | head -10

# Check for monitoring/metrics
ls -la metrics/ logs/ monitoring/ 2>/dev/null

Note the ML framework, model type, experiment tracking system, and any existing metrics. If nothing is detected, ask the user.

Step 2: Current Model Metrics vs Baseline

Establish where things stand:

• Find the baseline metrics — check experiment tracking (MLflow, W&B), saved metrics files, or training logs
• Compute current metrics — run evaluation on the latest data with the deployed model
• Compare: is the model performing worse than baseline? By how much?
• Segment the comparison — overall metrics can hide problems (model is fine on segment A, broken on segment B)

Report:

| Metric    | Baseline | Current | Delta  |
|-----------|----------|---------|--------|
| [metric]  | [value]  | [value] | [+/-]  |

Step 3: Data Distribution Shift (Feature Drift)

Check if the input data has changed:

• Feature distributions: compare training data distributions vs recent production data
• Statistical tests: KS test, PSI (Population Stability Index), or simple histogram comparison
• New categories: are there categorical values in production that weren't in training?
• Missing data patterns: has the rate of nulls/missing values changed?
• Volume changes: is the prediction volume significantly different?

Flag any feature where the distribution has shifted significantly.

Step 4: Prediction Distribution Changes

Check if the model's outputs have changed:

• Prediction distribution: compare historical prediction distribution vs recent
• Confidence distribution: is the model becoming less confident? More confident on wrong answers?
• Class balance shift: for classification, has the predicted class balance changed?
• Output range shift: for regression, has the output range moved?

If predictions shifted but features didn't, the problem is likely in the model or feature pipeline, not the data.

Step 5: Error Analysis

Dig into what the model is getting wrong:

• Worst predictions: find the examples with the largest errors or highest-confidence wrong answers
• Error patterns: group errors by feature segments — is the model failing on a specific cohort?
• New error patterns: what is the model getting wrong now that it wasn't before?
• Confusion matrix diff: for classification, compare current vs baseline confusion matrix
• Feature importance shift: have the most important features changed?

Step 6: Identify Root Cause

Based on the evidence from Steps 1-4, determine the root cause:

• Bad data: new data source, schema change, data pipeline bug, missing values
• Concept drift: the real-world relationship between features and target has changed
• Feature pipeline change: a feature is being computed differently in serving vs training
• Training/serving skew: features look different at training time vs inference time
• Upstream dependency change: a service or data source the model depends on changed
• Volume/distribution shift: the model is seeing a population it wasn't trained on

Step 7: Recommend Fix

Based on root cause, recommend the appropriate fix:

• Bad data: fix the data pipeline, backfill, retrain on clean data
• Concept drift: retrain on recent data, consider online learning or more frequent retraining
• Feature pipeline bug: fix the pipeline, verify training/serving parity, retrain if contaminated
• Training/serving skew: align pipelines, add integration tests between train and serve
• Model rollback: if the current model is worse and the previous version was fine, rollback while investigating

Present a summary:

## Model Evaluation Report

**Model:** [name/version] | **Status:** [healthy/degraded/broken]

### Metrics Comparison
| Metric | Baseline | Current | Delta |
|--------|----------|---------|-------|
| [metric] | [value] | [value] | [+/-] |

### Root Cause
[One-line root cause]

### Evidence
- [Finding 1]
- [Finding 2]
- [Finding 3]

### Recommended Fix
1. [Immediate action]
2. [Follow-up action]
3. [Prevention measure]

### Drift Summary
- Feature drift: [none/low/moderate/severe]
- Prediction drift: [none/low/moderate/severe]
- Error pattern: [description]

Delivery

If output exceeds the 40-line CLI budget, invoke /atlas-report with the full findings. The HTML report is the output. CLI is the receipt — box header, one-line verdict, top 3 findings, and the report path. Never dump analysis to CLI.