Note **The main artifact.** 51.4M documents paired with the first- and third-person constitution reflections the released models were trained on, plus the exact token offset each was inserted at. 155 GB.

Note Selection manifest for the 1T-token Dolma 3 subsample: which documents were chosen, in which order, and their safety scores. Rebuild the corpus from public data by document id — 16 GB instead of 2.6 TB.

Note Megatron `.idx` document boundaries, per-document token counts and sha256s — verify an independently rebuilt corpus byte for byte without downloading the 2.17 TB token streams.

Note SafeLM safety scores (0–5) and full class probabilities for 391M Dolma 3 documents. Publishing these removes the one stage of the pipeline that is not bit-reproducible.

Note Earlier 10M-document reflection run, superseded by `reflection-50m`. Kept for the ablations that used it.

Note 2,000-row sample of the reflection data in the identical schema — for inspecting the format without downloading 155 GB.