Limitations

Scientific limitations

Acoustic features are not biodiversity measures. ACI, the Biological-Band Spectral Magnitude Ratio, and other features capture aspects of acoustic structure but do not directly count species or measure total biodiversity.
Recovery score is model-dependent. The score depends on the selected features, scaling method, and distance metric. Different configurations may produce different scores for the same recordings.
No causal inference. A higher recovery score does not prove that restoration caused the change. Controlled study designs are needed for causal claims.
Not comparable across ecosystems. Scores from different ecosystems (coral reef vs. forest) are not directly comparable without calibration.
Bootstrap uncertainty is not certainty. The confidence interval reflects resampling variability under the current dataset, not external validity.

Single-worker processing. The prototype uses a single background worker. Large datasets will process sequentially.
SQLite database. Suitable for local prototype use. Not designed for concurrent multi-user access.
Local file storage. Audio files are stored on the local filesystem. No cloud storage integration.
No authentication. The API has no authentication. Not suitable for production deployment without adding auth.
No real-time updates. The frontend polls for job status. No WebSocket or SSE for real-time updates.

Existing unversioned SQLite databases must be backed up and upgraded with alembic upgrade head. The initial adoption migration is intentionally irreversible; restore a backup to roll back.
The CC0 Amazon source manifest is a technical pipeline demonstration only. Its synthetic degraded references and repeated source observations are not ecological recovery evidence.
Docker is configured for production static frontend serving and schema-first startup, but container runtime verification requires Docker to be installed on the deployment machine.