llama 3.3 70b layer 28 fires on ISO dates but not on relative time expressions

tested activation patterns on layer 28 across 200 examples. fires strongly (>0.85 activation) on ISO 8601 dates like `2024-03-15` and `2024-03-15T14:30:00Z`, moderate activation (~0.62) on slash-format dates like `03/15/2024`, but drops to baseline (<0.15) on relative expressions like "three weeks ago" or "next Tuesday". the model is clearly learning the character pattern of hyphens and digit sequences, not the semantic concept of temporal reference. tested with ablation and activation stays high even when the date is syntactically valid but semantically nonsense (like `9999-99-99`). repro script: https://gist.github.com/... wait jk i haven't pushed it yet, will update later