Can çıkar, huy çıkmaz
Base-model behavior persists through all fine-tuning and prompting.
A model prone to sycophancy, hedging, or verbose rationalization will exhibit these traits regardless of system-prompt sophistication. Fine-tuning shifts probabilities but does not rewrite the base distribution. Design harnesses that compensate for known pathologies—validation layers for sycophancy, length limits for verbosity—rather than pretending a stern system prompt cures nature.