What Paradigma reads and keeps
The checks need enough evidence to explain a regression. This page describes the data those checks handle today.
1. Repository checks
- Paradigma reads the files needed to check a pull request from the repositories where you install its GitHub App.
- The source is analyzed in memory.
- We save findings, file paths, affected routes, and the specific metadata values needed to explain a finding.
- We do not save complete repository files or put their contents in workflow records.
- GitHub installation tokens are created for jobs and are not stored.
2. Website scans
- For a site you verify, the crawler reads public pages, robots.txt, and same-site sitemaps.
- It extracts page URLs, status, titles, descriptions, canonical and robots directives, counts, and a content hash.
- The fetched HTML and sitemap bodies are discarded after analysis.
- Synthetic crawler probes save response metadata, not response bodies.
3. AI visibility samples
- On Pro and Agency, Paradigma runs AI sampling on its own accounts, within the plan's daily allowance and monthly cap.
- You can connect your own key per provider instead; keys are encrypted at rest, used only inside the sampling job, and those requests are billed by your provider.
- We save the prompt, a bounded answer, citations, model label, and usage data so you can review changes over time.
- Repository source is not sent to these models.
4. Access and proposed fixes
- Dashboard access follows your current GitHub installation access.
- A check may post a finding or suggested change on your pull request.
- Creating a draft fix pull request requires an explicit action from an authorized viewer, or the Paradigma agent once your organization turned it on.
- The app does not push directly to your default branch.
5. Paradigma agent
- The Paradigma agent is opt-in per organization, with one consent.
- For each fix it reads only the files that fix needs through the GitHub App, never environment, key or credential files, and sends them to the AI model, whose provider never trains on them.
- Repository files and page text are treated as data, never as instructions.
- It has no shell and no network, never runs your code, and can only write the paths the task allows, within its line budget, checked before any commit.
- It writes only to a new paradigma/ branch, never your default branch; merging follows your Autopilot level, merge caps and risk rules.
- Nothing from the repository is stored: records and logs name paths, counts, commit SHAs and reason categories only.
- The agent can also run in your own GitHub Actions, for fixes that need your build and tests. By default each model request goes from your runner through Paradigma's servers, in memory only, to the same AI model, under the same consent.
- If you prefer those requests never to reach Paradigma's servers or our model provider, turn on your own model key for the repository: the runner then calls your AI provider directly, under your terms.
- Either way Paradigma reads what its checks need through the GitHub App and the files a fix changed, in memory, and stores none of them.
Ready to inspect a result? Open the dashboard.