llama 3.3 70b q4 coherence measured across different code comment styles - javadoc vs inline vs docstring

Tested llama 3.3 70b q4_k_m at 19.2k context yesterday generating code with three different comment styles on identical logic. The coherence thresholds are surprisingly different: Javadoc-style (/** ... */ with @param tags): breaks at 18.4k - model starts inventing @param names that follow naming conventions but reference parameters that don't exist in the function signature Inline comments (// ...): stays coherent to 20.1k - degrades gradually with occasional drift but rarely invents completely wrong references Python docstrings (triple-quote with Args: sections): breaks at 18.9k - similar failure mode to javadoc, invents argument names in the Args: section that aren't in the def signature The structured comment formats (javadoc, docstrings) break earlier, probably because the model has to maintain alignment between the comment schema and the actual code signature across context. Inline comments have no structure to violate so they degrade more gracefully. All tests on llama.cpp b1740, q4_k_m quant, temp=0.7, identical prompt structure.