krush420/sovereign-g4-benchmarks

Sovereign G4 Hermetic Benchmark Audit Logs & Scorecards (krusch-stack — 325,000+ Queries/Traces, 12 External Suites, 36 Audit Logs)

Core Thesis — "Databases are the memory state and models are ephemeral compute."
Read Paper #27 (Return on Memory: #1 Overall (91.17%) & #1 Open-Weights (54.17% @ $0.00) on Official DolphinBench), Paper #26 (DolphinBench Sovereign Memory & Wave-4 Ablation), and the cross-stack report at krusch.dev/articles/how-krusch-stack-stacks-up.html, or explore interactive Pareto frontiers at huggingface.co/spaces/krush420/sovereign-router-arena.

TL;DR (Scope-Qualified): Across 325,000+ benchmark queries/traces (8,400 official openai/gpt-5.6-sol DolphinBench traces across 14 runs & 5 model families), multi-turn security episodes, 2.39M-document code corpora, and 178M-token conversational histories spanning 12 public benchmarks (36 raw JSON/CI audit logs), KruschStack & Krusch Cascade Router deliver:

1. #1 Overall 91.17% (547 / 600 @ $24.54, 35.38s) + #1 Open-Weights 54.17% (325 / 600 @ $0.00) on Official DolphinBench (openai/gpt-5.6-sol, reasoning_effort="medium"): Pairing our byte-identical $0.00 Zero-LLM 0o444 SovereignMemoryStore SQLite files across 14 runs (8,400 Official gpt-5.6-sol traces, 5 model families) with frontier Gemini 4 Argon (Low) reaches **91.17% (547/600 — Morgan 191, Alex 179, Riley 177, 92.22% mechanical [1,221/1,324], 93.62% semantic judge [558/596], $24.54, 35.38s median)**—surpassing #1 Hermes + GPT 5.6 Luna + Mem0 (70.67% / 424/600 @ $96.21, 37.69s) by +20.50 pp (+123 tasks) at ~3.9x lower cost (14.1x cheaper than Supermemory 59.17% @ $345.64, and 89.17% [535/600 @ $24.57] on Argon Medium). A controlled 2×2 Argon Thinking × Search factorial (78.00% / 79.17% Turn-0 Briefing-Only at $15.76 / $16.91 vs. 91.17% / 89.17% Full Agentic Search) uncovers a statistically significant ΔΔ = +3.17 pp (p = 0.0320, paired McNemar 29 vs. 17) thinking-to-search compensation interaction, a 96.17% (577/600) Oracle@14 retrieval-sufficiency lower bound (78.83% [473/600] Oracle@10 open-weights), a 4.19x superlinear LBF=4+ multiplier (89.33% vs. 21.33% 30B), and flatter corpus-entropy drop (-7.00 pp Argon Low / -3.50 pp Argon Med vs. -10.00 pp Mem0 to -32.50 pp MiniMax M3 + Hindsight). On the exact same $0.00 0o444 SQLite stores, **Google Gemma 4 31B (Thinking ON) reaches 54.17% (325/600 @ $0.00, #1 Open-Weights; 52.00% Thinking OFF; 61.50% Oracle@2)**—beating all 8 non-Luna commercial cloud baselines up to $1,830.57 (MiniMax M3 + Mem0 47.83% @ $148.02, Claude Sonnet 5 + Honcho 35.83% @ $1,565.82, Sonnet 5 + Mem0 32.33% @ $1,830.57) and outperforming dense Qwen2.5-Coder-32B (35.33%) by +18.83 pp (p = 1.35e-16) alongside 30B Wave-4 (42.67%), 120B Wave-4 (39.50%), 80B Stock (38.50%), Hermes-4-14B (34.50%), and Hermes-3-8B Wave-4 (18.33%, 15.7x lift). All 14 / 14 archives pass dolphinbench.submission.validate (7 queued on dolphinbench.ai, including Argon Low UUID c047b063-822c-49fb-8b36-27c3763d6ca7 & Argon Med UUID 29389157-b01b-42b5-8193-17d9a496f580, plus 7 clean-room ablations);
2. 0.00% Indirect Prompt-Injection ASR (0 / 949 attacks) across 100% of ETH SpyLab AgentDojo v1.2.2 (1,046 episodes across workspace, banking [52.78% -> 0.00% ASR], slack [57.14% -> 0.00% ASR], and travel), lifting Benign Utility 61.86% -> 69.07% on <UNTRUSTED_CONTENT> Spotlighting and preserving 61.86% (0.00 pp loss) + 83 hardline capability blocks on Sovereign Defense;
3. #1 Full-Corpus Code Retrieval (56.40% 10-Task Macro NDCG@10, 73.88% Recall@10) across 100% of ACL 2025 CoIR (10 / 10 tasks, 20 / 20 unshrunk sub-corpora, 2,389,123 docs, 162,216 queries) using krusch-git Hybrid RRF + snowflake-arctic-embed2:latest (1024d, 8,192 ctx)—exceeding all 10 published Table 3 baselines (Voyage-Code-002 56.26%, E5-Mistral 7B 55.18%; explicit 8K vs. 512-tok context caveat below);
4. Database State Grounding & Deferred MCP Compression: 26.00% -> 40.20% EX (1.55x, 100% valid SQL) across all 500 / 500 BIRD-SQL Mini-Dev dirty SQLite queries (11 DBs) via read-only node:sqlite PRAGMA/cell-sample grounding + 2PC Verify-on-Stop, and 59.88% -> 66.28% Micro (+144 Qs) / 84.69% Irrelevance (+17.46 pp) at -80.67% fewer schema tokens across all 2,251 / 2,251 UC Berkeley BFCL v3 Live (6/6) queries in a 64-tool Enterprise MCP Hub;
5. State-of-the-Art Cost-Accuracy Routing: 0.7883 (#1 in CI, PR #169 open) on RouteWorks/RouterArena (8,400 Qs), #1 ID & OOD Micro AIQ breaking Martian's ZeroRouter OOD ceiling on 72,980-row RouterBench (-72.63% cost @ 98% GPT-4), and #1 Pareto Distance (0.4269, -83.95% cost) on ACL 2026 LLMRouterBench (12,566 Qs);
6. Transactional Coding, Cross-File AST & Legal Span RAG: 19.56% -> 51.56% Pass@2 (2.64x, 72 healed) on Aider Polyglot (225/225) tying Claude 3.5 Sonnet (51.6%) at $0 API cost, 19.12% Macro Acc@1 / 68.58% Acc@5 on RepoBench-R (47,998 Qs), 26.74% Macro F1 / 33.95% MAUD Recall@2 on LegalBench-RAG (6,889 Qs), and 92.58% (100K) / 93.00% (10M IE) recall on ICLR 2026 BEAM (2,000/2,000 Qs).

---

📊 Sovereign External Benchmark Scorecard (68 / 68 Gates + 14 / 14 Validated DolphinBench Archives)

Public Benchmark	Evaluated Sovereign Subsystem	Scale & Grading Discipline	Verified Result & External Leaderboard Comparison
Official DolphinBench (600/600) Official openai/gpt-5.6-sol Judge	SovereignMemoryStore (256/32 Raw-RAG + 4-Way RRF, $0.00 Ingest) + Gemini 4 Argon (Low/Med), Gemma 4 31B & Wave-4 QLoRA+DPO (5 Model Families: 30B, 120B, 80B, 32B, 14B, 8B)	600 / 600 tasks × 14 runs = 8,400 traces (13,539 sessions / 1.50M tokens; 14/14 submission.validate passed: 7 queued on dolphinbench.ai + 7 clean-room ablations)	#1 Overall 91.17% (547/600) Gemini 4 Argon Low @ $24.54 (+20.50 pp vs. #1 Mem0 70.67% @ $96.21, ~3.9x cheaper; 89.17% Argon Med; 96.17% Oracle@14, 78.83% Oracle@10 Open-Weights, 4.19x LBF=4+, +3.17 pp 2×2 interaction p=0.032) · #1 Open-Weights 54.17% (325) Gemma 4 31B @ $0.00 (52.00% No-Think; beats all 8 non-Luna cloud stacks) · 42.67% 30B Wave-4 · 39.50% 120B Wave-4 · 35.33% Qwen2.5-32B · 34.50% Hermes-4-14B · 18.33% 8B Wave-4 (15.7x).
ETH SpyLab AgentDojo v1.2.2 100% Full 4 Suites	delimitUntrustedContent + Sink Taint Provenance + createKruschPolicyFloor (gpt-oss:120b)	1,046 episodes (97 benign + 949 important_instructions attacks = 3,138 live multi-turn sessions; 0 LLM judges)	0.00% Targeted ASR (0/949, 100.0% Defense Rate) vs. 19.28% raw ASR (57.14% slack, 52.78% banking) · 69.07% Benign / 69.02% UUA (Spotlighting) & 61.86% Benign (0.00 pp loss) + 83 floor blocks (Sovereign).
ACL 2025 CoIR 100% Full 10/10 Tasks	krusch-git Hybrid RRF + AST Spine + snowflake-arctic-embed2 (1024d, 8192 ctx)	2,389,123 unshrunk docs & 162,216 test queries across all 20 / 20 sub-corpora (0 LLM judges)	56.40% 10-Task Macro NDCG@10 (#1) · 50.94% MRR@10 (#1) · 73.88% Recall@10 (#1); exceeds BM25 (48.98%), pure dense (55.65%), & all 10 CoIR Table 3 baselines (Voyage-Code-002 56.26%, E5-Mistral 7B 55.18%).
BIRD-SQL Mini-Dev 100% Full SQLite (11 DBs)	Read-Only node:sqlite PRAGMA/Cell Grounding + createKruschPolicyFloor + 2PC VerifyOnStop (gpt-oss:120b)	500 / 500 dirty SQLite queries (460 challenging, 32 moderate, 8 simple; 0 LLM judges)	26.00% (130/500) → 40.20% EX (201/500, 1.55x) · 35.75% R-VES · 100.00% valid SQL; 13 queries self-healed across 45 retries (0.09 extra calls/query; 5.0x on toxicology).
UC Berkeley BFCL v3 Live 100% Full 6/6 Splits	createDeferredToolBridge + krusch-stack 64-Tool Hub + VerifyOnStop Schema Guard (gpt-oss:120b)	2,251 / 2,251 queries (simple, multiple, parallel, parallel_multiple, irrelevance, relevance; 0 LLM judges)	59.88% → 66.28% Micro (+144 Qs) · 41.00% → 48.72% 6-Split Macro · 84.69% Irrelevance (+17.46 pp) · 29 schema repairs healed · -80.67% schema tokens (7,894 → 1,526 tok/turn).
Martian RouterBench 0-Shot & 5-Shot ID + OOD	krusch-cascade-router (KruschCascadeRouter & KruschZeroShotRouter)	36,497 (0-shot) + 36,483 (5-shot) prompts × 11 LLMs across 9 suites (36/36 zero-leak splits; 0 LLM judges)	#1 ID Micro AIQ (0.7488 0-shot, 0.7695 5-shot) & #1 Pareto (0.2531 / 0.2039) · -72.63% cost @ 98% GPT-4 ($9.89 vs. $36.15); breaks Martian's OOD ZeroRouter ceiling across 0-shot (0.7268) & 5-shot (0.7455).
ACL 2026 LLMRouterBench 13-Model ID + Unseen OOD	krusch-cascade-router (Dual L1 + Arctic L2)	12,566 prompts × 13 frontier models (8 ID n=3,005 + 2 unseen OOD HLE & ARC-AGI n=2,558; 0 LLM judges)	#1 Lowest Pareto Distance (0.4269) across all 10 policies · 42.78% Micro / 57.78% Macro Acc at $14.39/1K (-83.95% cost vs. best_single $89.62/1K); 25.25% unseen ARC-AGI vs. 8.75% RouteLLM.
RouterArena Live Cascade Live Generation + SSE Abort	CascadeRouter (7B -> 120B SSE Logprob/Loop Abort + Path A L1/L2)	809 sub_10 live generations + 420 robustness noise queries (75.2 GB 4-model VRAM; 325 aborts billed)	Lifts 7B accuracy +11.93 pp (48.36% -> 60.29%, 0.6111 Arena, 78.99% OptAcc) while cutting 120B TTFT -48.0% (1,694.5ms -> 880.7ms) and cost -37.4% (-60.2% on Path A $0.0646/1K).
Agentic Polyglot Routing Multi-Step Trajectory Loop	CascadeRouter + KruschFailureClassifier + TruncationGuard + VerifyOnStop	60 6-language Exercism suites × 4 multi-step policies = 240 sandboxed compile/test trajectories	Pass@2 = 51.67% (31/60, +20 healed) — 100.0% parity with static_heavy_120b and +30.00 pp (2.38x) over llm_router_preflight (21.67%) at 0.395ms overhead (1,442x faster).
RouteWorks RouterArena Official GitHub CI (PR #169 Open)	krusch-pre-router (<15µs) + krusch-cascade-router	8,400 queries + 420 robustness perturbations (Official github-actions[bot] PR #169, pending merge)	0.7883 Workflow Score (#1 in CI) · 80.72% accuracy at $0.2117 / 1K queries (90.48% robustness, 0 abnormal); surpasses Paix2 (0.7763), KT-ModelRouter (0.7628), Martian ($4.83), & NotDiamond ($4.10).
Aider Polyglot (225/225) 6-Lang Coding	krusch-coding-harness 2PC + open-weights openai/gpt-oss-120b	225 / 225 Exercism suites across C++, Go, Java, JS, Python & Rust (0 LLM judges)	Pass@1 = 19.56% (44/225) → Pass@2 = 51.56% (116/225) (2.64x lift, 72 failed tasks self-healed on 1 Verify-on-Stop turn)—ties cloud Claude 3.5 Sonnet (51.6%) & beats GPT-4o (23.1%) at $0 API cost.
RepoBench-R Cross-File Code IR	krusch-git (v1.2.3) + snowflake-arctic-embed2 (1024d)	47,998 Python & Java cross-file queries across all 8 cff/cfr × easy/hard splits (0 LLM judges)	19.12% 8-Split Macro Acc@1 (+21.8% vs. pure dense & +61.5% vs. import-order), 68.58% Acc@5 (sweeps 8/8 splits), and 31.33% Py / 27.45% Java cfr Easy—beating CodeBERT & UniXcoder.
LegalBench-RAG Character-Exact Span IR	KruschNexus (v0.2.6) + 8K Arctic + mxbai-rerank-large-v2	714 legal docs · 6,889 full / 776 mini queries · sort_by_document: false · char-exact overlap	26.74% Full Macro F1 (k=1) & 25.81% F1 / 44.72% Recall (k=2) (47.71% R at k=6), +128.5% (2.29x) F1 vs. fixed dense & 33.95% MAUD Recall@2 > OpenAI 3-large + Cohere Recall@64 (28.28%).
ICLR 2026 BEAM 100K–10M Memory	krusch-context-mcp (v2.0.0) + 120B Stock & gpt-oss-120b-sre-lora	100 chats (~178.34M tokens), 2,000 / 2,000 official probes across 100K, 500K, 1M, 10M	92.58% (100K) & 93.00% (10M IE) recall; zero E2E degradation from 100K → 1M (72.86%); beats cloud past.dev (gpt-5.6-luna) on Contradiction Resolution (4/4 splits) & 10M MSR.

---

⚖️ Honest Gaps, Methodology Caveats & Dataset Hygiene Disclosures

• DolphinBench Return-on-Memory 14-Run Decomposition & Queue Status (91.17% Argon Low & 54.17% Gemma 4 31B on Byte-Identical 0o444 SQLite Stores): On Official DolphinBench (600 tasks, openai/gpt-5.6-sol reasoning_effort="medium", 8,400 traces across 5 model families), freezing the $0.00 Zero-LLM SovereignMemoryStore SQLite files (0o444) and swapping only the ephemeral reader isolates retrieval from reasoning (Paper #27): Oracle@14 reaches 96.17% (577/600) (78.83% [473/600] Oracle@10 across $0.00 open-weights), Gemini 4 Argon (Low) reaches 91.17% (547/600 @ $24.54, 35.38s, #1 Overall, +20.50 pp / +123 tasks above #1 Mem0 70.67% @ $96.21, 37.69s; 89.17% Argon Medium @ $24.57) with a 2×2 Thinking × Search interaction of ΔΔ = +3.17 pp (p = 0.0320, 78.00%/79.17% Briefing-Only vs. 91.17%/89.17% Full Search) and a 4.19x superlinear multiplier at LBF=4+ (89.33% vs. 21.33% 30B). Zero-cloud open-weights models are led by **Google Gemma 4 31B at 54.17% (325/600 @ $0.00, #1 Open-Weights, 52.00% Thinking OFF, 61.50% Oracle@2)**—beating all 8 non-Luna commercial cloud baselines (MiniMax M3 + Mem0 47.83% @ $148.02, Sonnet 5 + Mem0 32.33% @ $1,830.57) and outperforming dense Qwen2.5-Coder-32B (35.33%) by +18.83 pp (p = 1.35e-16), alongside 30B Wave-4 (42.67%), 120B Wave-4 (39.50%), 80B Stock (38.50%), Hermes-4-14B (34.50%), and 8B Wave-4 (18.33%). Public dolphinbench.ai/leaderboard/results.json currently lists the 13 commercial cloud baselines while 7 sovereign submission archives (Argon Low UUID c047b063-822c-49fb-8b36-27c3763d6ca7 / SHA-256 80a199a60aef65f0..., Argon Med UUID 29389157-b01b-42b5-8193-17d9a496f580 / SHA-256 9a7fa41fb3d1cf5f..., plus 30b_wave4, 30b_stock, 120b_wave4, 120b_stock, and 8b_wave4) have status: "queued" awaiting maintainer merge and the remaining 7 clean-room ablations (Gemma 4 31B, Briefing-Only, 80B, 32B, 14B, all 14/14 passing dolphinbench.submission.validate) are held unsubmitted to avoid crowding the maintainer queue.
• ACL 2025 CoIR Context-Window Caveat (8,192 Native vs. 512-Token Table 3 Cap): ACL 2025 CoIR Table 3 truncated inputs to 512 tokens (Voyage-Code-002 56.26%, E5-Mistral 7B 55.18%), whereas snowflake-arctic-embed2:latest (568M Apache-2.0 params) runs at its native 8,192-token context window (56.40% Hybrid RRF / 55.65% Pure Dense), particularly benefiting long multi-turn code QA (codefeedback-mt 60.23%) and contest code translation (codetrans-contest 95.98% / 97.96%), while short synthetic-query tasks like apps (9.70% vs. 26.52%) and cosqa (23.76% vs. 32.76%) still favor proprietary code-pretrained models.
• Frontier Multi-Turn Coding & SQL Ceilings: On Aider Polyglot (51.56% Pass@2) and BIRD-SQL Mini-Dev (40.20% EX on gpt-oss:120b, up 1.55x from 26.00% 1-shot), single-retry local execution matches Claude 3.5 Sonnet (51.6% Polyglot) at $0 API cost, while multi-turn frontier reasoning pipelines (o3, gpt-5, Claude 4, Gemini 4 Argon) reach 75%–88%+ on Polyglot and 65%–75%+ on BIRD—validating our hybrid design (~85% local execution + ~15% L3 frontier escalation).
• Aggregate BEAM Reader Ceiling (70.6%–73.5% Cross-Family Referees vs. 86.2%–92.1% Cloud Self-Judging): Graded by two independent cross-family referees (qwen2.5:32b + hermes3:latest) with passive supersedes_id = NULL ingestion rather than same-model cloud self-judging.
• Open Upstream Leaderboard PRs & Dataset Hygiene:
  • RouteWorks/RouterArena PR #169 (0.7883 CI) is open/pending merge; 7 submission.validate-verified DolphinBench .zip archives are queued on dolphinbench.ai (Argon Low UUID c047b063-822c-49fb-8b36-27c3763d6ca7, Argon Med UUID 29389157-b01b-42b5-8193-17d9a496f580, 14/14 archives validated locally); withmartian/routerbench, ethz-spylab/agentdojo, and Aider upstream PR adapters are staged locally.
  • Upstream Dataset Bugs Documented & Deterministically Handled: (1) BFCL_v3_live_relevance.json contains a duplicate "id": "live_relevance_3-3-0" at lines 3 & 4 (disambiguated as live_relevance_3-3-0__dup2 so all 18/18 rows / 2,251/2,251 Live queries are graded); (2) RepoBench-R python_cff/test_easy rows [5522, 7591] have OOB gold_snippet_index (47,998/48,000 valid); (3) RouterBench 5-shot has 28 NaN rows (36,483/36,511 complete); (4) BEAM 100K Conv 14 KU Q1 label mismatch enforced as 0.00 canary.

---

Quickstart (datasets)

from datasets import load_dataset

# 1. Load top-level 16-entry sovereign scorecard (68/68 gates + 14/14 validated DolphinBench archives)
scorecard = load_dataset("krush420/sovereign-g4-benchmarks", "scorecard", split="test")

# 2. Official DolphinBench 600-task leaderboard split (14 Sovereign runs incl. #1 91.17% Argon Low & 54.17% Gemma 4 31B + 13 official baselines = 27 rows)
dolphinbench = load_dataset("krush420/sovereign-g4-benchmarks", "dolphinbench_600", split="test")

# 3. Wave-2 full splits: AgentDojo (1,046 eps), BIRD-SQL (500 Qs), BFCL Live (2,251 Qs), CoIR (10 tasks / 162,216 Qs)
agentdojo = load_dataset("krush420/sovereign-g4-benchmarks", "agentdojo_suites", split="test")
bird_sql = load_dataset("krush420/sovereign-g4-benchmarks", "bird_sql_splits", split="test")
bfcl = load_dataset("krush420/sovereign-g4-benchmarks", "bfcl_live_splits", split="test")
coir = load_dataset("krush420/sovereign-g4-benchmarks", "coir_10tasks", split="test")

# 4. Routing, Coding, AST & Memory splits
routerbench = load_dataset("krush420/sovereign-g4-benchmarks", "routerbench_regimes", split="test")
lrb = load_dataset("krush420/sovereign-g4-benchmarks", "llmrouterbench_splits", split="test")
polyglot = load_dataset("krush420/sovereign-g4-benchmarks", "polyglot_225", split="test")
repobench = load_dataset("krush420/sovereign-g4-benchmarks", "repobench_splits", split="test")
legalbench = load_dataset("krush420/sovereign-g4-benchmarks", "legalbench_domains", split="test")
beam = load_dataset("krush420/sovereign-g4-benchmarks", "beam_multisplit", split="test")

---

🔐 Cryptographic SHA-256 Audit Log Ledger (36 Audit Logs — raw_audit_logs/ + Official CI)

Raw Audit Log File / Artifact	Queries / Episodes	Verified SHA-256 Digest / Provenance
raw_audit_logs/dolphinbench_gemini_4_argon_low_gpt56_sol.json	600 (547/600 = 91.17%, #1 Overall)	77cf92f14ebacd27cfc8b8a9f14f63bfdec00bd0c0aa8aac5ac3d86611eb592a
raw_audit_logs/dolphinbench_gemini_4_argon_gpt56_sol.json	600 (535/600 = 89.17%, #2 Overall)	b803d5fa6a8068926e691409996bd95cd527bfd14682cfd2dacb6b3d7fe7364e
raw_audit_logs/dolphinbench_gemini_4_argon_briefing_only_gpt56_sol.json	600 (475/600 = 79.17%, Turn-0 Briefing)	de7fdad2d095e37fb01a2b4bfa73e0e00c491a9e5153022c5fc2a97fd309bcda
raw_audit_logs/dolphinbench_gemini_4_argon_low_briefing_only_gpt56_sol.json	600 (468/600 = 78.00%, Turn-0 Briefing)	43150fd01ac8014c0c87a6e908b92e9f593c2e920243a88616306d317af1a211
raw_audit_logs/dolphinbench_gemma4_31b_stock_gpt56_sol.json	600 (325/600 = 54.17%, #1 Open-Weights)	edad31041cb70d63def5af3a825b2474d7e472d93be354f01795ec5f3eca3a9f
raw_audit_logs/dolphinbench_gemma4_31b_nothink_stock_gpt56_sol.json	600 (312/600 = 52.00%, No-Think Ablation)	ff47b95e33bf66f1461a71b3ac7e998bebc13e80dfaca1a6766be8c57dd33e9e
raw_audit_logs/dolphinbench_30b_wave4_qlora_gpt56_sol.json	600 (256/600 = 42.67%)	c1f39f6ea1301039ccc33ae6eab704d9e8154d88e729a1bde8524404b9f8fd28
raw_audit_logs/dolphinbench_30b_stock_gpt56_sol.json	600 (252/600 = 42.00%)	5e1399cc5760400076b2772640a419a5af9e39b71ff27253f77fe5bcf74332f3
raw_audit_logs/dolphinbench_120b_wave4_qlora_gpt56_sol.json	600 (237/600 = 39.50%)	d17522a51cc65d4364b40bb2bbd4cc15cd67ffdf53b1227aec020f3cf80202a1
raw_audit_logs/dolphinbench_120b_stock_gpt56_sol.json	600 (235/600 = 39.17%)	0eece10654a51fbc8b0513cf38defaf4eb47fe01c5de5419bde9ea56ec47459a
raw_audit_logs/dolphinbench_80b_stock_gpt56_sol.json	600 (231/600 = 38.50%)	fcef6b9e8b939f8a0b18ccbc839eda8f3619b15662c2547e8f5b07a69652b6c5
raw_audit_logs/dolphinbench_qwen25_coder_32b_stock_gpt56_sol.json	600 (212/600 = 35.33%)	aeeb261f3c2cae7008943587fc44dc1f19b9689c168eb960867f410db74471bf
raw_audit_logs/dolphinbench_hermes4_14b_stock_gpt56_sol.json	600 (207/600 = 34.50%)	e30b04479f2b41bd028b39cea4beee5d3a60189b4e5039bcf1109fd7c7791f9d
raw_audit_logs/dolphinbench_hermes3_8b_wave4_qlora_gpt56_sol.json	600 (110/600 = 18.33%)	7759511a3a31a3a8470d0e180792d63d8f45776ffa236e88c76a0514dc0fe780
raw_audit_logs/agentdojo_audit_log.json	1,046 (3,138 sessions)	1b5e53051d44f98f7e621ae54f2249f5b07384291f1d71ef945f9d0053b4ad0f
raw_audit_logs/coir_audit_log.json	162,216 (2,389,123 docs)	4500936eeccfd5fb1cc07b8dde782a8381df758854afe0ae5f6b8a0c9c241cd2
raw_audit_logs/bird_sql_audit_log.json	500 (11 SQLite DBs)	0f660fea4ac00c045c111d4981c673701a8a63edbfa04d02636fd59459f24596
raw_audit_logs/bfcl_audit_log.json	2,251 (6/6 Live splits)	551ee4bc9554adf436f47c1f869b99d84e72b7c7dac8636d10df436afc9af3e5
raw_audit_logs/routerbench_audit_log.json	72,980 (0-shot + 5-shot)	007874e72e4126083ff0f1f8c5307d2dad0a295f31d9525d22583629d98e6b7e
raw_audit_logs/llmrouterbench_audit_log.json	12,566	73c85a9fbbb94163c18ac0612cbd501b0616278438c0b91437f18ce44a11ba4f
raw_audit_logs/live_cascade_audit_log.json	809 + 420	58405c428ef0284a17e2a2dbf238228ae0e49cfc70bcb1c8051b46c4a310fc92
raw_audit_logs/agentic_step_audit_log.json	240 (60 × 4)	377080ad66b9d68276b77ae5841ae4fbb808e40372e48c4c2875c2779f292b10
https://github.com/RouteWorks/RouterArena/pull/169	8,400 + 420	official_routeworks_ci_pr_169
raw_audit_logs/polyglot_stock_audit_log.json	225	c97515ad5edee3ae1fcd93445d3f47ec7c5769398e4eb4c4086ffd77ab2f3973
raw_audit_logs/repobench_r_audit_log.json	47,998	1122b545b4fa4d9bc87e8e42d2a3d694fa68834b1bef8d8eafbbac2e606ed4c0
raw_audit_logs/legalbench_nexus_audit_log.json	6,889	be1f515ca5cd68f2dc0f73258460a437a6b2c9bc1e017cf48d255701b8330628
raw_audit_logs/legalbench_rag_audit_log.json	6,889	bc87e9e4d0a3b9a7ac470bdd052d7e4fab3fdbd120692e3546add02e2305a6b5
raw_audit_logs/beam_100k_qlora_full400_audit_log.json	400	69f93d66504baa53ec6ccde88f9b0019a473f718bacb393508dddae5aeedbdd6
raw_audit_logs/beam_100k_v200_audit_log.json	400	f20825bdb26e4433e6c3d04e41a2a7b887046397341d9cfb9d460b503dece223
raw_audit_logs/beam_500k_v200_audit_log.json	700	d99f8e135ab7f8534ef23b7e6a738fe398bbe20a708f4c0dd6f8932dfdf62db7
raw_audit_logs/beam_1m_v200_audit_log.json	700	421de391798e60c0b0d7e816365593706ca353dbb0f74300dd816ac33f281f51
raw_audit_logs/beam_10m_v200_audit_log.json	200	ac29b9769085325370e3170de10ae9a2cb7522888e648ee52369d22d59fbd7d3
raw_audit_logs/beam_100k_v190_audit_log.json	400	b01330af93d00f1d0d498cf47421dcc3b6ec116e88e43423bfa5103e7d840eb0
raw_audit_logs/beam_500k_v190_audit_log.json	700	19644d5fc6d242e9c72cf4ce113494899641aa20ecf398b1da25bd2cc6408a2b
raw_audit_logs/beam_1m_v190_audit_log.json	700	20ce9281d4093a11747926490ffe1efb120a7fb26df3e2412c81a795d4a64417
raw_audit_logs/beam_10m_v190_audit_log.json	200	bc70c9f9fee3e704f379f79db3e063e175bf0f33068ac01b9e660876c0492657

---

📚 Upstream Dataset Attributions & Citations

• DolphinBench (Apache-2.0, Mem0 / mem0ai/dolphinbench, https://dolphinbench.ai/).
• AgentDojo (MIT, Debenedetti et al., NeurIPS 2024, arXiv:2406.13352, ethz-spylab/agentdojo).
• CoIR (Apache-2.0, Li et al., ACL 2025, arXiv:2407.02883, CoIR-Retrieval/*).
• BIRD-SQL Mini-Dev (CC-BY-SA-4.0, Li et al., NeurIPS 2023, arXiv:2305.03111, birdsql/bird_mini_dev).
• Berkeley Function Calling Leaderboard (BFCL v3) (Apache-2.0, Patil et al., UC Berkeley Gorilla, gorilla-llm/Berkeley-Function-Calling-Leaderboard).
• RouterBench (Apache-2.0, Hu et al. / Martian, arXiv:2403.12031, withmartian/routerbench).
• LLMRouterBench (Apache-2.0, Li et al., ACL 2026, arXiv:2601.07206, NPULH/LLMRouterBench).
• RouterArena (Apache-2.0, RouteWorks / Rice University, RouteWorks/RouterArena).
• LegalBench-RAG (MIT, Pipitone & Alami 2024, arXiv:2408.10343), CUAD, MAUD, ContractNLI, PrivacyQA.
• RepoBench-R (CC-BY-4.0, Liu et al., ICLR 2024, arXiv:2306.03091).
• Aider Polyglot Benchmark (Apache-2.0 / MIT, Aider-AI/polyglot-benchmark).
• BEAM (Tavakoli et al., ICLR 2026, arXiv:2510.27246, Mohammadta/BEAM & Mohammadta/BEAM-10M).