From 3d6e33d154951622e5969e95633fcff737a32fd0 Mon Sep 17 00:00:00 2001
From: Lin Jia <linj@nvidia.com>
Date: Wed, 19 Aug 2026 15:35:52 -0700
Subject: [PATCH] docs(benchmark): add v0.2.0 TB-2.1 escalation routing profile

Signed-off-by: Lin Jia <linj@nvidia.com>
---
 .../tb21-escalation-opus-glm-deepseek.toml    | 57 +++++++++++++++++++
 1 file changed, 57 insertions(+)
 create mode 100644 benchmark/routing-profiles/tb21-escalation-opus-glm-deepseek.toml

diff --git a/benchmark/routing-profiles/tb21-escalation-opus-glm-deepseek.toml b/benchmark/routing-profiles/tb21-escalation-opus-glm-deepseek.toml
new file mode 100644
index 000000000..c2c13a0dd
--- /dev/null
+++ b/benchmark/routing-profiles/tb21-escalation-opus-glm-deepseek.toml
@@ -0,0 +1,57 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Escalation-router deployment behind the v0.2.0 Terminal-Bench 2.1 efficiency results
+# (https://developer.nvidia.com/blog/route-ai-agent-workloads-across-models-with-nvidia-nemo-switchyard/).
+#
+# This is the configuration as benchmarked, with one change: the published runs were
+# served through NVIDIA-internal inference endpoints, replaced here with OpenRouter
+# equivalents so the deployment is publicly runnable. Routing-algorithm parameters
+# are exactly as run; absolute solve rates may shift slightly across serving stacks.
+# As run, the strong tier used Anthropic adaptive thinking with output effort `high`,
+# the weak tier ran with reasoning effort `high`, and the judge ran with thinking
+# disabled — mirrored below with OpenRouter's `reasoning` controls.
+#
+# The `escalation` block is part of the v0.2.0 server config schema only; it is not
+# accepted by the current `switchyard-server`. Run from the tag:
+#   git checkout v0.2.0-rc.1
+#   switchyard-server --config benchmark/routing-profiles/tb21-escalation-opus-glm-deepseek.toml --port 4000
+
+schema_version = 1
+
+[llm_clients.openrouter]
+format = "openai_chat"
+base_url = "https://openrouter.ai/api/v1"
+api_key_env = "OPENROUTER_API_KEY"
+
+[targets.strong]
+id = "anthropic/claude-opus-4.8"
+llm_client = "openrouter"
+
+[targets.strong.extra_body.reasoning]
+effort = "high"
+
+[targets.weak]
+id = "z-ai/glm-5.2"
+llm_client = "openrouter"
+
+[targets.weak.extra_body.reasoning]
+effort = "high"
+
+[targets.judge]
+id = "deepseek/deepseek-v4-flash"
+llm_client = "openrouter"
+
+[targets.judge.extra_body.reasoning]
+enabled = false
+
+# Every conversation starts on the weak tier. The judge reviews the trajectory each
+# turn, and two consecutive escalate verdicts latch the session onto the strong tier.
+[routes.switchyard]
+id = "switchyard"
+type = "llm_classifier"
+classifier_target = "judge"
+strong_target = "strong"
+weak_target = "weak"
+base_threshold = 0.5
+escalation = { confirmations = 2, recent_turn_window = 28, window_message_chars = 500 }