# Threat model The toolkit treats the public network, forwarding headers, request targets, cookies, credentials, and stored request records as untrusted. Application code, the configured trusted-proxy set, server filesystem permissions, and explicitly selected storage adapters are trusted. Controls include explicit proxy trust, bounded parsing, cryptographic request and session identifiers, digest-only session storage, Argon2id passwords, constant-time comparisons, same-origin and CSRF primitives, fail-closed storage errors, separate safe/sensitive analytics projections, organization-scoped bindings, single-use invitation digests, and short-lived audited break-glass grants. Unsafe methods without an exact Origin or trustworthy same-origin Fetch Metadata fail the origin check. Authentication middleware fails closed when its service or `__Host-` cookie policy is invalid. Imported request records have bounded byte and duration fields before analytics sums them. The toolkit does not sandbox application handlers, secure an incorrectly configured reverse proxy, authorize application routes automatically, encrypt a compromised host, or decide how long an operator may lawfully retain personal request evidence. Applications must pass the authenticated user and requested resource hierarchy to `access.Authorize`; possessing a platform-level `auth` role does not bypass that decision. Team membership is resolved by the repository rather than accepted from request input. Break-glass access lasts at most one hour and is not a substitute for ordinary role policy. Local storage adapters assume the parent directory and host account are trusted. They reject a symlink at the configured final path and apply private file modes, but they do not defend against a concurrent privileged actor replacing path ancestors during an open. The synchronous JSONL adapter deliberately favors durable, bounded evidence over maximum request throughput; the application owns rotation, retention, disk monitoring, and health escalation.