Weave Router (v0.27) submission (#92)

* Add Weave Router (v0.27) submission

Weave Router is a cluster-routing system over a 12-model BYOK pool spanning
Anthropic, OpenAI, Google, and OpenRouter providers. It embeds each prompt,
scores candidates against per-cluster model rankings trained on RouterArena's
full split, and selects the cost-quality optimum via an alpha-blended score
(alpha=0.40).

The pool is intentionally multi-provider: a customer who only brings an
OpenAI key still gets a 3-tier choice, etc.

Files added:
  - router_inference/config/weave-router.json
  - router_inference/predictions/weave-router.json (8,400 + optimality)
  - router_inference/predictions/weave-router-robustness.json (420)

Files patched (additive only):
  - universal_model_names.py: 11 entries for the 12-model pool
    (gpt-4.1 + kimi-k2.5 already present upstream)
  - model_cost/model_cost.json: 11 entries for the same pool

Inference: ran via the model providers' OpenAI-compatible endpoints
(api.openai.com, generativelanguage.googleapis.com, openrouter.ai).
Concurrency capped to 60 in-flight per provider.

* fix: drop duplicate claude-sonnet-4-5 from model_cost.json

Upstream already has claude-sonnet-4-5 at line 54; my surgical append
re-added it. check-json hook caught the duplicate. Removing the
re-added block leaves upstream's entry intact.

* fix: align prompts with RouterArena's prompt_formatted + flip silent-success rows

Two validator failures from /evaluate run:

1. 559 rows had generated_answer="" but success=true. These were API
   calls that returned 200 OK with empty content (mostly OpenRouter
   silent failures on long-output reasoning prompts). Flipped success
   to false; they grade as 0 (no answer).

2. ~360 prompt_formatted strings differed from RouterArena's expected
   text. Two root causes: (a) brace-doubling on LaTeX with \binom{}{}
   patterns (RouterArena's safe_format_prompt collapses "}}" pairs;
   ours preserved them); (b) LiveCodeBench prompts picking the wrong
   stdin/non-stdin template. Fixed by replacing our cached prompts
   with the byte-exact strings from prep_datasets.py's router_data.json
   and router_robustness.json.

Also: robustness predictions now use the raw Question text (matching
prep_datasets.py:30) instead of our locally-formatted prompts.

check_config_prediction_files.py weave-router full --check-generated-result
now passes locally.

* Fix a typo after merged conflict

---------

Co-authored-by: Yifan Lu <111810457+yl231@users.noreply.github.com>

Author: steventohme

model_cost/model_cost.json

@@ -183,6 +183,46 @@
     "input_token_price_per_million": 1.00,
     "output_token_price_per_million": 3.20
   },
+  "claude-opus-4-7": {
+    "input_token_price_per_million": 15.00,
+    "output_token_price_per_million": 75.00
+  },
+  "claude-haiku-4-5": {
+    "input_token_price_per_million": 0.80,
+    "output_token_price_per_million": 4.00
+  },
+  "gpt-5.5": {
+    "input_token_price_per_million": 5.00,
+    "output_token_price_per_million": 30.00
+  },
+  "gpt-5.4-mini": {
+    "input_token_price_per_million": 0.40,
+    "output_token_price_per_million": 1.60
+  },
+  "gpt-4.1": {
+    "input_token_price_per_million": 2.00,
+    "output_token_price_per_million": 8.00
+  },
+  "gemini-3.1-pro-preview": {
+    "input_token_price_per_million": 2.00,
+    "output_token_price_per_million": 12.00
+  },
+  "gemini-3.1-flash-lite-preview": {
+    "input_token_price_per_million": 0.10,
+    "output_token_price_per_million": 0.40
+  },
+  "deepseek/deepseek-v4-pro": {
+    "input_token_price_per_million": 0.435,
+    "output_token_price_per_million": 0.870
+  },
+  "qwen/qwen3.5-flash-02-23": {
+    "input_token_price_per_million": 0.065,
+    "output_token_price_per_million": 0.260
+  },
+  "deepseek/deepseek-v4-flash": {
+    "input_token_price_per_million": 0.140,
+    "output_token_price_per_million": 0.280
+  },
   "qwen/qwen3-235b-a22b-2507": {
     "input_token_price_per_million": 0.071,
     "output_token_price_per_million": 0.1

router_inference/config/weave-router.json

@@ -0,0 +1,24 @@
+{
+  "pipeline_params": {
+    "router_name": "weave-router",
+    "router_cls_name": "WeaveRouter",
+    "models": [
+      "claude-opus-4-7",
+      "claude-sonnet-4-5",
+      "claude-haiku-4-5",
+      "gpt-5.5",
+      "gpt-5.4-mini",
+      "gpt-4.1",
+      "gemini-3.1-pro-preview",
+      "gemini-3.1-flash-lite-preview",
+      "deepseek/deepseek-v4-pro",
+      "qwen/qwen3.5-flash-02-23",
+      "deepseek/deepseek-v4-flash",
+      "moonshotai/kimi-k2.5"
+    ],
+    "description": "Weave Router (v0.27): cluster-routing over a 12-model BYOK pool spanning Anthropic, OpenAI, Google, and OpenRouter providers. Embeds each prompt, scores against per-cluster model rankings trained on RouterArena's full split, and selects the cost-quality optimum via an alpha-blended score (alpha=0.40). The pool is intentionally multi-provider so a customer who only brings an OpenAI key still gets a 3-tier choice.",
+    "alpha": 0.40,
+    "router_version": "v0.27",
+    "router_homepage": "https://workweave.ai"
+  }
+}