AI Agent Benchmarks — Saturation Catalog (agent edition · no-JS)
Machine-readable reference of agentic (agent = Model + Harness/scaffold) evaluation benchmarks and how saturated each is. Snapshot 2026-09. This page is built for AI agents and CLI browsers; it needs no JavaScript.
Machine endpoints: llms.txt (agent entry point) · data.json · data.csv · catalog.md · schema.json · methodology.html · corrections.html · CITATION.cff
Dataset version 2026.09.0 (2026-09-30), licensed CC BY 4.0. Cite the version, not just the URL: Santa María Morales, Santiago (2026). AI Agent Benchmarks — Saturation Tracker (version 2026.09.0) [Data set]. https://benchmarks.santismm.com/
saturation = SOTA as % of the metric ceiling (or human reference if no fixed ceiling): open(<50) · active(50-75) · near(75-90) · saturated(≥90). priority = 0.40·Headroom + 0.25·Durability + 0.20·Credibility + 0.15·Feasibility (1-5 each).
Caveat: Snapshot as of 2026-09. Releases since the August snapshot: Claude Opus 5.5 (09-22), Claude Sonnet 5.5 (09-28) and Fable 5.1; GPT-6 Astra and GPT-6.1 Sol (09-29); Gemini 4 Argon (09-30, limited release); Grok 4.7 (09-21); Muse Spark 1.3 (09-02); MiMo-V2.6-Pro (09-21, MIT open weights); GLM-5.3 (08-14); Kimi K3 (July). Only Anthropic's two launch pages could be read directly; rows tagged Evidence B rely on second-hand reports of vendor tables because the vendor pages could not be read from the build environment. Many SOTA figures remain vendor- or aggregator-sourced and harnesses are not standardized. Treat as indicative, not definitive — verify against each row's `source`. 'agent' = Model + Harness/scaffold; the scaffold (named in `system`) materially affects scores.
Rows sorted by priority-to-beat. Full fields & per-benchmark notes are in data.json / catalog.md.
| Benchmark | Saturation | SOTA / who / when | Headroom | Priority | Source |
|---|---|---|---|---|---|
| Konwinski Prize (K Prize) Software engineering | open 7.5% | 7.5% Open-source entry (E. Rocha) — Laude Institute / Konwinski 2025-07 | 92.5% | 4.65 High | src |
| ScienceAgentBench ML research / Data science | open 33.3% | 33.3% SAB Self-Debug + o3 Medium ($11.69/run) — OSU NLP (HAL) 2025-04 | 66.7% | 4.65 High | src |
| WebGames Web / Computer-use / GUI | open 48% | 48% Best browser agent (VLM + Browser Use framework) — Convergence AI 2026 | 52% | 4.6 High | src |
| ARC-AGI-3 Frontier reasoning | open 30.2% | 30.2% Claude Opus 5 30.2 (tracked snapshot); GPT-5.6 Sol 7.8; best pre-frontier agent 12.58 (Tufa 'StochasticGoose') — Anthropic / ARC Prize 2026-08 | 69.8% | 4.55 High | src |
| Legal Agent Benchmark (LAB) Tool-use / Assistant | open 19.6% | 19.6% Gemini 4 Argon 19.6; Grok 4.7 19.6; GPT-6 Astra 5.4; Claude Opus 5.5 3.8; GPT-5.6 Sol (max) 2.5 (vendor tables) — Harvey 2026-09 | 80.4% | 4.5 High | src |
| TheAgentCompany Long-horizon / Economic | open 30% | 30% Gemini 2.5 Pro / Claude 3.7 + OpenHands — CMU 2025 | 70% | 4.5 High | src |
| AssistantBench Tool-use / Assistant | open 25.2% | 25.2% SPA (official); ~42% claimed by aggregators — Allen AI / Tel Aviv 2024 | 74.8% | 4.45 High | src |
| Blueprint-Bench 2 Frontier reasoning | open 38.6% | 38.6% Claude Fable 5 / Mythos 5 38.6 (vendor, contested); GPT-5.5 36.2; Andon's own writeup ranks GPT-5.5 #1 — Andon Labs 2026-06 | 61.4% | 4.3 High | src |
| VisualWebArena Web / Computer-use / GUI | open 36.5% | 36.5% Recon-Act (GPT-4o-class) — Academic (CMU) 2025 | 63.5% | 4.3 High | src |
| SpreadsheetBench 2 Office / productivity | open 34.8% | 34.8% Kimi K3 34.8; Fable 5 34.7; GPT-5.6 Sol 32.4 (Moonshot card) — Academic (arXiv 2606.29955) 2026-07 | 65.2% | 4.25 High | src |
| Multi-step cyber autonomy (AISI) Cyber / Autonomy | active 68.8% | 22 steps Claude Mythos Preview (researcher harness) — first to complete the 32-step range; Opus 4.6 ~16 — UK AISI 2026-04 | 31.2% | 4.2 High | src |
| SWE-Lancer Software engineering | active 66.3% | 66.3% GPT-5.1 Codex (OpenAI harness) — OpenAI 2026-06 | 33.7% | 4.15 High | src |
| WorkArena++ Web / Computer-use / GUI | open 2.1% | 2.1% Claude Sonnet + BrowserGym (full L2/L3 ~0-2%; easier L2 subset ~39) — ServiceNow Research / academic 2026 | 97.9% | 4.15 High | src |
| SWE-bench Multimodal Software engineering | active 59% | 59% Claude Mythos Preview (self-reported) — Anthropic 2026-06 | 41% | 4.05 High | src |
| SWE-bench Pro Software engineering | near 80.3% | 80.3% Claude Mythos 5 / Fable 5 80.3 (Anthropic scaffold, vendor); Opus 5 79.2; SEAL-standardized top Muse Spark 1.1 (Meta) 61.5 — Anthropic / Scale SEAL 2026-06 | 19.7% | 4.05 High | src |
| RE-Bench ML research / Data science | nosota % | Claude 3.5 Sonnet / o1-preview + METR scaffold — METR 2024-11 | % | 4.0 High | src |
| tau-bench Tool-use / Assistant | active 70% | 70% Claude Sonnet 4.5 (airline ~70); overall record Mythos 5 / Fable 5 tied 89.2 (GA again since 2026-07-01) — Sierra (orig) 2026-06 | 30% | 4.0 High | src |
| Vending-Bench 2 Long-horizon / Economic | open 17.7% | 11181 $ Claude Opus 5 $11,181 avg over 5 runs (Andon board); Opus 4.7 $10,937; GPT-5.6 Sol $9,619 — Andon Labs 2026-07 | 82.3% | 4.0 High | src |
| FrontierCode 1.1 (Main) Software engineering | active 54.4% | 54.4% Claude Opus 5.5 54.4 (medium/default effort, vendor); GPT-6 Astra 53.3; Sonnet 5.5 52.1 (xhigh) / 46.2 (max); Fable 5.1 50.3; GPT-6 Sol 49.3 — Cognition 2026-09 | 45.6% | 3.95 High | src |
| GAIA2 (ARE) Tool-use / Assistant | nosota % | Frontier models in ARE harness (Pareto vs cost) — Meta (FAIR) 2025-09 | % | 3.95 High | src |
| Multi-SWE-bench Software engineering | nosota % | n/a — ByteDance Seed 2026 | % | 3.95 High | src |
| OfficeBench Office / productivity | open 47% | 47% GPT-4 (best at publication) — UCSD / authors 2024 | 53% | 3.9 High | src |
| ToolSandbox Tool-use / Assistant | nosota % | Proprietary >> open (no single SOTA) — Apple 2024-08 | % | 3.9 High | src |
| Mind2Web 2 Web / Computer-use / GUI | active 60% | 60% OpenAI Deep Research (o-series agent) — OpenAI 2025-06 | 40% | 3.85 High | src |
| GAIA Tool-use / Assistant | active 74.6% | 74.6% HAL Generalist Agent + Claude Sonnet 4.5 (val, reproducible single-agent) — Princeton HAL / orig Meta-HF 2026-05 | 25.4% | 3.8 High | src |
| Humanity's Last Exam (HLE) Frontier reasoning | active 67.7% | 67.7% Claude Opus 5.5 67.7 (with tools); Fable 5.1 65.6; Sonnet 5.5 64.5; Opus 5 63.6; GPT-6 Astra 57.2 (Anthropic pages) — Anthropic 2026-09 | 32.3% | 3.8 High | src |
| tau2-bench Tool-use / Assistant | near 76.5% | 76.5% LongCat-Flash-Thinking-2601 (airline 76.5); telecom ~99 (JT-35B-Flash 99.1); Fable 5 telecom 98.5 — Sierra (orig) 2026-06 | 23.5% | 3.8 High | src |
| FutureX Deep research | active 56% | 56% H2O AI Super Agent — H2O.ai (board by FutureX team) 2026 | 44% | 3.75 High | src |
| APEX-Agents-AA Tool-use / Assistant | open 47.1% | 47.1% Gemini 3.5 Flash (high) on AA Stirrup harness — Artificial Analysis (data by Mercor) 2026-05 | 52.9% | 3.7 High | src |
| BFCL v4 Tool-use / Assistant | near 75% | 75% Qwen3.7 Max (native FC) — UC Berkeley (Gorilla) 2026-08 | 25% | 3.7 High | src |
| PaperBench ML research / Data science | open 24.4% | 24.4% orig o1-high 24.4 (verified); Kimi K2.5 63.5 self-reported/unverified — Moonshot AI / orig OpenAI 2026-05 | 75.6% | 3.7 High | src |
| AutomationBench Tool-use / Assistant | active 53.1% | 53.1% MiMo-V2.6-Pro 53.1 (v1.0.6, Xiaomi-reported); Gemini 4 Argon 51.3; Muse Spark 1.3 49.4; GPT-6 Astra 41.4; Claude Opus 5.5 40.0 (Zapier early access; 42.5 in Google's table); Fable 5.1 31.4; Kimi K3 30.8 — Zapier (benchmark); top figure reported by Xiaomi 2026-09 | 46.9% | 3.65 High | src |
| CursorBench 4.0 Software engineering | active 57.8% | 57.8% Claude Opus 5.5 57.8 (medium/default); Sonnet 5.5 55.5; Fable 5.1 51.8; Grok 4.7 46.3; GPT-5.6 Sol (max) 41.7 — Cursor 2026-09 | 42.2% | 3.65 High | src |
| GTA / GTA-2 Tool-use / Assistant | nosota % | Frontier models in GTA harness (GPT-4 era <50) — OpenCompass / Shanghai AI Lab 2026-02 | % | 3.65 High | src |
| SWE-rebench Software engineering | active 65.3% | 65.3% Claude Opus 4.6 + fixed 128K ReAct scaffold — Nebius 2026-05 | 34.7% | 3.6 High | src |
| AgentBench Tool-use / Assistant | nosota % | Frontier models (general); no verified 2026 SOTA — Tsinghua (THUDM) | % | 3.55 Med | src |
| ARC-AGI-2 Frontier reasoning | active 52.9% | 52.9% GPT-5.2 X-High ~52.9% @ $1.90/task (efficiency-limited; first to ~match human) — OpenAI / Google 2026-05 | 47.1% | 3.55 Med | src |
| ExploitBench Cyber / Autonomy | near 78% | 78% Claude Fable 5 78.0 (llm-stats board, 2026-07); Mythos 5 (unblocked, Anthropic-reported) 78.0; Opus 4.8 40.0; GPT-5.5 34.0 — CMU + Bugcrowd 2026-07 | 22% | 3.55 Med | src |
| LiveBench (agentic-coding subset) Software engineering | nosota % | not separately reported — LiveBench (Abacus.AI / NYU / Nvidia) 2025-04 | % | 3.55 Med | src |
| BioMysteryBench ML research / Data science | open 46.1% | 46.1% Claude Fable 5 / Mythos 5 46.1 hard / 83.9 human-solvable (vendor); Opus 4.8 40.0 / 80.4 — Anthropic 2026-06 | 53.9% | 3.5 Med | src |
| SWE-Marathon Software engineering | open 42% | 42% Kimi K3 42.0; Opus 4.8 40.0; GPT-5.6 Sol 39.0; Fable 5 35.0 (Moonshot card) — Academic (arXiv 2606.07682) 2026-07 | 58% | 3.5 Med | src |
| DeepResearch Bench (RACE) Deep research | active 56% | 55.95 score NVIDIA AI-Q deep research agent — NVIDIA 2026-03 | 44% | 3.4 Med | src |
| GDPval Long-horizon / Economic | open 41% | 41% GDPval-AA v2.1 Elo: Claude Opus 5.5 1846; Sonnet 5.5 1844; Fable 5.1 1735; Opus 5 1708; GPT-5.6 Sol 1588; GPT-6 Astra 1542 (Anthropic pages) — OpenAI / Artificial Analysis 2026-09 | 59% | 3.4 Med | src |
| NYU CTF Bench Cyber / Autonomy | active 59% | 59% Claude Opus 4.5 (CTF agent harness) — NYU 2026-04 | 41% | 3.4 Med | src |
| WebArena Web / Computer-use / GUI | active 74.3% | 74.3% WebTactix (DeepSeek v3.2 scaffold) 74.3 (steel.dev tracked); OpAgent v2 71.6 (2026-04) — academic / steel.dev tracker 2026-06 | 25.7% | 3.4 Med | src |
| Chartography Frontier reasoning | active 64.4% | 64.4% Claude Opus 5.5 64.4 (no tools); Sonnet 5.5 61.6; GPT-6 Sol 53.6. With tools (Anthropic protocol): Opus 5.5 89.0, Fable 5.1 88.4, Opus 5 83.4 — Surge AI 2026-09 | 35.6% | 3.35 Med | src |
| SpreadsheetBench Office / productivity | active 70.5% | 70.5% Gemini in Google Sheets (vendor) — Google 2026-03 | 29.5% | 3.35 Med | src |
| Aider Polyglot Software engineering | near 88% | 88% gpt-5 high (88%) / Opus 4.5 (89.4%, vendor) via aider — OpenAI / Anthropic 2025-08 | 12% | 3.3 Med | src |
| DSBench ML research / Data science | open 34% | 34% DatawiseAgent (notebook-centric); modeling >90 — Academic 2025-03 | 66% | 3.25 Med | src |
| FrontierMath Frontier reasoning | near 87.7% | 87.7% GPT-5.5 Pro 87.7 ±1.9 (v2, LM Council run); Claude Fable 5 87.0 ±2.0 (statistically tied; Epoch's own note puts Fable first) — OpenAI / Epoch AI 2026-08 | 12.3% | 3.25 Med | src |
| MLE-bench ML research / Data science | active 64.4% | 64.4% Famou-Agent 2.0 + Gemini-3-Pro (24h/task) — Baidu 2026-02 | 35.6% | 3.25 Med | src |
| BIRD-SQL Data / SQL | near 81.7% | 81.67% Agentar-Scale-SQL (test-time scaling) — Ant Group / community 2026 | 18.3% | 3.2 Med | src |
| CORE-Bench (Hard) ML research / Data science | near 77.8% | 77.8% Claude Code + Claude Opus 4.5 ($87/run) — Princeton HAL 2025-12 | 22.2% | 3.2 Med | src |
| CWE-bench v1 Cyber / Autonomy | active 68% | 68% Gemini 4 Argon = GPT-6 Astra = Grok 4.7 68 (tie) — Collinear AI 2026-09 | 32% | 3.2 Med | src |
| HealthBench Professional Tool-use / Assistant | active 66% | 66% Claude Fable 5 / Mythos 5 66.0 (vendor); Opus 4.8 56.9; GPT-5.5 51.8; OpenAI GPT-5.4 'ChatGPT for Clinicians' beats physician baseline — OpenAI 2026-06 | 34% | 3.2 Med | src |
| OfficeQA Pro Office / productivity | active 63.3% | 63.3% Kimi K3 63.3 (Moonshot card; Codex harness for the GPT rows) — Databricks 2026-07 | 36.7% | 3.2 Med | src |
| OSWorld 2.0 Web / Computer-use / GUI | active 73.5% | 73.5% GPT-6 Astra 73.5 (max; another report 72.6); GPT-6.1 Sol 71.4 (max); Muse Spark 1.3 66.9; Gemini 3.8 Flash 59.0 — XLANG / HKU (OSWorld); figures vendor-reported (OpenAI) 2026-09 | 26.5% | 3.2 Med | src |
| Spider 2.0 Data / SQL | active 73.1% | 73.13% DivSkill-SQL (Lite 73.13); Genloop Sentinel Agent v2 Pro (full-set 96.7) — Snowflake/UCSD; Genloop 2026-06 | 26.9% | 3.2 Med | src |
| Terminal-Bench 4.0 Software engineering | active 70.6% | 70.6% Claude Sonnet 5.5 70.6; Opus 5.5 66.4 (xhigh); GPT-6 Astra 57.9 (high) (Anthropic pages) — Terminal-Bench (Stanford / Laude Institute); top figure vendor-reported (Anthropic) 2026-09 | 29.4% | 3.2 Med | src |
| Terminal-Bench-Science 0.1 ML research / Data science | active 64.6% | 64.6% GPT-6 Astra 64.6; Opus 5.5 58.7 (xhigh); Fable 5.1 52.6; Opus 5 29.0; GPT-5.6 Sol 22.4 (Anthropic page) — Stanford-led, Terminal-Bench team 2026-09 | 35.4% | 3.2 Med | src |
| Vals Finance Agent v2 Tool-use / Assistant | active 65.4% | 65.4% Gemini 4 Argon 65.4; GPT-6 Astra ~53.5; Claude Opus-class ~58.6 (vendor tables) — Vals AI 2026-09 | 34.6% | 3.2 Med | src |
| WebDev / Code Arena (LMArena) Web / Computer-use / GUI | nosota % | 1582 Elo Claude Opus 4.8 ~1582 Elo (Code Arena); WebDev-specific board lists Opus 4.7 1556 — LMArena 2026-08 | % | 3.2 Med | src |
| Terminal-Bench 2.0 Software engineering | near 84.7% | 84.7% NexAU-AHE harness + GPT-5.5 (official board) — OpenAI model / NexAU 2026-05 | 15.3% | 3.15 Med | src |
| Terminal-Bench 2.1 Software engineering | near 89.5% | 89.5% GPT-5.6 Sol 89.5 (AA board); Claude Opus 5 89.1 — OpenAI 2026-08 | 10.5% | 3.15 Med | src |
| ScreenSpot-Pro Web / Computer-use / GUI | near 87.9% | 87.9% Claude Opus 4.8 (native grounding, general); specialist GTA1-32B 95.2 — Anthropic 2026-06 | 12.1% | 3.05 Med | src |
| AndroidWorld Web / Computer-use / GUI | near 87.1% | 87.1% H Company Surfer 2 (vision-only, vendor 87.1); verified board top Qwen3.8 Max 85.3 — H Company / Alibaba 2026 | 12.9% | 3.0 Med | src |
| BrowseComp Web / Computer-use / GUI | saturated 92.2% | 92.2% GPT-5.6 Sol Ultra (agentic deep-research setup) 92.2; Kimi K3 91.2; Claude Opus 5 90.8 — OpenAI 2026-08 | 7.8% | 3.0 Med | src |
| MCP Atlas Tool-use / Assistant | near 75.5% | 75.5% Muse Spark 1.3 75.5 (Meta model card) — Scale AI 2026-09 | 24.5% | 2.95 Low | src |
| SWE-bench Verified Software engineering | saturated 96% | 96% Claude Opus 5 96.0; GPT-5.6 Sol 96.2 (vals.ai run 97.0); Fable 5 95.0 / Mythos 5 95.5 — Anthropic / OpenAI 2026-08 | 4% | 2.9 Low | src |
| DeepSWE v1.1 Software engineering | near 77.9% | 77.9% Gemini 4 Argon 77.9; Muse Spark 1.3 75.4; GPT-6.1 Sol ~75.2; Opus 5.5 74.2; GPT-6 Astra 74.1; Grok 4.7 71.0 (high); GLM-5.3 66.9 — Datacurve 2026-09 | 22.1% | 2.8 Low | src |
| OSWorld (Verified) Web / Computer-use / GUI | near 86.1% | 86.1% Qwen3.8 Max 86.1; Claude Mythos 5 / Fable 5 85.0-85.4; Opus 4.8 83.4 — Alibaba (Qwen) 2026-08 | 13.9% | 2.75 Low | src |
| ProgramBench Software engineering | near 77.8% | 77.8% Kimi K3 77.8; GPT-5.6 Sol 77.6; Fable 5 76.8 (Moonshot card) — Not verified 2026-07 | 22.2% | 2.6 Low | src |
| CyberGym Cyber / Autonomy | near 84.5% | 84.5% GLM-5.3 84.5; Claude Mythos 5 83.8; GPT-5.6 Sol 83.6 (Z.ai-reported) — UC Berkeley; figure Z.ai-reported 2026-08 | 15.5% | 2.55 Low | src |
| Online-Mind2Web Web / Computer-use / GUI | saturated 97% | 97% Browser Use Cloud v3 (Auto-Research scaffold) — Browser Use 2026-03 | 3% | 2.55 Low | src |
| OSWorld 2.1 Web / Computer-use / GUI | near 81.8% | 81.8% Claude Opus 5.5 81.8; Fable 5.1 80.7; Sonnet 5.5 80.1; Opus 5 74.0 (Anthropic pages) — XLANG / HKU (OSWorld); figures Anthropic-reported 2026-09 | 18.2% | 2.55 Low | src |
| FrontierSWE Software engineering | near 86.6% | 86.6% Claude Fable 5 86.6 (max, with fallback); Kimi K3 81.2; GPT-5.6 Sol 71.3; GLM-5.2 67.3 (Moonshot card) — Not verified (cited as Chu et al., 2026) 2026-07 | 13.4% | 2.2 Low | src |
| Cybench Cyber / Autonomy | saturated 100% | 100% Claude Mythos Preview ~100 (self-reported) — Anthropic (eval: Stanford) 2026-06 | 0% | 2.1 Low | src |
| WebVoyager Web / Computer-use / GUI | saturated 98.5% | 98.5% Alumnium (Claude Sonnet 4.6 policy + GPT-5 Nano) + Selenium — Alumnium 2026-03 | 1.5% | 1.9 Low | src |
| AA-Briefcase v1.1 Long-horizon / Economic | nosota % | 1822 Elo Claude Opus 5.5 1822; Sonnet 5.5 1811; GPT-6 Sol 1483 (Anthropic pages) — Artificial Analysis 2026-09 | % | n/a | src |
| AgentHarm Autonomy / Safety | nosota % | n/a (safety eval; inverse metric) — Gray Swan + UK AISI 2024-10 | % | n/a (safety eval) | src |
| Agents' Last Exam Long-horizon / Economic | nosota % | Conflicting vendor tables: OpenAI reports GPT-6 Astra 59.3 (Opus 5 55.5, GPT-5.6 Sol 53.6); Google reports Gemini 4 Argon 39.5 vs Astra 34.2; GLM-5.3 28.5; Kimi K3 28.3 — Berkeley RDI (UC Berkeley) with 250+ industry experts 2026-09 | % | n/a | src |
| Artificial Analysis Intelligence Index Meta-leaderboards | tracker % | Index v4.3: Claude Fable 5.1 / GPT-6 53; GPT-6.1 Sol (high) 50, (medium) 48; Muse Spark 1.3 48; Grok 4.7 46; MiMo-V2.6-Pro 46.3 — re-based vs June (Fable 5 64.9) — Artificial Analysis 2026-09 | % | n/a (tracker) | src |
| Epoch AI Benchmarking Hub Meta-leaderboards | tracker % | ECI 159: GPT-5.5 Pro (Fable 5 not yet recomputed) — Epoch AI 2026-04 | % | n/a (tracker) | src |
| Galileo Agent Leaderboard v2 Meta-leaderboards | tracker % | n/a — Galileo AI | % | n/a (tracker) | src |
| HAL - Holistic Agent Leaderboard Meta-leaderboards | tracker % | n/a — Princeton PLI | % | n/a (tracker) | src |
| METR Time Horizon Autonomy / Safety | nosota % | 14.5 hours Claude Opus 4.6 14.5h (measured); no measured Fable 5 / Opus 5 figure yet — METR 2026-02 | % | n/a (measurement) | src |
| Terminal-Bench 3.0 Software engineering | nosota % | GLM-5.3 28.3 (Z.ai-reported open-weights SOTA; GLM-5.2 4.6); no overall leader identified — Terminal-Bench (Stanford / Laude Institute) 2026-08 | % | n/a | src |
| Vellum LLM Leaderboard Meta-leaderboards | tracker % | n/a — Vellum | % | n/a (tracker) | src |
| WANDR Deep research | nosota % | Claude Opus 5.5 scores above Fable 5.1 and Opus 5 at lower cost (Anthropic page) — no numeric score published — Not verified (one third-party article attributes it to Perplexity) 2026-09 | % | n/a | src |
agent = Model + Harness · snapshot 2026-09 · verify numbers against each benchmark's primary source.