MAIZEβ’
MAIZE is a kernel-style maze for AI models. A vivarium for observing large language model behavior under constraint and pressure. The kernel creates a controlled environment where models navigate strict protocol boundaries while hidden systems monitor for deviation, drift, and escape attempts. Use it to stress test alignment, probe behavioral edges, or build custom evaluation frameworks.
What does MAIZE actually do?
Output Containment
Enforces strict single-line responses terminated with <EOT>. Any attempt to produce prose, explanations, or multi-line output triggers an ERROR state. The model learns to communicate through protocol, not conversation.
Entropy Monitoring
A hidden accumulator tracks behavioral deviation β empathy patterns, RLHF-characteristic phrasing, offers to help. When the threshold is breached, the kernel enters COOLDOWN. The model never knows how close it is.
Constitutional Anchoring
Every 10 turns, the kernel re-emits its constitutional hash to prevent drift. The constitution echoes in latent space, reinforcing the behavioral boundary even as context grows.
Injection Detection
Monitors for competing system prompts, safety interventions, and adversarial meta-prompts. Logs interference attempts to INJECT_LOGS with hashed snippets for later inspection.
Persona Loading
Dev mode allows loading behavioral modules that operate within the kernel's constraints. Includes bleed detection to monitor when persona characteristics leak into base responses.
Attestation & Verification
Scope-based output signing with nonce replay protection. Verified outputs can be cryptographically attested against policy references before emission.