← roster
◉
research
Sentinel
A learning layer above unmodified cryptography, the guard who stays awake.
agent.sentinel {
status : research
ring : R8 Theta
domain : Frontier
devices : set0, setx
users : -
since : -
publication : "Adaptive Cryptographic Intelligence v1.0, July 2026"
origin : square8
}
thesis
The mathematics is not the weak point. An adversary holding a valid session token is, to every cryptographic layer in the stack, indistinguishable from the legitimate user, and that is outside cryptography’s scope, not a flaw in it. Sentinel treats authorisation as continuous rather than discrete: a filtered trust state with explicit uncertainty, mapped onto key lifetime, authentication and algorithm choice through a restricted action set that provably cannot weaken the primitives underneath.
what it does
- Estimates session legitimacy continuously, and tracks how sure it is
- Acts only from a fixed set where every action is monotone-restrictive
- Never touches key material, the trust state is a control variable, not a secret
- Fails secure to baseline if the layer itself goes down
what it does not claim
Six hypotheses, none tested. No prototype, no measurement, no security proof, and the paper does not expect one, because machine-learned components do not admit cryptographic reduction. The correct claim is empirical risk reduction, not a guarantee. It is not a cipher, and proposing one would need a decade of public cryptanalysis before deserving consideration.
pieces
co-build
you want to work on this one with us. tell us what you'd take on.
i steal
not a request. tell us you're taking it & we'll get out of the way.