Gemma4Bolmor replaces generic LLM German with a verifiable, trained voice. Seven LoRA adapters on a shared Gemma-4 31B base, each tuned to one pipeline role. The target was not "passable" — it was output indistinguishable from human writer references in blind A/B review (val-loss < 0.25, n=40). Getting there meant rebuilding the training-data stack from the ground up.
The Writer S1 and S2 loaders searched for fields named draft and prose. The actual data used text and the nested repair_patch.patched_text. Both writer LoRAs — the models that actually produce prose — trained on zero pairs. Silently.
augment_data.py called randomize_protagonist(). That function never existed — the real name was randomize_all_names(). The pipeline had never successfully augmented anything.
curate_source_data.py printed statistics and exited 0 without writing output. Every run looked green; no curation happened.
Eval test T01 produced "Ich habe ihn gefunden" twenty times in sequence. Cause: training without response-masking — the model was learning to predict the instructions as well as the response.
Ran upstream of everything else. Without fixing it first, no amount of retraining would converge.
harvest_project_artifacts.py extracts 267 training pairs from real production pipeline runs (Lena_Schatten, Ankerbrecher). Total dataset grew from 262 to 1,055 pairs — a 4× step up without any synthetic generation.
train_on_responses_only enabled across all LoRAs. The model learns to produce responses, not to parrot instructions back.
Deterministic regression tests for POV repair, meta-suppression, speech-tag variety, sensory openings, markdown bleed, and overall quality. Each training run is gated against them.
Seven LoRAs on one gemma4:31b base. Ollama Cloud cold-swap in under 800ms means the pipeline can switch roles mid-chapter without a reload penalty.
Quality is not "good enough" — it is val-loss < 0.25 on the repair-writer set, confirmed by n=40 blind A/B against human references. The bar is measurable, not aesthetic.