Continuous Adaptive Risk and Trust Assessment (CARTA)

Kest implements the CARTA model by propagating and attenuating trust scores throughout the execution graph.

Trust Scores (0–100)

A trust score is an integer between 0 (untrusted) and 100 (fully trusted). Using integers makes thresholds immediately legible in policies without floating-point comparison issues.

Origin	Score	Description
`"system"`	100	Internal system components, cron jobs
`"internal"`	100	API gateway, verified internal services
`"verified_rag"`	90	Verified RAG pipeline sources
`"third_party_api"`	60	Trusted external APIs (Stripe, GitHub)
`"user_input"`	40	Direct human user input
`"internet"`	10	Untrusted public web sources
`"llm"`	0	Raw LLM output

Trust Evaluators

Trust propagation is handled by a TrustEvaluator. The default uses a weakest-link model:

python

from kest.core.models import DefaultTrustEvaluator
 
evaluator = DefaultTrustEvaluator()
# trust = (min(parent_scores) * self_score) // 100

Policy Integration

Policies receive trust_score as an integer. For high-value operations, set a high minimum threshold. For public entry points, set a lower threshold.

Rego Example

rego

package kest.allow
 
import future.keywords
 
default allow := false
 
allow if {
    input.trust_score >= 80
    input.workload_id == "trusted-service"
}

Cedar Example

cedar

permit(
    principal,
    action == Action::"TransferFunds",
    resource
) when {
    context["trust_score"] >= 80
};

Common Thresholds

Threshold	Meaning
`>= 100`	Fully trusted; internal system only
`>= 50`	Internal workload or verified delegation
`>= 10`	Minimum viable trust (internet entry with valid identity)
`= 0`	Blocked; raw LLM output or completely untrusted