Card: state corpus provenance and what is released

README.md

@@ -126,7 +126,13 @@ characters it introduces, where the training corpus abandons 0.0%.
 
 The corpus is synthetic, generated by two teacher models: **ALLaM 2.7B** (75%) and
 **Qwen3.6-27B** (25%), with quality gates rejecting foreign characters and template
-openings.
+openings. Both teachers are Apache-2.0.
+
+The generated stories are **not redistributed**: they were produced through a
+commercial inference API whose terms restrict building redistributable databases
+from model output. What is released instead is the *generator* — seed space, prompt
+templates, quality gates and sampling procedure. Generation is deterministic given a
+seed, so the corpus is reproducible from what is published.
 
 ---
 
@@ -148,7 +154,8 @@ Capacity buys coherence, not vocabulary.
 ## Paper
 
 <!-- TODO: arXiv link on release -->
-*In preparation.* Corpus, metrics, checkpoints and every measurement are released.
+*In preparation.* Metrics, checkpoints, tokenizer, generation seeds and every
+measurement are released.
 
 ```bibtex
 @misc{binsafi2026ruqlm,