From 79e575e9f3f32d096bfeafe94d072b6923a2ae64 Mon Sep 17 00:00:00 2001 From: moe-berri Date: Fri, 4 Sep 2026 20:14:33 -0700 Subject: [PATCH 1/2] docs(blog): add stall-escalation solve-rate result --- blog/autorouter_stall_escalation/index.md | 2 ++ 1 file changed, 2 insertions(+) diff --git a/blog/autorouter_stall_escalation/index.md b/blog/autorouter_stall_escalation/index.md index d82390a69..5ed19e956 100644 --- a/blog/autorouter_stall_escalation/index.md +++ b/blog/autorouter_stall_escalation/index.md @@ -31,6 +31,8 @@ A cheap model can get stuck mid-task: it calls the same tool with the same argum `stall_escalation_enabled: true` gives the router that judgment on its own. +On a 24-task test set escalating `gpt-4.1-nano` to `gpt-5.4-mini`, stall escalation achieved 52.8% solve rate vs 38.9% without it—a 14 percentage point jump. Both models are modest, but the improvement is real. + ## How it decides - **Reads the assistant's own tool calls, not the human's messages.** A task counts as stalled once the newest call repeats, or errors, at least `stall_escalation_repeat_threshold` times across the last `stall_escalation_window` calls From d68bfc1e4d861824250d8758843d658ddc31f307 Mon Sep 17 00:00:00 2001 From: moe-berri Date: Sat, 5 Sep 2026 10:35:27 -0700 Subject: [PATCH 2/2] Update index.md --- blog/autorouter_stall_escalation/index.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/blog/autorouter_stall_escalation/index.md b/blog/autorouter_stall_escalation/index.md index 5ed19e956..28f550e9a 100644 --- a/blog/autorouter_stall_escalation/index.md +++ b/blog/autorouter_stall_escalation/index.md @@ -31,7 +31,7 @@ A cheap model can get stuck mid-task: it calls the same tool with the same argum `stall_escalation_enabled: true` gives the router that judgment on its own. -On a 24-task test set escalating `gpt-4.1-nano` to `gpt-5.4-mini`, stall escalation achieved 52.8% solve rate vs 38.9% without it—a 14 percentage point jump. Both models are modest, but the improvement is real. +On a 24-task test set escalating `gpt-4.1-nano` to `gpt-5.4-mini`, stall escalation achieved 52.8% solve rate vs 38.9% without it, a 14 percentage point jump. Both models are not that powerful, but the improvement is good proof that it works. ## How it decides