225 CHAT files fail chatter validation in the 2026-07 TalkBank mirror (13 corpora; skipped in 2026.1)

[mcfrank]

The validation sweep over the 2026-07 TalkBank mirror (438 corpora, 56,804 .cha files across the childes and phon banks; full per-corpus table in pipeline/validation/report.tsv) found 225 files that fail chatter validation, concentrated in 13 corpora. All other corpora validate 100%.

Per-corpus breakdown:

bank	corpus	n_cha	valid	invalid
childes	German/Leo	500	417	83
phon	Biling/Nulahan	89	41	48
phon	Spanish/PAIDUS-Spanish	77	54	23
phon	Biling/French-EP	52	31	21
phon	Dutch/CLPF	213	193	20
phon	Clinical/Cummings	126	111	15
phon	Eng-NA/Penney	52	45	7
phon	Clinical/NeumannFoxBoyer	29	26	3
phon	Biling/Almeida	110	109	1
phon	Clinical/Cattini	20	19	1
phon	Clinical/McAllister	17	16	1
phon	Japanese/Ota	63	62	1
phon	Other/Setswana/Matlhaku	33	32	1
	Total			225

Notes:

• German/Leo is the largest single contributor (83 of 500 files, 17% of the corpus) and the only affected corpus in the childes bank; the other 12 are all PhonBank corpora.
• These 225 files were skipped in the 2026.1 release (0.4% of all files), now published on Redivis: https://stanford.redivis.com/datasets/b6q6-14tmdjgw8 (v1.4). Skips are logged per file during the pipeline run; they are exactly the invalid set from report.tsv (see pipeline/validation/full_run_report.md §2).
• Consequence for users: e.g. Dutch/CLPF loses ~4% of transcripts vs 2021.1, and Eng-NA/Penney (already the subject of a phonology report, get_tokens: Empty actual_phonology and model_phonology columns  childesr#61) is missing 7 files in 2026.1.
• The import is resumable per corpus: if TalkBank repairs these files upstream, deleting the 13 corpora's marker files under parquet_full/markers/ re-imports them in minutes.

Suggested actions:

1. Report the per-file validation failures upstream to TalkBank (Leo first).
2. Document the skips in the 2026.1 release notes, and re-check the 13 corpora against a future mirror for a point release.