Limitations

Found by reading the code; file:line links open the exact line.

US-English assumptions

Assumption	Where
ASR: Whisper large-v3 with language="en"	scripts/run_whisper_asr.py:36, :66
Audio encoder openai/whisper-small.en (English-only)	kirad/base_models.py:849, kirad/datasets_and_dataloaders.py:380, all YAMLs
Text model google-bert/bert-base-uncased	kirad/base_models.py:922, kirad/datasets_and_dataloaders.py:546
LLaMA prompt in English (PHQ-9 only)	research/stable/ordinal_regression_mtl_jun_2024/config_llama.yaml
Cutoffs PHQ-9 (9, 14), GAD-7 (4, 9, 14); maxima 27 / 21	kirad/constants.py:37, :41
Fixed 80-bin log-mel profile IDEAL_LOGMEL_ENERGIES	kirad/constants.py:172, used at kirad/feature_extraction_whisper.py:94
US survey schema for subject merge (zip_code, state, english_preferred, …)	kirad/dataset_utils.py:663
16 kHz assumed, never resampled (only a warning)	kirad/constants.py:14, kirad/datasets_and_dataloaders.py:479

What blocks a run today

#	Blocker	Where	Fix
1	wandb_entity / wandb_project are None and raise	scripts/launcher.py:172	set them; W&B is also required by the score tables (kirad/base_models.py:448, wandb.run.id)
2	get_package_info("kipy"): private package, find_spec returns None	scripts/launcher.py:255 → kirad/utils.py:740	drop "kipy" from the list
3	df_or_dataset_version: "v2.x" downloads private splits with gsutil; URIs are None	kirad/dataset_utils.py:183, kirad/constants.py:116	point it to a CSV per split
4	DAM 1 config.yaml train uses merge_rows_by_subject: many_to_one; the merge drops audio_filename, which the audio dataset then reads	merge: kirad/dataset_utils.py:692; read: kirad/datasets_and_dataloaders.py:465	remove the option (or recompute the path list after merging)
5	audio_dir_or_feature_store_bucket is not a field of KintsugiAudioDataset (extra="forbid") → validation error	config_dam1.0.yaml, config_dam2.0.yaml, config_llama.yaml, config_dam3.0.yaml, config_llma.yaml, config_llma_dam3.0.yaml	use audio_root under metadata=
6	text_dir is not a field of KintsugiTextDataset (extra="forbid")	every config with text= (DAM 2 config.yaml, config_dam2.0.yaml, config_llama.yaml, config_llma_dam3.0.yaml)	use text_root under metadata=
7	ideal_logmel_energies: "logmel_energies.npy" is not in the repo → np.load fails	kirad/datasets_and_dataloaders.py:403, also scripts/launcher.py:274; all DAM 2/3 configs and config_dam1.0.yaml	provide the file or remove the key
8	Placeholders <wandb_url_to_…> (LLMA, LLaMA, backbone cache, DAM 2 checkpoint)	DAM 3 configs, config_llama.yaml	train the upstream stage first and paste its W&B URL
9	Text runs read row.asr_output_id	kirad/datasets_and_dataloaders.py:600	add the column (any value)

Only DAM 1 config.yaml avoids 5–8; with 1–4 fixed it is the shortest path to a first run.

Missing data and metrics

• Data: metadata CSVs, audio buckets and logmel_energies.npy are private (GCS/W&B); no checkpoints or teacher scores (score_targets_csv) are included.
• Preprocessing: VAD and resampling are upstream and not in the repo; only the VAD parameters remain (kirad/constants.py:22).
• Outcomes: only PHQ-9 and GAD-7 are wired in configs; a binary real task (exact2, cutoff 0) exists in kirad/constants.py but no config uses it (candidate for SCID).
• Metrics: no calibration, no per-country / per-language breakdown by default (analyze_dataframe_with_scores groups by any column, but is not called in training); no confidence intervals (metrics.bootstrap_confusion_matrix exists but is never called).
• Tests: none in this copy.