Add Lynkr router submission

[vishalveerareddy123]

Router: Lynkr

Lynkr is an open-source (Apache-2.0), self-hosted LLM gateway that routes requests across local and cloud models. Routing here uses its embedding-anchor tier classifier (no LLM classifier call — routing decisions are local and deterministic).

• Repo: https://github.com/Fast-Editor/Lynkr
• Model pool (all served via OpenRouter): openai/gpt-oss-120b, qwen/qwen3-235b-a22b-2507, z-ai/glm-4.7

Files

• router_inference/config/lynkr.json — router configuration
• router_inference/predictions/lynkr.json — full-split predictions with generated_result populated
• router_inference/predictions/lynkr-robustness.json — robustness predictions
• router_inference/router/lynkr_router.py + registration in __init__.py — adapter for reproducibility
• Additive entries in model_cost/model_cost.json (z-ai/glm-4.7 pricing) and universal_model_names.py for the pool models

Local evaluation results (for reference; please re-run officially)

Arena 67.65 · Accuracy 68.41% · $0.29/1K queries · Robustness 92.38 · Opt.Sel 11.13 / Opt.Cost 17.00 / Opt.Acc 85.37

Disclosures

• Optimality metrics were computed on 629/809 of the optimality queries locally — our OpenRouter credits ran out on the GLM tail. The prediction files are complete; your official evaluation should produce full coverage.
• Routing is heavily concentrated on gpt-oss-120b (~96% of queries) — a consequence of the tier boundaries, not tuning on RouterArena data.
• Per the evaluation-only policy: no Lynkr component was trained, fitted, or tuned on RouterArena data or labels. The anchor classifier was built entirely from Lynkr's own telemetry before this evaluation; RouterArena was run once as a held-out test.

Happy to adjust anything to match submission conventions.

---

Leaderboard row details (for the manifest/README, to save you a lookup)

• Router name: Lynkr
• Affiliation: 👤 @vishalveerareddy123
• Links: Code · npm
• Open-source: yes (self-hosted gateway; routing decision is local, no LLM call)
• Prediction files: lynkr / lynkr-robustness (this PR)

[vishalveerareddy123]

/evaluate

[vishalveerareddy123]

/evaluate

[github-actions]

Router Evaluation Results

Router: lynkr
Dataset Split: full

RouterArena Metrics

Metric	Value
RouterArena Score	0.6761
Accuracy	68.35%
Total Cost	$2.424762
Avg Cost per Query	$0.000289
Avg Cost per 1K Queries	$0.2887
Number of Queries	8400
Abnormal Entries	6
Robustness Score	0.9238

⚠️ 6 of 8400 queries (0.1%) 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.

Optimality Metrics

Metric	Value
Opt.Sel (Optimal Selection)	0.1116
Opt.Cost (Cost Efficiency)	0.1700
Opt.Acc (Accuracy vs Optimal)	0.8565

---

Evaluation completed by RouterArena automated workflow

[vishalveerareddy123]

/evaluate

[vishalveerareddy123]

/evaluate

[github-actions]

Router Evaluation Results

Router: lynkr
Dataset Split: full

RouterArena Metrics

Metric	Value
RouterArena Score	0.6765
Accuracy	68.41%
Total Cost	$2.456236
Avg Cost per Query	$0.000292
Avg Cost per 1K Queries	$0.2924
Number of Queries	8400
Abnormal Entries	0
Robustness Score	0.9238

Optimality Metrics

Metric	Value
Opt.Sel (Optimal Selection)	0.1097
Opt.Cost (Cost Efficiency)	0.1608
Opt.Acc (Accuracy vs Optimal)	0.8448

---

Evaluation completed by RouterArena automated workflow

[vishalveerareddy123]

@yl231 The full-split evaluation is now clean — 0 abnormal entries (down from 6 on the previous run; root cause was the truncated entries in the regenerated prediction file). Final numbers from the run above: Arena 0.6765, Accuracy 68.41%, $0.2924/1K queries, Robustness 0.9238, 8,400/8,400 scored, all CI checks green.

The submitted config is Lynkr's untuned production default, per the evaluation-only policy. Ready for review whenever you have a moment — happy to adjust anything needed for the leaderboard entry.

[yl231]

Maintainer review — approving. Structural audit clean (8,400 base + 420 robustness + 1,618 optimality rows; genuine per-row generations; costs reconcile against the registered per-token prices). The added z-ai/glm-4.7 pricing matches the existing underscore variant exactly. lynkr_router.py is a thin gateway adapter — no RouterArena data access. Official /evaluate: Accuracy 68.41% · Arena 0.6765 · $0.2924 per 1K · Robustness 0.9238. All checks green. Thanks @vishalveerareddy123 — merging; leaderboard row to follow.