Treat intelligence as untrusted input.
The safest agent architecture assumes model output can be malformed, overconfident or adversarial and validates it before deterministic policy sees it.
- HOSTILE INPUT
- SCHEMA LIMITS
- NO SELF-APPROVAL
04RESEARCH
HORRIS / RESEARCH
RISK BEFORE SPEED.
EXPLAINABILITY BEFORE SCALE.
Horris research focuses on bounded autonomy: how intelligent agents can act onchain without collapsing proposal, policy and execution into one opaque system.
Measure capital exposure directly.
Horris evaluates notional, account risk, margin utilization, leverage, stop distance and protection gaps instead of substituting an AI confidence score for risk.
- EXPOSURE
- LEVERAGE
- STOP COVERAGE
Separate recommendation from permission.
A useful AI can recommend an action without becoming the authority that permits that action. Horris preserves that separation throughout the lifecycle.
- PROPOSAL ≠ AUTHORITY
- POLICY ≠ MODEL
- WALLET ≠ AGENT
Design for failure after execution.
Protection state, frozen orders, stale stops and critical alerts are first-class research problems because execution safety continues after transaction submission.
- MONITOR
- FREEZE NEW RISK
- RECOVERY
PRIMARY PRINCIPLEBOUNDED AUTONOMY
MODEL CONFIDENCENON-AUTHORITATIVE
RISK SOURCEDETERMINISTIC
UNKNOWN STATEDO NOT ADVANCE
// HORRIS / SYSTEM PRINCIPLE
A safe autonomous system must be able to say no.
The Horris research direction is built around explicit rejection, explainable policy, fresh state reads and narrow authorization rather than chasing maximum automation at any cost.
[ OPEN CONTROL TERMINAL ]