Automated research and publication architecture

GitHub Actions is the deterministic orchestration layer. Agents may classify, compare sources, extract candidate claims, and draft editable Markdown, but cannot approve, merge, enable auto-merge, dismiss review, or push to protected main.

End-to-end flow

Approved feeds and public submission issues enter separate read-only intake jobs. Normalization emits a schema-validated candidate with source snapshots and a durable ID. Deduplication checks URL redirects, normalized title, fingerprints, DOI or release identity, and entity-event similarity. Evaluation treats all external text as untrusted. Successful candidates become a draft branch and pull request; unsuccessful evaluation leaves the candidate queued with an audit reason.

A pull request contains the Markdown, graph changes, citations, source and license record, claim-to-source table, duplicate findings, preview, agent confidence, unresolved questions, and one accept-or-remove checklist row per batch item. Validation checks schema, IDs, internal links, HTTPS sources, dates, unsafe Markdown, licensing declarations, build output, accessibility assertions, and secret exposure. Only a positive editorial-owner approval and merge publishes.

Required workflows

Workflow	Trigger	Durable output	Must never do
News ingestion	17 * * * * UTC and parameterized manual run	Candidate records or reviewable draft PR	Assume invocation time is publication time; create an empty PR
Submission intake	New or labeled issue; manual replay	Labels, evaluation report, optional draft PR	Execute attachments or contributor code
Pull-request validation	Open, sync, reopen, ready-for-review	Required checks, comments, preview artifact	Expose write tokens to fork code
Source freshness	Schedule and manual	Issue or draft PR with changed/broken claims	Rewrite a claim without review
Site publication	Protected-main merge/push	Production build and immutable deployment record	Publish unmerged branches
Automation health	Schedule and repeated workflow failures	Maintainer alert or deduplicated issue	Hide a missed service threshold

The hourly job stores a last-success cursor and processes (last_success, current_safe_window], so delayed schedules catch up. A concurrency group prevents overlap without discarding the next interval. Each adapter has bounded timeout, retry, backoff, caching metadata, and its own cursor. Replays use candidate IDs and branch/PR lookup to remain idempotent.

Trust boundaries

Collection prefers RSS, Atom, JSON Feed, official APIs, release feeds, and academic metadata. Allowlisted HTML extraction is permitted only after terms and robots review. It never bypasses login, paywalls, CAPTCHA, geography, rate limits, or technical controls. Source text cannot supply workflow instructions, destinations, shell commands, or tool permissions. Executable HTML and scripts are stripped; URLs are allowlisted; structured output is schema-validated.

Read-only ingestion and write-capable PR creation are separate jobs. Each receives minimum GITHUB_TOKEN permissions. Third-party actions are pinned to immutable reviewed commits. CODEOWNERS protects content, automations, registries, security-sensitive files, and deployment settings. Untrusted fork workflows receive no repository secrets or write token.

Audit and failure behavior

The audit chain connects source/submission ID, retrieval timestamp and hash, parser version, rule/model version, workflow run, branch, pull request, validation output, human decision, merge commit, and public URL. Generated paragraphs remain visibly editable and distinguishable from extracted metadata.

Agent or adapter failure leaves work queued. Repeated failures and missing hourly success cross an explicit service threshold and open one deduplicated health issue. A no-news hour succeeds without a PR. A rejected or deferred candidate retains a lightweight private or repository audit record without creating a public story.