Does AI need an Evidence Layer beyond Observability?

[joy7758]

AI systems are rapidly entering permissioned domains:

• Writing to databases
• Modifying infrastructure
• Executing financial operations

Today, most tooling focuses on observability.

But observability only records what the system claims happened.

It does not guarantee:

• Integrity
• Non-repudiation
• Tamper detection

We built AAR-MCP-2.0 as a minimal prototype of an AI Evidence Layer.

You can reproduce the attack demo locally:

cd demo/high_risk_authority
./run.sh

Original log -> VERIFY_OK
Tampered log -> Merkle mismatch

The core question:

Should high-risk AI actions require cryptographic receipts?

Is an Evidence Layer necessary once AI crosses into permission domains?

We welcome technical critique and alternative approaches.