From 716ba301eb5e7f490697a394b5c048c3fe1dc2bf Mon Sep 17 00:00:00 2001 From: Felmon Fekadu Date: Tue, 14 Jul 2026 11:41:31 -0600 Subject: [PATCH] docs: link the runnable refund-agent demo from the README felmonon/arh-demo-refund-agent shows the harness blocking a real double-refund regression in CI, with an intentionally-open PR as living proof. --- README.md | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/README.md b/README.md index 2a3d706..b888065 100644 --- a/README.md +++ b/README.md @@ -36,6 +36,15 @@ It may: Final-answer evals miss all of this. This harness evaluates the **trajectory**, and turns each of these concerns into a versioned policy rule with a stable ID. +## See it block a real regression + +[felmonon/arh-demo-refund-agent](https://github.com/felmonon/arh-demo-refund-agent) +is a complete runnable demo: a refund agent, a trace recorder, this harness as +an 11-line CI gate, and [an open pull request](https://github.com/felmonon/arh-demo-refund-agent/pull/1) +where a harmless-looking prompt edit makes the agent refund customers twice — +and the gate blocks the merge with `ARH-FLW-003` (duplicate side effect) while +the agent's final answer still looks perfect. + ## 60-second example ```bash