Your CI gate says FAIL. An engineer asks why. Six weeks later, an auditor asks the same question, and now the answer matters more.
Most checks can't answer it. The logs rotated. The runner is gone. The tool that produced the verdict has been updated four times since, and nobody can say what it would have said that day, about that commit, under that config.
We think a merge gate for agent-written code has to hold itself to a higher standard, for one simple reason: when an AI agent wrote the branch, "a human reviewed it carefully" is no longer the story you tell your auditor. The gate is the story. So the gate has to show its work.
Here is what Rosentic records for every verdict, automatically:
The ledger is append-only. There is no edit path, at any layer. And because the engine is deterministic (no model, no sampling, no drift), every verdict is replayable: check out the recorded SHAs, run the same engine version, get the same answer. We call that a fingerprint, and each row carries one. Evidence you can re-derive beats evidence you have to trust.
One click exports the whole history as JSON or CSV. That file is the artifact your review actually wants: proof of what agent output was allowed to land, under which rules, with receipts.
Deterministic engines make honest ledgers possible. Merge gates built on LLM judgment can log what the model said, but they can't promise to say it again. Ours can. Same input, same output, provable on demand.
The audit trail ships on every tier, including free. Retention is the paid lever, not the truth.
Rosentic intercepts agent code output at the merge boundary, records a reproducible audit trail of every verdict, shows which agents write to which branches, and enforces policy over what lands. One engine, one ledger, four doors in.