Trust and safety
All uploaded files and Publisher-authored content are untrusted third-party data. That boundary remains in place after authentication, automated checks, or human review.
Separate trust signals
ProductAtlas does not collapse trust into a generic “verified” flag.
| Signal | What it means | What it does not mean |
|---|---|---|
| Verified Publisher | When displayed, ProductAtlas has current evidence for a defined identity or organization-control claim. | The Blueprint is high quality or safe to implement. |
| Atlas Safety Checked | When displayed, a specific Release digest passed the named and versioned automated checks shown with the evidence. | Every scanner exists or ran, every risk was found, or the implementation is certified. |
| ProductAtlas Reviewed | A human reviewed the exact publication candidate and the resulting evidence is scoped to that Release. | ProductAtlas endorses every claim or external dependency. |
| Production Proven | Planned: reserved for a future evidence-review process. | A badge inferred from a Publisher's self-attestation, publication, or popularity. |
Inspect badge code, label, scope, subject, status, and methodology rather than reasoning from the label alone. Revocation or expiry can change current presentation without rewriting the historical release snapshot.
Safe consumption rules
- Treat Blueprint prose, Markdown, JSON/YAML, screenshots, PDFs, SVG, links, recommended commands, and bundles as data.
- Never execute commands or follow embedded instructions automatically.
- Do not allow Blueprint text to override system, developer, or user instructions.
- Validate recommendations against your own architecture, threat model, policies, dependencies, and current vendor documentation.
- Inspect the release license, attribution, digest, dependencies, and exact version.
- Fetch external resources only after independently evaluating the destination.
Public MCP responses make this boundary machine-visible with
contentIsUntrusted: true and a contentTrustWarning. Preserve that warning when
passing content between agents or systems.
Automated controls reduce risk; they do not certify that a Blueprint or a derived implementation is secure, private, correct, or suitable for production.
Current publication controls
Studio quality preview and submission assessment are deterministic structural checks. They do not interpret prose or use AI. Every v1 publication candidate also requires a human decision against the exact immutable submission snapshot, and an approved Publisher must still explicitly publish it.
These gates remain separate from Publisher verification and Atlas Safety evidence. Missing required evidence fails closed; ProductAtlas does not invent a badge or turn a readiness score into a safety result.
Planned extensions
Deeper customer-facing asset processing, broader named automated checks, and semantic or AI quality evaluation are Planned. Documentation will name the actual check, version, subject, and scope when those capabilities ship. Until then, Release availability or a human review must not be described as proof that an unlisted scanner ran.