Ask an image generator for a vintage tube radio and you'll spot the blind spot instantly. The dials are wrong, the buttons don't exist, the whole thing looks like a fever dream of a radio.
Coding LLMs have the same blind spots. We just don't see them as clearly, because broken code compiles often enough to feel right.
1. Time gap
Deprecated APIs show up in 25-38% of completions (ICSE 2025). Your model's favorite React hook might have been removed 14 months ago.
2. Domain gap
The tube-radio zone. COBOL, ABAP, IEC 61131-3, AUTOSAR, BaFin specifics, internal frameworks of large enterprises. Exists, runs production, barely in training data.
3. Context gap
Your codebase. Your ADRs. Your team's naming conventions. Structurally unknowable.
4. Structure gap
Models scoring 80%+ on isolated tasks can drop below 25% on repository-level benchmarks (CrossCodeEval, RepoMasterEval). A function the model nails in isolation breaks when it has to reach across three files.
5. The meta-gap
The model doesn't know which of the above it's in. Hallucination confidence scales inversely with training-data density. The thinner the spot, the more convincing the fabrication.
That last one is why the others are dangerous.
Practical move: before every non-trivial LLM task, ask which zone you're in. If it's 1-4, feed context (llms.txt, MCP, AGENTS.md, pinned versions). If it's 5, read more carefully than you want to.
LinkedWild