Shunt submission hand-off — RouterArena (sub_10)

REPRODUCE.md

@@ -0,0 +1,43 @@
+# Shunt submission — RouterArena (sub_10) — how to reproduce
+
+Router **shunt** (@KookaS, https://github.com/KookaS/shunt, open-source) is a thin
+adapter over Shunt's real router engine: every prompt is routed through
+`RouterEngine.decide` under Shunt's shipped `session_cascade` strategy over
+Shunt's live pool (`deepseek-v4-flash`, `deepseek-v4-pro`). RouterArena's
+stateless single-turn / no-verification regime provides no verified-outcome
+channel, so Shunt's verified-failure escalation ladder never fires and the
+router deterministically picks its cheapest live model
+(`deepseek/deepseek-v4-flash`).
+
+## Pin
+
+The router code was generated against Shunt at this commit (Shunt is not on
+PyPI — install from git):
+
+```
+SHUNT_PIN = 7ee66af5ddcbd78886a58e5336b21ffc5291da56
+pip install "git+https://github.com/KookaS/shunt@7ee66af5ddcbd78886a58e5336b21ffc5291da56"
+```
+
+or point `SHUNT_SRC` at a Shunt `src/` checkout of that commit.
+
+## Regenerate the routing decisions
+
+From this repository root (a Shunt install or `SHUNT_SRC` must be available):
+
+```
+uv run python router_inference/generate_prediction_file.py shunt sub_10 --no-optimality
+uv run python router_inference/generate_prediction_file.py shunt robustness --no-optimality
+```
+
+The recorded answers in `predictions/shunt.json` are the models' own outputs
+(provider generations are not byte-reproducible); the router CHOICES are
+deterministic. The two prediction files are required by RouterArena's workflow
+(exactly one base file + one `-robustness` file per PR).
+
+## Honest framing
+
+This is a model baseline produced by Shunt's real `session_cascade` machinery:
+without a verification channel the router stays on its cheapest live model.
+Training/tuning on RouterArena data was not used — the engine routes from Shunt's
+shipped configuration only.

router_inference/config/shunt.json

@@ -0,0 +1,17 @@
+{
+  "pipeline_params": {
+    "router_name": "shunt",
+    "router_cls_name": "ShuntSessionCascade",
+    "models": [
+      "deepseek/deepseek-v4-flash",
+      "deepseek/deepseek-v4-pro"
+    ],
+    "default_model": "deepseek/deepseek-v4-flash"
+  },
+  "description": "Shunt (https://github.com/KookaS/shunt, open-source) running its shipped `session_cascade` strategy over its live pool (deepseek/deepseek-v4-flash, deepseek/deepseek-v4-pro) via the real RouterEngine. RouterArena's stateless single-turn regime gives Shunt no verified-outcome channel (no repo to re-test, no confirmation step), so the verified-failure escalation ladder never fires and the router stays on its cheapest live model (deepseek/deepseek-v4-flash) — the measured QA outcome that the cheap rung dominates. Provider routing uses the native DeepSeek API for the menu models.",
+  "affiliation": "@KookaS",
+  "github": "https://github.com/KookaS/shunt",
+  "homepage": "https://github.com/KookaS/shunt",
+  "contact": "https://github.com/KookaS/shunt/issues",
+  "open_source": true
+}