Deterministic Self-Improvement Loops

The Problem

The project agent will NOT reliably execute self-improvement loops on its own. Proven failure modes:

1. Skips steps — reports deficits without acting on them
2. Halts early — checks the math, says "insufficient," and waits for human direction
3. Hallucinated results — invents thresholds and metrics instead of running real training
4. No retry — tries once, fails, reports failure as final answer

This was proven in the Classifier Lab session (2026-03-29 to 2026-03-31):

• Agent reported "data insufficient" 3 times without searching HuggingFace
• Agent fabricated evaluation thresholds instead of reading eval-results.json
• Agent required human correction at every step to continue the loop
• Once the loops were coded as scripts, they ran correctly every time

The Rule