box C in_lang: 1,641 rows, 68 languages, orthography scored

270m_in_lang/README.md

@@ -0,0 +1,25 @@
+# Box C generated text — `in_lang`
+
+1,641 rows across 68 languages, generated on CPU with `google/gemma-3-270m`
+and scored with this project's rule-based metrics, **including `tier2_orthography`**
+(diacritic density against each language's own corpus rate — the strongest single
+discriminator in the rule set, AUC 0.955 separating generated from real text).
+
+## Prompt system: `in_lang`
+
+Each prompt is a short (12–35 word) sentence of **real corpus text in the target language**, which the base model continues. Every snippet passed two independent language checks before being used — diacritic rate against the language's own corpus rate, and overlap with that language's 500 commonest words — because an earlier attempt at corpus-derived seeds harvested task-formatted rows and labelled Latvian as Yoruba.
+
+Measured on the output: english-ness **0.10** (clean corpus 0.063, the English-prompt runs 0.45), vocabulary **0.34** (clean 0.705), repetition **0.30** (clean 0.013). In-language, and clearly below real text on every quality dimension.
+
+## Columns
+
+`prompt_system` and `seed_source` identify the generation regime; `source_box`
+records the machine. Every other column matches the main dataset's schema.
+
+## Honest limits
+
+- `gemma-3-270m` is a base checkpoint, not instruction-tuned, and has no chat
+  template. It continues text; it does not follow instructions.
+- Rows where `tier1_is_text` is false have no `tier2_*` scores: the gate short-circuits
+  grading. They are still informative — they failed on digit runs and repetition,
+  which is exactly the degenerate behaviour worth labelling.

270m_in_lang/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:073ab8dd3f97b09c76297d95f54ce022a4953b815f854a14f90345563ad0a862
+size 917991