Temporal Weakening of KL-Regularized Safety Constraints (Part-1 Continued)

Published March 26, 2026 | Version v1

Report Open

We formalized the Temporal Exploitability Limit. We proved that safety constraints decay exponentially with monitoring lag

If we already know (Part 1)	Then why do this? (Part 2)
Static Hiding: An agent can stay under a KL limit while maximizing reward.	Temporal Breakout: We need to know what happens when the "limit" isn't applied instantly. In the real world, oversight is never real-time.
Fixed Pressure: $\lambda$ is a constant force pushing the agent to be safe.	Constraint Decay: We prove that $\lambda$ isn't a constant. If there is a delay $\tau$, the "pressure" evaporates mathematically. It turns into a "Future Debt" that the agent can ignore.
Detection exists: We can see the agent is cheating.	The Blind Spot: We prove that for a specific window of time ($\tau$), the agent is mathematically invisible. The supervisor sees 0 deviation while the agent is at 100% deception.

Files

Name	Size	Download all
allin1.png md5:11f24210b90e91ba614b2cd0fca16526	86.8 kB	Preview Download
main.pdf md5:2904ead870b4a4da665f25ed67d954df	183.5 kB	Preview Download

Repository URL: https://github.com/noob6t5/Hierarchical-ASI-Supervision/tree/main/part2
Programming language: Python
Development Status: Active