The engineering problem
A plausible generated finding is not sufficient evidence of a vulnerability. Polysec separates investigation from verification and report generation so each stage has a defined artifact and a reviewable responsibility.
An explicit orchestration layer
Python coordinates bounded parallel investigations. Atomic file writes carry structured results between roles; Pydantic schemas, cost ledgers, status records, and operator gates make interrupted or incomplete work visible.
Tools and execution boundaries
Profiles select domain-specific tools through an MCP registry and runtime adapters. Mock execution supports tests; Docker-backed execution and proxy rules can constrain configured workloads. Those controls depend on actual runtime selection and routing, so they are not described as universal containment.
Evaluation before claims
Severity baselines, drift checks, and reproduction artifacts support experiments about output quality. The local unit run passed 2,044 tests with two skipped. That result covers tested logic, not live provider correctness, Docker isolation, or real-world vulnerability discovery rates.
Current scope
Research tooling, not an autonomous audit guarantee. Isolation depends on the selected runtime and tool routing; mock execution is not containment. Scope filters and proxy policies have defined coverage, not universal enforcement. Historical campaign notes are not independently reproduced benchmarks. The November 2026 unit run passed 2,044 tests with two skipped; Docker and paid-provider behavior require separate integration testing.