# Source policy

## Admission rules

A built-in source must be one of:

1. A primary paper endpoint operated by the repository itself, currently arXiv Atom.
2. A laboratory-controlled domain publishing research, engineering, safety, model, or product information.
3. A verified organization page hosting code, models, datasets, or model cards.
4. A named founder/core person's own blog or X account, with a separate first-party lab page establishing current identity and role.

The plugin does not ingest media, SEO summaries, anonymous reposts, generic X search, influencer accounts, or arbitrary URLs supplied by the model.

Each source can declare a bounded quality contract: category allow/deny lists, required publication dates, excluded index paths, known boilerplate titles, and a staleness window. These checks run before technical relevance scoring and persistence. Rejections are counted by reason so stricter rules cannot silently erase records.

## Provenance fields

Every record retains its source id/class, lab, canonical URL, source endpoint, collection time, and any person identity evidence. Ranking never removes these fields. Cross-source duplicates retain the first normalized record and list corroborating source ids.

For arXiv records, the plugin preserves the stable paper id for deduplication and the observed versioned id/URLs for execution evidence. It may query Hugging Face Paper Pages for linked public repositories, models, datasets, and Spaces. These community-maintained links are artifact leads, not independent verification. Public GitHub repository roots may be resolved through the official REST API to a full default-branch commit plus license/archive metadata. A `github_org` source additionally observes public, non-fork, non-archived repositories and a bounded set of Releases/Tags; draft Releases are excluded, tag targets are pinned to commits, and GitHub-provided release-asset digests are retained. Public Hugging Face model and dataset repositories may likewise be pinned to their full Hub SHA. Mutable stars never contribute to truth or source-authority scores.

Release evidence clustering is lab-scoped and entity-keyed. It never uses vague embedding similarity to merge different organizations. A recognized model family/version can join official publications, artifacts, papers, evaluations, and verified person signals; unknown names remain isolated one-record bundles. A bundle's substantive digest excludes observation timestamps and mutable popularity, so watches react only to evidence-bearing changes.

## Source health

Successful and failed collection observations update a small persistent health record: last attempt/success, newest accepted publication, accepted/raw/rejected counts, count delta, structure fingerprint, failure streak, staleness, and last error. A material count drop, changed structure, repeated failure, or stale newest item raises an alert. Health is operational evidence about the adapter and upstream publication cadence; it is not a truth or project-quality score.

Person identities can change. The built-in catalog is conservative and should be re-audited on releases. When a reliable, lab-verifiable public channel does not exist, absence is preferred to a guessed account.

## Source classes

`official_lab` and `official_artifact` receive the highest source score, followed by primary papers, person blogs, then person X posts. This order reflects suitability for implementation evidence, not a judgment that a laboratory claim is automatically true.

## X

X collection uses `GET /2/users/by/username/:username` and `GET /2/users/:id/tweets` with a user-provided bearer token. There is no cookie reuse, browser automation, proxy, RSS bridge, or scraping fallback. Account access, rate limits, and pricing are external conditions and are reported as source failures without stopping other sources.

## Network safety

Only HTTPS is accepted. Literal loopback, link-local, and private IP targets plus embedded credentials are rejected; redirects are revalidated; every response has a time and byte limit. Custom source files are trusted administrator input and cannot be supplied through model tool arguments. Deployments with stronger egress requirements should enforce an outbound allowlist at the OS/container boundary as well.

## Content use

The corpus stores bounded titles, metadata, summaries, and outbound artifact URLs. It does not archive full articles or PDFs. Follow each source's terms, robots policy, API agreement, artifact license, and data restrictions.