# Sources

> Every claim on CachedDebt.com cites one of these papers or official tool documents, with the quoted passage, what it supports, and its limits.

Canonical page: https://www.cacheddebt.com/sources/

Every external claim on this site points here. Each entry shows the quoted passage, what it supports, and where it stops.

Quotes last checked against the live pages on October 22, 2026. A script re-fetches every page and fails when a quote disappears. Spot an error? See the [corrections policy](https://www.cacheddebt.com/about/#corrections).

## Research

### [1] Hidden Technical Debt in Machine Learning Systems

- Citation ID: `sculley-2015`
- Publisher: Sculley et al., Advances in Neural Information Processing Systems 28 (NeurIPS 2015), 2015
- URL: https://papers.nips.cc/paper_files/paper/2015/hash/86df7dcfd896fcaf2674f757a2463eba-Abstract.html
- Quote: "It is common to incur massive ongoing maintenance costs in real-world ML systems."
- Supports: The canonical prior art. Names undeclared consumers, unstable data dependencies, legacy features, dead experimental codepaths, and configuration debt.
- Limits: Written about production ML pipelines, before LLM agents and local model caches were common.

### [2] Context Rot: How Increasing Input Tokens Impacts LLM Performance

- Citation ID: `chroma-context-rot`
- Publisher: Hong, Troynikov, and Huber, Chroma, 2025
- URL: https://www.trychroma.com/research/context-rot
- Quote: "Model performance consistently degrades with increasing input length."
- Supports: Retained context has a quality cost, not only a storage cost.
- Limits: Vendor technical report, not peer reviewed. Measures input length, not staleness.

### [3] Lost in the Middle: How Language Models Use Long Contexts

- Citation ID: `liu-lost-in-the-middle`
- Publisher: Liu et al., Transactions of the Association for Computational Linguistics, vol. 12, 2024
- URL: https://aclanthology.org/2024.tacl-1.9/
- Quote: "Performance is often highest when relevant information occurs at the beginning or end of the input context."
- Supports: Adding more context does not guarantee the model uses it.
- Limits: Studies position within context, not outdated content.

## Tool documentation

### [4] Migrate to a New Embedding Model

- Citation ID: `voyage-embedding-migration`
- Publisher: Voyage AI by MongoDB documentation
- URL: https://www.mongodb.com/docs/voyageai/tutorials/migrate-embedding-model/
- Quote: "Embeddings that different models generate aren't comparable, so a migration involves more than changing a model name."
- Supports: Vectors from an old model silently degrade retrieval until the old index is removed.
- Limits: Covers embedding model changes. Does not cover source documents changing under an unchanged model.

### [5] Understand caching

- Citation ID: `hf-cache`
- Publisher: Hugging Face Hub documentation
- URL: https://huggingface.co/docs/huggingface_hub/guides/manage-cache
- Quote: "At the moment, cached files are never deleted from your local directory: when you download a new revision of a branch, previous files are kept in case you need them again."
- Supports: Model caches grow by default. Old revisions stay on disk.

### [6] Ollama FAQ: where are models stored

- Citation ID: `ollama-faq`
- Publisher: Ollama documentation
- URL: https://docs.ollama.com/faq
- Quote: "If a different directory needs to be used, set the environment variable OLLAMA_MODELS to the chosen directory."
- Supports: The override variable. The same FAQ entry lists the default model location for macOS, Linux, and Windows.

### [7] docker system prune

- Citation ID: `docker-prune`
- Publisher: Docker documentation
- URL: https://docs.docker.com/reference/cli/docker/system/prune/
- Quote: "By default, volumes aren't removed to prevent important data from being deleted."
- Supports: A routine prune leaves volumes behind, so orphaned service data persists.

### [8] docker system df

- Citation ID: `docker-df`
- Publisher: Docker documentation
- URL: https://docs.docker.com/reference/cli/docker/system/df/
- Quote: "Displays information regarding the amount of disk space used by the Docker daemon."
- Supports: The read-only inspect command for images, containers, volumes, and build cache.

### [9] pip cache

- Citation ID: `pip-cache`
- Publisher: pip documentation
- URL: https://pip.pypa.io/en/stable/cli/pip_cache/
- Quote: "Inspect and manage pip's wheel cache."
- Supports: Inspect and purge commands for the pip cache.

### [10] npm cache

- Citation ID: `npm-cache`
- Publisher: npm documentation
- URL: https://docs.npmjs.com/cli/v11/commands/npm-cache
- Quote: "It should never be necessary to clear the cache for any reason other than reclaiming disk space."
- Supports: Package caches cost disk, not correctness. Treat them as low blast radius.

### [11] uv caching

- Citation ID: `uv-cache`
- Publisher: uv documentation, Astral
- URL: https://docs.astral.sh/uv/concepts/cache/
- Quote: "uv cache prune removes all unused cache entries."
- Supports: A safe, periodic cleanup command exists for the uv cache.

### [12] torch.hub

- Citation ID: `torch-hub`
- Publisher: PyTorch documentation
- URL: https://docs.pytorch.org/docs/2.14/hub.html
- Quote: "If set_dir() is not called, default path is $TORCH_HOME/hub where environment variable $TORCH_HOME defaults to $XDG_CACHE_HOME/torch."
- Supports: Where PyTorch Hub stores downloaded models. No cleanup command is documented.

### [13] Browsers

- Citation ID: `playwright-browsers`
- Publisher: Playwright documentation
- URL: https://playwright.dev/docs/browsers
- Quote: "When there are no more clients that require a particular version of the browser, that version is deleted."
- Supports: Playwright cleans up its own browsers. Copies installed outside its control are not covered.

### [14] Claude Code settings: cleanupPeriodDays

- Citation ID: `claude-code-settings`
- Publisher: Claude Code documentation, Anthropic
- URL: https://code.claude.com/docs/en/settings-reference
- Quote: "Set how many days Claude Code keeps session transcripts and other application data before deleting them."
- Supports: Agent transcripts have a retention setting. The default is 30 days.

### [15] The .claude directory

- Citation ID: `claude-code-directory`
- Publisher: Claude Code documentation, Anthropic
- URL: https://code.claude.com/docs/en/claude-directory
- Quote: "These files are plaintext. Anything that passes through a tool is written to a transcript on disk."
- Supports: Agent transcripts can retain secrets and file contents. Auto memory files are not swept. The same page documents claude purge and its --dry-run preview.
