feat(router): submit NuRoute to the leaderboard

[vibre06]

NuRoute (https://nuroute.ai), a provider-agnostic inference control plane.
Measured in production on the full 8,400-query split. Every routing decision
and every answer came from a real production request.

Metric	Value
Arena Score	69.30
Accuracy	76.51%
Cost / 1K	$4.36
Robustness	70.48
Abnormal entries	99 (1.2%)

Model pool (7) — all served by the live router, none pinned:
moonshotai/Kimi-K2.6 (75.9%), MiniMax-M3 (15.2%), claude-sonnet-5 (7.0%),
ministral-8b-latest (0.9%), gpt-5.5 (0.9%), o4-mini, qwen3-30b-a3b-instruct-2507.

Catalog additions. moonshotai/Kimi-K2.6, claude-sonnet-5 and
ministral-8b-latest were not in universal_model_names.py; added there with
provider list prices in model_cost/model_cost.json. Dated variants the API
returns (gpt-5.5-2026-04-23, o4-mini-2025-04-16) are mapped to their
canonical names.

Disclosures

• Generated at max_tokens=8192 rather than the pipeline default of 2048. At
  2048 the served reasoning models exhausted the output budget mid-reasoning and
  returned empty content on ~18% of prompts, billed in full. Measured p50 to
  recover them was ~4,400 output tokens.
• Optimality entries omitted (--no-optimality), so Opt.Sel / Opt.Cost /
  Opt.Acc will show as not reported.

No training, fitting or tuning was done on RouterArena data at any point.

[vibre06]

/evaluate

[github-actions]

Router Evaluation Results

Router: nuroute
Dataset Split: full

RouterArena Metrics

Metric	Value
RouterArena Score	0.6930
Accuracy	76.51%
Total Cost	$36.593006
Avg Cost per Query	$0.004356
Avg Cost per 1K Queries	$4.3563
Number of Queries	8400
Abnormal Entries	99
Robustness Score	0.7048

⚠️ 99 of 8400 queries (1.2%) had no valid generation (inference failed / empty answer) and were scored as incorrect (0). These queries still count toward the denominator, so accuracy and cost reflect the full query set. Please regenerate predictions for these queries and resubmit for a complete evaluation.

---

Evaluation completed by RouterArena automated workflow