Who ensures it followed the specs?
The more context an LLM gets, the more likely it will start to ignore instructions.
If the LLM runs a context compression, all bets are off. There's a reason Anthropic upped the context to 1M tokens to reduce the chance of this from happening.