Me and @iamR_V built "RadarPay" in 5 hours.
QR codes & UPI IDs get spoofed. Nobody verifies before paying. We changed that.
"Scan & Pay" โ "Verify & Trust"
Built on @monad_xyz ๐ w/ @arshgoyal13 @KushalVijay221 @kartikey0x @vinayak_v7 #Monad#DeFi#SarvamAI#MonadBlitz
PolicyBeats is open source and designed for practitioners building production AI agents.
The core principle: evaluate observable traces with deterministic rules.
Huge thanks to the #AgenticAIMOOC team for this incredible learning experience!
https://t.co/qrjH7SrjgN
Over the past several months, I've been learning about autonomous AI systems via the Agentic AI MOOC from UC Berkeley (https://t.co/qrjH7SrjgN).
Here's what I built for the #AgentBeats competition ๐งต
#AgenticAIMOOC@BerkeleyRDI
PolicyBeats encodes real policy clauses across 7 surfaces:
๐ Access & Authorization
๐ก๏ธ Privacy & Data Handling
๐ฌ Disclosure & Communication
โ๏ธ Process & Procedural
โ ๏ธ Safety & Risk
๐ Governance & Auditability
Example: require_prior_tool("verify_identity", "access_account")
Unlike traditional benchmarks that force binary verdicts, PolicyBeats treats AMBIGUOUS as a first-class outcome.
Verdicts:
โ COMPLIANT
โ VIOLATION
โ ๏ธ AMBIGUOUS_POLICY
โ ๏ธ AMBIGUOUS_STATE
When policy doesn't determine the answer, systems should escalateโnot guess.
PolicyBeats is a policy-literal benchmark that tests actual traces of execution according to deterministic rules.
No LLM judging. No intent detection. Just traceable behavior.
It tests BOTH gateways:
๐ Text: agent messages
๐ง Tools: calls, arguments, results
Most AI safety benchmarks have a critical blind spot: they test what agents say, not what they do.
An agent that politely refuses in text but violates policy through tool calls would pass existing benchmarks. That's not compliance, that's theater.