Datasets:

---

SlayerLab

/

polish-dynaword

like 10

Follow

Slayer Labs 19

What

Follow-up to v0.2.3. Three related fixes:

1. Complete the phrase-frequency report. The v0.2.3 CHANGELOG states the report was recomputed, but the README carried a "recount pending" note and the committed snippet was a stale 15-source run (missing nkjp1m). Regenerated across all 16 sources with src/pattern_frequency_report.py: total 6,881,277,362 tokens, refreshed tables + 10 charts, and replaced the pending note in the README "Results" section.

2. Fix SOURCES registry drift. 1000_novels and global_voices shipped in data/ but were unregistered in src/sources.py, so make_docs.py silently dropped them (saw 14/16 sources, wrong totals) and would regress the README on the next run. Registered both, with a new custom_datasheet guard in make_docs.py so their hand-authored datasheets and per-source "Added" dates are not clobbered by template regeneration.

3. Add AGENTS.md. Documents the add-source and release ritual plus the real footguns (make_docs is scaffolding, not a round-trip source of truth; the global ADDED restamp; unregistered sources are dropped; the README narrative is hand-maintained). Several recent contributions added data but skipped the doc/version bump; this should reduce that.

Notes for review

• No new data and no version bump — same 16 sources, totals unchanged (2,710,974 docs / 6.88B tokens).
• Datasheets and the README header/tables were intentionally left untouched (not regenerated via make_docs, which would restamp dates and drop curated prose).
• Judgment call: global_voices (PR #7) was added with no in-repo build script, so build_dynaword --all cannot rebuild it — it may have been left out of SOURCES deliberately. The custom_datasheet guard keeps it safe, but revert that one registration if the omission was intentional.

🤖 Generated with Claude Code