Skip to content

Security & Design

TrustSight reads hostile AUR content. This section explains the guarantees behind that, how they are enforced, how evidence is grouped, and what the analysis cannot see.

Page Answers
Security Model What TrustSight guarantees, and which gate proves each guarantee
Security Boundaries The A1-A15 invariants that keep the reading machine safe
How TrustSight Works The analysis pipeline from parse to verdict
Evidence and Confidence The tier and layer axes, and what confidence means
Assurance Layers The L1-L8 evidence groups, coverage states and a worked report
Threat Model and Limitations The adversary, the trust boundary, and the inherent limits

Where each subject is owned

Subject Authoritative location
Security guarantees Security Model
Enforcement mechanisms Security Model and the Enforcement Map
Evidence layers Assurance Layers
Evidence classification Evidence Categories
Analysis limitations Security Model and Threat Model and Limitations
Detection interpretation Rule Interpretation

Other pages link to these sources instead of restating them in full. A change to a guarantee, a B clause or an enforcement gate is a security-documentation change and needs the same review as the code it describes.

Supporting pages

These are reachable from the pages above and remain part of the security record, but they are not primary navigation entries:

The rule that keeps this section honest

A security claim without a gate is a claim not made. Every guarantee in the security model maps to an executable check in scripts/security_gates.py or one of the CI workflows, and a change to a guarantee is reviewed as a security-documentation change, not as editorial cleanup.