Track

VT Verification & trust

Evals, testing agent output, benchmarks, security, governance and guardrails.

From the issues

#01 · Myth vs Reality

The Approve button is a receipt for the wrong thing

No. A human approving each command doesn't make a coding agent safe, because the risk sits in what the command sets off. In 40 of 111 approval records studied, the fields a human sees left effects uncovered. Sandboxes and effect checks do the protecting. The click is just a receipt.

Numbers on this track

NumberWhat it measuresSourceCheckedTrack · issue
40 of 111Agent approval records where the fields a human sees left effects uncoveredAgent Approval Laundering, arXiv:2609.28586
✓verified at sourceVTVerification & trust
Issue #01

Papers on this track

Other tracks

ISIntent & specs CKContext & knowledge ABAgents that build UMUnderstanding & modernisation POPeople & operating model

Last updated . Reuse with credit under CC BY 4.0: “The Dabbawala Protocol, thedabbawalaprotocol.com”.