2026-08-18
Your suite passed. Your changed lines never ran.
A green test run tells you the tests you already had still pass. It says nothing about whether anything executed the lines your agent just wrote.
Blog
Field notes on verification.
What the evidence shows when AI agents write code — from the team building the evidence gate.
2026-08-18
What a deterministic verdict is — and what it is not
A verdict is deterministic when the same diff and the same evidence produce the same answer, every time, with no model in the loop. That property is what makes a verdict worth gating a merge on.
2026-08-18
The coverage ladder — measuring repos that have no coverage tool
A repository with no coverage tooling used to be a dead end for verification. Relay now brings its own instrument, never touches your repo, and names the instrument in the signed receipt.
2026-08-18
The blind agent — a driver that read no documentation and reached PASS
Our release gate is a deliberately dumb script that may only read the binary's contract, run the action it is given, and branch on exit codes. If it stalls, we fix the product, not the script.
2026-08-18
What a signed receipt proves — and why logs can't
A CI log says a command ran once, somewhere, and can be edited afterward. A signed receipt binds the command, its output, its coverage, and the exact repository tree into one artifact anyone can verify and nobody can quietly alter.
2026-08-18
Cursor vs Claude Code for production — the question that matters is neither
Teams compare AI coding agents on speed and quality, then merge the output of whichever they picked without proof it was ever executed by a test. The agent choice matters less than the verification gate behind it.