analyze-results

name	analyze-results
description	Analyze YOLO training runs — compares to baseline/best, checks per-class regression, analyzes training dynamics and tune convergence, writes actionable recommendations.

Analyze Results

Single-pass analysis run after each experiment or on-demand via /analyze.

Data

Read:

• Latest experiment's report + results.csv or tune_results.csv
• Current best metrics from experiments/summary.md journal entries
• experiments/dataset_profile.yaml
• training-plan.md (thresholds, goals, constraints)

Results.csv Lookup Order

1. experiments/exp_NNN_name/train/results.csv
2. experiments/exp_NNN_name/results.csv
3. Fall back to metrics.yaml — note that training dynamics analysis is unavailable

For tune runs, also check:

• experiments/exp_NNN_name/tune/tune_results.csv

Analysis

Comparison: Latest vs baseline AND current best (mAP50-95, per-class AP)

Per-class regression check: Flag any class exceeding training-plan.md threshold vs current best

Training dynamics (if results.csv available):

• Convergence speed: epochs to reach 90% of final mAP
• Overfitting onset: epoch where val loss diverges from train loss
• Loss plateau: epochs with <0.1% change in val loss
• LR effectiveness: correlation between LR schedule phase and metric improvement

Tune analysis (if tune_results.csv available):

• Tune convergence: are later iterations improving or stalled?
• Best params found vs defaults — which params moved the most?
• Top-N trial spread: how tight is the performance range?

Diminishing returns detection (when 3+ experiments show <0.5% change):

• Same-category stall (all augmentation, or all LR variations): that lever is exhausted, try a different category
• Cross-category stall (multiple lever types tried): model may be at ceiling for this dataset/architecture
• Recommendation differs per stall type — surface this explicitly

Recommendations: Ordered by expected impact. Each cites evidence from the analysis.

Output

Write experiments/analysis.md with findings and recommendations. Print top 3 findings + top 3 recommendations.

Findings also inform the journal entry's learning field and the next DIAGNOSE step in the reasoning loop.

Important

• The reasoning loop's DIAGNOSE step reads analysis.md from the previous experiment or last explicit /analyze. It does NOT read analysis of the experiment that just ran before it has been analyzed.
• Keep analysis concise — focus on what changed and what to try next, not restating all historical metrics.