Add BARouter submission

[zulk2002]

BARouter: Buget Adaptive Router## BARouter Submission

This PR adds the BARouter prediction artifacts for RouterArena leaderboard evaluation.

Router

BARouter uses model pool:

• qwen3-235b-a22b-instruct-2507
• gpt-4.1
• deepseek-v3.1

The submission includes final routing decisions and generated outputs only; no router implementation code or local predictor artifacts are included.

Files

• router_inference/config/BARouter.json
• router_inference/predictions/BARouter.json
• router_inference/predictions/BARouter-robustness.json

Local Validation

I locally evaluated the full split and checked the submission files.

• Full split: 8,400 regular entries + 1,618 optimality entries
• Robustness split: 420 entries
• All full entries include generated_result, accuracy, cost, and usable token usage
• Abnormal entries: 0

Local full-split metrics:

• RouterArena Score: 0.6714
• Accuracy: 68.86%
• Total Cost: $5.2636
• Avg Cost per Query: $0.000627
• Avg Cost per 1K Queries: $0.6266
• Opt.Sel: 0.5862
• Opt.Cost: 0.5919
• Opt.Acc: 0.8793

[zulk2002]

/evaluate

[github-actions]

Router Evaluation Results

Router: BARouter
Dataset Split: full

RouterArena Metrics

Metric	Value
RouterArena Score	0.6709
Accuracy	68.80%
Total Cost	$5.263578
Avg Cost per Query	$0.000627
Avg Cost per 1K Queries	$0.6266
Number of Queries	8400
Abnormal Entries	0
Robustness Score	0.5238

Optimality Metrics

Metric	Value
Opt.Sel (Optimal Selection)	0.5856
Opt.Cost (Cost Efficiency)	0.5903
Opt.Acc (Accuracy vs Optimal)	0.8791

---

Evaluation completed by RouterArena automated workflow