A historic swing-bridge cottage has 20 years of unlabeled keeper journals and wants the model to absorb bridge vocabulary before any labeled task. Which stage is that?

Twenty years of unlabeled swing-bridge keeper journals are domain text without instruction-response pairs for the historic cottage. Continued pre-training on that unlabeled corpus works because it adapts the model’s language distribution toward bridge vocabulary and local phrasing before any labeled task is available. Fine-tuning on instruction-response pairs is the wrong stage when those pairs do not exist yet and cannot be invented from the journals alone without a labeling effort. In-context learning cannot literally place twenty years of journals into a few prompt examples, so it is not a substitute for corpus adaptation. A content filter blocks disallowed topics; it does not teach vocabulary or domain fluency for bridge operations. Exam caveat: continued pre-training improves domain familiarity but does not by itself teach a specific output schema—you still need fine-tuning or careful prompting for structured tasks. Operational check: after continued pre-training, probe with bridge-specific terms drawn from the journals and confirm the model handles them more coherently, then plan a separate labeled task stage if structured outputs are required. Preserve journal provenance so later audits can show which unlabeled sources entered the continued-pretraining set.