Only necessary capability enters the run
Models, sources and tools are filtered before context construction.
Security and authority
Probabilistic intelligence. Deterministic boundaries.
LowToHi assumes models can be useful and still be wrong, manipulated or overconfident. Security begins by preventing cognition from becoming authority.
Security principles
Identity and scope remain server-resolved
Browser-provided tenant, workspace or capability claims are not trusted as authority.
Tools expose bounded operations
A tool invocation cannot silently become unrestricted filesystem, network or production access.
Consequential actions leave receipts
Versions, approvals, sources, outputs and effects remain inspectable after execution.
Control layers
Pre-context admission
Unauthorized data and capability are removed before the working set exists.
Bounded cost and continuation
Token, time, tool and operational limits constrain the execution.
Human control over high-consequence effects
Sensitive operations pause until the configured authority grants exact approval.
Restart, backup and rollback
Operational state and evidence are designed to survive process failure and controlled restoration.
SECURITY
Responsible disclosure
Security reports should identify the affected version, boundary and reproducible impact without exposing secrets or user data. Configure the public security contact before launch.