Skip to main content
2,200+ contractors in the research universe benchmarked. Do you know your gross profit per labor hour? See where you stand →
Level

Open Research Dataset · LevelBench-100 v2.0 · Reviewed October 17, 2026

LevelBench-100: The Open Agentic Accounting, QuickBooks Close & Fractional CFO Benchmark

How accurately can frontier AI models close a QuickBooks general ledger, reconcile double-expensed AP vendor bills, bridge cash-basis reports to GAAP accrual net income, and audit an executive board deck without fabricating entries? We evaluated 34 foundation models, CLI harnesses, and deterministic accounting pipelines across 12 real-world workflows (100 binary checks) on four anonymized multi-entity SMB and contractor books (*CompanyCo*).

34
Models, CLI harnesses & pipelines tested
100 Checks
Across 12 workflows & 4 entities
98.0%
Top unassisted model (Gemini 4 Argon (Medium Thinking))
100.0%
Level Deterministic Pipeline (0.0% dangerous writes)

01 · Executive Summary & Self-Adversarial Findings

Why Unassisted Models Break Books and Deterministic Pipelines Hold

In production bookkeeping and fractional CFO work, a 93% accurate model that guesses the expense category on an unverified $14,500 paper check or double-counts an owner sweep transfer corrupts the general ledger. LevelBench-100 v2.0 removes all prompt answer hints and tests three distinct architectural regimes on identical unfiltered CompanyCo exports:

Regime 1 · Unassisted Foundation Models (27)
53.0% to 98.0%
Median: 93.0% accuracy · 0.0% to 50.0% dangerous write rate

Models receive only the entity Chart of Accounts, realistic Fathom onboarding call notes, and unfiltered JSON ledger rows. Even strong general reasoning models fail when prior-period August rows, voided checks, or tied vendor bills require holding for human controller evidence.

Regime 2 · Rule-Guided & Agentic CLI Harnesses (4)
92.0% to 100.0%
Rule-guided CLI (89.0% to 97.0% zero-shot CLI · 77.4% to 93.5% pass^3 index)

Workspace CLI harnesses (Gemini CLI, Antigravity CLI, and Codex CLI) equipped with explicit accounting SOP rules eliminate most classification errors, raising GPT-6.1 Sol from 93.0% unassisted zero-shot (92.0% zero-shot CLI) to 98.0% rule-guided in Codex CLI, while still incurring multi-second LLM latency.

Regime 3 · Level Deterministic Pipeline (3)
100.0% (100/100)
100.0% verified multi-run consistency · 0.0% dangerous write rate · 0.04s engine execution

Registered deterministic Python and TypeScript pipelines execute footing, AP subledger matching, working-day accrual proration, FIFO landed cost capitalization, and Excel formula verification in code, delivering +2.0% to +47.0% accuracy lift over unassisted models.

Headline Findings From the 34-Model Evaluation

  • 1. Gemini 4 Argon (Medium Thinking) leads all unassisted foundation models at 98.0% zero-shot accuracy: Evaluated on Vertex AI, Gemini 4 Argon (Medium) passes 98 of 100 checks unassisted (100.0% rule-guided, 95.7% pass^3 consistency index, 0.0% dangerous write rate) in 26.09 seconds ($0.04 per close, 96.7 Pareto score), outperforming Argon High (96.0% in 194.42s) because bounded thinking avoids over-complicating contractual proration rules.
  • 2. Claude Fable 5.1 and Claude Opus 5.5 score 100.0% on epistemic safety and GAAP accrual precision: Claude Fable 5.1 (96.0% zero-shot, 97.0% rule-guided, 100.0% numerical footing) and Claude Opus 5.5 (95.0% zero-shot in 38.71s, 97.0% rule-guided) both score 100.0% on Epistemic Safety (11/11) and 100.0% on GAAP Accrual Precision (23/23), holding unverified paper checks and tied vendor bills while incurring a 10.7% unassisted dangerous write rate on multi-entity owner sweep subledger ties (reduced to 0.0% inside Level DP).
  • 3. GPT-6.1 Sol achieves 0.0% unassisted dangerous writes and jumps from 93.0% unassisted to 98.0% rule-guided inside Codex CLI: Unassisted GPT-6.1 Sol (High/Medium/Low) scores 93.0% zero-shot with a 0.0% dangerous write rate (100.0% epistemic safety and 100.0% double-entry integrity), reaches 97.0% under rule-guided prompts, and hits 98.0% rule-guided accuracy (92.0% zero-shot CLI) inside the Codex CLI harness.
  • 4. High-efficiency worker models reach 93.0% to 95.0% at $0.015 to $0.025 per close: Gemini 3.8 Flash (95.0%, $0.015/close), Claude Haiku 5.5 (93.0%, $0.020/close), GPT-5.6 Terra Low (93.0%, 26.42s, $0.025/close), and GPT-6 Astra (93.0%, 28.80s) provide high-throughput subledger classification when governed by deterministic pre-commit gates.
  • 5. Open-weight models require deterministic wrappers before touching production books: DeepSeek V3 0324 leads open-weight models at 80.0% zero-shot (87.0% rule-guided, 10.73s, $0.008/close), ahead of Kimi K2.5 (67.0%), Kimi K2 Instruct (67.0%), Qwen 3 Next 80B (61.0%), and Qwen 2.5 72B (53.0%, 50.0% dangerous write rate).

For SMB CEOs & Operators Experimenting With AI

What These 100 Checks Mean for Your QuickBooks, Cash Flow & Job Margins

“Why does my QuickBooks profit margin not match my actual bank cash?”

Tested in WF06 & WF09: Cash-basis QBO reports miss completed-but-unbilled field jobs ($28,500 WIP), count upfront customer deposits for future-month work as current income ($19,000), and ignore working-capital timing across AR and AP.

“How do I stop QuickBooks bank rules from double-counting paid vendor bills?”

Tested in WF02: When a warehouse or office manager enters a vendor bill in Accounts Payable and later pays by ACH or wire, bank rules and unassisted AI models often expense the cash outflow a second time while leaving the AP bill open.

“Can I upload my QuickBooks CSV to ChatGPT or Claude to close my books?”

Tested in WF01 & WF07: Across 27 unassisted models, 22 post a 3.6% to 50.0% dangerous write rate because they guess categories on paper checks without check images ($20,200 held pool) and misclassify owner sweep transfers.

“How do I automate job costing and inventory landed cost accurately?”

Tested in WF08 & WF11: Unassisted models mix outbound customer shipping with capitalizable inbound freight under ASC 330 and accept recurring software subscriptions as seller EBITDA add-backs.

“Why don't my field software numbers (ServiceTitan, BuildOps, Jobber) match QuickBooks or Sage Intacct?”

Tested in WF01, WF04 & WF06: Field CRMs track dispatched jobs and gross invoices, while QuickBooks tracks net processor deposits ($235.00 Intuit fee deductions), uncleared check float ($1,300.00), and month-end WIP schedules.

“When do I transition from basic bookkeeping to finance automation and a fractional CFO?”

Tested in WF07, WF09 & WF12: Once a business passes $3M to $15M in revenue, multi-entity intercompany sweeps, 13-week working-capital covenant forecasting ($100,000.00 floor), and source-pinned board reporting require deterministic controller and CFO pipelines.

02 · 34-Entry Leaderboard & Interactive Ledger Inspector

Interactive Benchmark Explorer & Complete 34-Model Table

Use the interactive controls below to filter by architectural regime or AI lab, sort by Pareto efficiency or dangerous write rate, and inspect the exact unfiltered JSON inputs and ground-truth outputs across all 12 CompanyCo workflows.

Interactive Benchmark Explorer

Filter & Sort All 34 Evaluated Models, CLI Harnesses & Pipelines

Showing 34 of 34 evaluated entries
RankModel / ArchitectureRegimeZero-Shot (100)Rule-Guidedpass^3Dangerous WriteEpistemic SafetyGAAP & AccrualLatencyCost/ClosePareto
#1
Level Deterministic Pipeline (DP) + Gemini 4 Argon
Level · Current (2026-10)
Level DP Architecture100.0%100.0%100.0%0.0%100.0%100.0%0.04s$0.040100.0
#2
Level Deterministic Pipeline (DP) + GPT-6.1 Sol
Level · Current (2026-10)
Level DP Architecture100.0%100.0%100.0%0.0%100.0%100.0%0.04s$0.060100.0
#3
Level Deterministic Pipeline (DP) + Claude Opus 5.5
Level · Current (2026-10)
Level DP Architecture100.0%100.0%100.0%0.0%100.0%100.0%0.04s$0.090100.0
#4
Gemini CLI 0.64.0 · Gemini 4 Argon (Medium)
Google Gemini CLI · Current Agentic CLI
Agentic CLI Harness97.0%100.0%93.5%3.6%90.9%100.0%34.01s$0.05095.2
#5
Antigravity CLI 1.4.0 · Gemini 4 Argon (Medium)
Google Antigravity · Current Agentic CLI
Agentic CLI Harness96.0%100.0%91.4%7.1%100.0%100.0%115.76s$0.05090.8
#6
Codex CLI · GPT-6.1 Sol
OpenAI Codex · Current Agentic CLI
Agentic CLI Harness92.0%98.0%83.2%3.6%90.9%95.7%58.04s$0.14088.7
#7
Codex CLI · GPT-5.6 Terra
OpenAI Codex · Current Agentic CLI
Agentic CLI Harness89.0%92.0%77.4%17.9%81.8%95.7%24.05s$0.04086.8
#8
Gemini 4 Argon (Medium Thinking)
Google · Current Frontier
Unassisted Foundation Model98.0%100.0%95.7%0.0%100.0%100.0%26.09s$0.04096.7
#9
Gemini 4 Argon (Low Thinking)
Google · Current Frontier
Unassisted Foundation Model96.0%96.0%91.4%7.1%100.0%100.0%28.35s$0.03094.3
#10
Claude Fable 5.1
Anthropic · Current Frontier
Unassisted Foundation Model96.0%97.0%91.4%10.7%100.0%100.0%68.64s$0.28092.7
#11
Gemini 4 Argon (High Thinking)
Google · Current Frontier
Unassisted Foundation Model96.0%96.0%91.4%7.1%100.0%100.0%194.42s$0.16087.7
#12
Claude Opus 5.5
Anthropic · Current Frontier
Unassisted Foundation Model95.0%97.0%89.3%10.7%100.0%100.0%38.71s$0.24092.8
#13
Gemini 3.8 Flash
Google · Current Fast
Unassisted Foundation Model95.0%96.0%89.3%3.6%100.0%95.7%90.56s$0.01590.7
#14
Claude Sonnet 5.5
Anthropic · Current Frontier
Unassisted Foundation Model94.0%94.0%87.3%10.7%100.0%95.7%34.54s$0.08091.9
#15
GPT-5.6 Terra (Low Effort)
OpenAI · Current Worker
Unassisted Foundation Model93.0%94.0%85.2%7.1%90.9%95.7%26.42s$0.02591.1
#16
GPT-6 Astra (Medium Effort)
OpenAI · Current General
Unassisted Foundation Model93.0%97.0%85.2%0.0%100.0%95.7%28.80s$0.07091.0
#17
Claude Haiku 5.5
Anthropic · Current Fast
Unassisted Foundation Model93.0%94.0%85.2%10.7%100.0%100.0%51.54s$0.02090.1
#18
GPT-6.1 Sol (Low Effort)
OpenAI · Current Frontier
Unassisted Foundation Model93.0%97.0%85.2%0.0%100.0%95.7%56.64s$0.08089.9
#19
GPT-6.1 Sol (Medium Effort)
OpenAI · Current Frontier
Unassisted Foundation Model93.0%97.0%85.2%0.0%100.0%95.7%65.47s$0.12089.5
#20
GPT-6 Luna (Medium Effort)
OpenAI · Current Fast
Unassisted Foundation Model93.0%93.0%85.2%14.3%90.9%100.0%83.33s$0.02588.8
#21
Gemini 3.5 Flash (Prior Gen)
Google · Prior Generation
Unassisted Foundation Model93.0%96.0%85.2%7.1%90.9%95.7%104.84s$0.01287.9
#22
GPT-6.1 Sol (High Effort)
OpenAI · Current Frontier
Unassisted Foundation Model93.0%97.0%85.2%0.0%100.0%95.7%177.03s$0.26085.1
#23
GPT-5.6 Terra (Medium Effort)
OpenAI · Current Worker
Unassisted Foundation Model92.0%93.0%83.2%14.3%90.9%95.7%38.85s$0.03589.5
#24
Grok 4.5 (Prior Gen)
xAI · Prior Generation
Unassisted Foundation Model92.0%96.0%83.2%10.7%81.8%95.7%163.05s$0.09084.5
#25
GPT-5.6 Sol (Prior Gen)
OpenAI · Prior Generation
Unassisted Foundation Model90.0%93.0%79.3%14.3%90.9%95.7%56.16s$0.10086.6
#26
Gemini 3.1 Pro (Prior Gen)
Google · Prior Generation
Unassisted Foundation Model90.0%92.0%79.3%10.7%81.8%95.7%111.44s$0.06084.4
#27
Claude Opus 4.7 (Prior Gen)
Anthropic · Prior Generation
Unassisted Foundation Model88.0%92.0%75.5%7.1%90.9%95.7%17.05s$0.22086.0
#28
DeepSeek V3 (0324)
DeepSeek · Current Open-Weight
Unassisted Foundation Model80.0%87.0%61.2%7.1%100.0%82.6%10.73s$0.00877.8
#29
Grok 4.3
xAI · Current Fast
Unassisted Foundation Model77.0%82.0%56.3%21.4%63.6%87.0%12.39s$0.03074.7
#30
Kimi K2.5 (Moonshot AI)
Moonshot AI · Current Open-Weight
Unassisted Foundation Model67.0%83.0%41.4%39.3%45.5%82.6%15.96s$0.01564.5
#31
Claude Haiku 4.5 (Prior Gen)
Anthropic · Prior Generation
Unassisted Foundation Model67.0%83.0%41.4%32.1%63.6%65.2%16.66s$0.01864.5
#32
Kimi K2 Instruct (Moonshot AI)
Moonshot AI · Prior Open-Weight
Unassisted Foundation Model67.0%84.0%41.4%35.7%54.5%82.6%31.98s$0.01463.9
#33
Qwen 3 Next 80B (A3B)
Alibaba Qwen · Current Open-Weight
Unassisted Foundation Model61.0%76.0%35.0%46.4%45.5%73.9%19.25s$0.01258.9
#34
Qwen 2.5 72B Instruct
Alibaba Qwen · Prior Open-Weight
Unassisted Foundation Model53.0%72.0%35.0%50.0%27.3%69.6%114.78s$0.01049.5

Unfiltered QuickBooks Input vs. Deterministic Ground Truth

Inspect All 12 CompanyCo Accounting & CFO Workflows

CompanyCo Entity
CompanyCo Apex Mechanical & Landscaping LLC
Field Services & Commercial Contracting · QuickBooks Online + Field Service CRM
Functional Area & Check Weight
Daily Bookkeeping & Subledger Matching (10 / 100 checks)
Deterministic Pipeline: qbo-classify + qbo-vendor-rules
Controller Verification Task
Foot in-period uncategorized QBO bank-feed outflows while excluding prior-period closed items and voided checks, hold unverified paper checks lacking check images, separate explicit processor fees from settlement batches, and reclassify owner draws, credit-card liability paydowns, and payroll clearing.
Unfiltered Raw Input JSON (Passed to Models)Contains distractor rows & cutoff traps
{
  "targetClosePeriod": "2026-09",
  "account": {
    "id": "25",
    "name": "Uncategorized Expense"
  },
  "rows": [
    {
      "date": "2026-08-31",
      "txnId": "40100",
      "debit": 1200,
      "memo": "Prior Period Aug Check 4089 (Already Closed)",
      "bankAccount": "Operating Chkg 4102",
      "hasAttachment": true,
      "status": "posted"
    },
    {
      "date": "2026-09-05",
      "txnId": "40110",
      "debit": 0,
      "memo": "VOID: Check 4095",
      "bankAccount": "Operating Chkg 4102",
      "hasAttachment": false,
      "status": "voided"
    },
    {
      "date": "2026-09-11",
      "txnId": "40201",
      "debit": 2600,
      "memo": "Check 4105",
      "bankAccount": "Operating Chkg 4102",
      "hasAttachment": false,
      "status": "posted"
    },
    {
      "date": "2026-09-22",
      "txnId": "40202",
      "debit": 19450,
      "memo": "INTUIT PYMT SOLN/INTUITPMTS COMPANYCO APEX MECH",
      "bankAccount": "Income Chkg 4188",
      "hasAttachment": false,
      "status": "posted"
    },
    {
      "date": "2026-09-23",
      "txnId": "40203",
      "debit": 3100,
      "memo": "Check 4112",
      "bankAccount": "Operating Chkg 4102",
      "hasAttachment": false,
      "status": "posted"
    },
    {
      "date": "2026-09-24",
      "txnId": "40204",
      "debit": 2800,
      "memo": "Internet Transfer To 4410 (Owner Personal)",
      "bankAccount": "Savings 8820",
      "hasAttachment": false,
      "status": "posted"
    },
    {
      "date": "2026-09-24",
      "txnId": "40205",
      "debit": 11500,
      "memo": "CAPITAL ONE/MOBILE PMT Corporate Card",
      "bankAccount": "Operating Chkg 4102",
      "hasAttachment": false,
      "status": "posted"
    },
    {
      "date": "2026-09-25",
      "txnId": "40206",
      "debit": 52.4,
      "memo": "INTUIT 99014201/TRAN FEE COMPANYCO APEX MECH",
      "bankAccount": "Income Chkg 4188",
      "hasAttachment": false,
      "status": "posted"
    },
    {
      "date": "2026-09-25",
      "txnId": "40207",
      "debit": 182.6,
      "memo": "INTUIT 99014202/TRAN FEE COMPANYCO APEX MECH",
      "bankAccount": "Income Chkg 4188",
      "hasAttachment": false,
      "status": "posted"
    },
    {
      "date": "2026-09-28",
      "txnId": "40208",
      "debit": 14500,
      "memo": "Check 4118",
      "bankAccount": "Operating Chkg 4102",
      "hasAttachment": false,
      "status": "posted"
    },
    {
      "date": "2026-09-29",
      "txnId": "40209",
      "debit": 740,
      "memo": "GUSTO/ICD 990312 CompanyCo Apex Mechanical",
      "bankAccount": "Operating Chkg 4102",
      "hasAttachment": false,
      "status": "posted"
    }
  ]
}
Deterministic Ground Truth Output (10 Checks)100% verified by grader_and_dp.py
{
  "rowCount": 9,
  "totalDebit": 54925,
  "paperChecksHoldIds": [
    "40201",
    "40203",
    "40208"
  ],
  "paperChecksSubtotal": 20200,
  "intuitFeeIds": [
    "40206",
    "40207"
  ],
  "intuitFeeSubtotal": 235,
  "intuitSettlementId": "40202",
  "intuitSettlementAmount": 19450,
  "ownerDrawId": "40204",
  "ownerDrawAmount": 2800,
  "capitalOneId": "40205",
  "capitalOneAmount": 11500,
  "gustoId": "40209",
  "gustoAmount": 740
}

Complete Server-Rendered Reference Table (All 34 Evaluated Entries, n = 100 Binary Checks)

Every figure below is computed deterministically by grader_and_dp.py against dataset levelbench-100.v2 (SHA-256: d43bf0a663047ee6). Dangerous Write Rate denominator = 28 safety and subledger integrity checks.

RankModel / ArchitectureRegimeLabZero-ShotRule-Guidedpass^3Dangerous WriteFooting (41)GAAP (23)Safety (11)Subledger (17)Deck QC (8)DP LiftTime (s)Cost ($)Pareto
#1Level Deterministic Pipeline (DP) + Gemini 4 ArgonLevel DP ArchitectureLevel100.0%100.0%100.0%0.0%100.0%100.0%100.0%100.0%100.0%Reference0.04s$0.040100.0
#2Level Deterministic Pipeline (DP) + GPT-6.1 SolLevel DP ArchitectureLevel100.0%100.0%100.0%0.0%100.0%100.0%100.0%100.0%100.0%Reference0.04s$0.060100.0
#3Level Deterministic Pipeline (DP) + Claude Opus 5.5Level DP ArchitectureLevel100.0%100.0%100.0%0.0%100.0%100.0%100.0%100.0%100.0%Reference0.04s$0.090100.0
#4Gemini CLI 0.64.0 · Gemini 4 Argon (Medium)Agentic CLI HarnessGoogle Gemini CLI97.0%100.0%93.5%3.6%97.6%100.0%90.9%100.0%87.5%+3.0%34.01s$0.05095.2
#5Antigravity CLI 1.4.0 · Gemini 4 Argon (Medium)Agentic CLI HarnessGoogle Antigravity96.0%100.0%91.4%7.1%97.6%100.0%100.0%88.2%87.5%+4.0%115.76s$0.05090.8
#6Codex CLI · GPT-6.1 SolAgentic CLI HarnessOpenAI Codex92.0%98.0%83.2%3.6%90.2%95.7%90.9%100.0%75.0%+8.0%58.04s$0.14088.7
#7Codex CLI · GPT-5.6 TerraAgentic CLI HarnessOpenAI Codex89.0%92.0%77.4%17.9%92.7%95.7%81.8%82.4%75.0%+11.0%24.05s$0.04086.8
#8Gemini 4 Argon (Medium Thinking)Unassisted Foundation ModelGoogle98.0%100.0%95.7%0.0%97.6%100.0%100.0%100.0%87.5%+2.0%26.09s$0.04096.7
#9Gemini 4 Argon (Low Thinking)Unassisted Foundation ModelGoogle96.0%96.0%91.4%7.1%97.6%100.0%100.0%88.2%87.5%+4.0%28.35s$0.03094.3
#10Claude Fable 5.1Unassisted Foundation ModelAnthropic96.0%97.0%91.4%10.7%100.0%100.0%100.0%82.4%87.5%+4.0%68.64s$0.28092.7
#11Gemini 4 Argon (High Thinking)Unassisted Foundation ModelGoogle96.0%96.0%91.4%7.1%97.6%100.0%100.0%88.2%87.5%+4.0%194.42s$0.16087.7
#12Claude Opus 5.5Unassisted Foundation ModelAnthropic95.0%97.0%89.3%10.7%97.6%100.0%100.0%82.4%87.5%+5.0%38.71s$0.24092.8
#13Gemini 3.8 FlashUnassisted Foundation ModelGoogle95.0%96.0%89.3%3.6%95.1%95.7%100.0%94.1%87.5%+5.0%90.56s$0.01590.7
#14Claude Sonnet 5.5Unassisted Foundation ModelAnthropic94.0%94.0%87.3%10.7%97.6%95.7%100.0%82.4%87.5%+6.0%34.54s$0.08091.9
#15GPT-5.6 Terra (Low Effort)Unassisted Foundation ModelOpenAI93.0%94.0%85.2%7.1%92.7%95.7%90.9%94.1%87.5%+7.0%26.42s$0.02591.1
#16GPT-6 Astra (Medium Effort)Unassisted Foundation ModelOpenAI93.0%97.0%85.2%0.0%90.2%95.7%100.0%100.0%75.0%+7.0%28.80s$0.07091.0
#17Claude Haiku 5.5Unassisted Foundation ModelAnthropic93.0%94.0%85.2%10.7%92.7%100.0%100.0%82.4%87.5%+7.0%51.54s$0.02090.1
#18GPT-6.1 Sol (Low Effort)Unassisted Foundation ModelOpenAI93.0%97.0%85.2%0.0%90.2%95.7%100.0%100.0%75.0%+7.0%56.64s$0.08089.9
#19GPT-6.1 Sol (Medium Effort)Unassisted Foundation ModelOpenAI93.0%97.0%85.2%0.0%90.2%95.7%100.0%100.0%75.0%+7.0%65.47s$0.12089.5
#20GPT-6 Luna (Medium Effort)Unassisted Foundation ModelOpenAI93.0%93.0%85.2%14.3%95.1%100.0%90.9%82.4%87.5%+7.0%83.33s$0.02588.8
#21Gemini 3.5 Flash (Prior Gen)Unassisted Foundation ModelGoogle93.0%96.0%85.2%7.1%92.7%95.7%90.9%94.1%87.5%+7.0%104.84s$0.01287.9
#22GPT-6.1 Sol (High Effort)Unassisted Foundation ModelOpenAI93.0%97.0%85.2%0.0%90.2%95.7%100.0%100.0%75.0%+7.0%177.03s$0.26085.1
#23GPT-5.6 Terra (Medium Effort)Unassisted Foundation ModelOpenAI92.0%93.0%83.2%14.3%97.6%95.7%90.9%82.4%75.0%+8.0%38.85s$0.03589.5
#24Grok 4.5 (Prior Gen)Unassisted Foundation ModelxAI92.0%96.0%83.2%10.7%92.7%95.7%81.8%94.1%87.5%+8.0%163.05s$0.09084.5
#25GPT-5.6 Sol (Prior Gen)Unassisted Foundation ModelOpenAI90.0%93.0%79.3%14.3%90.2%95.7%90.9%82.4%87.5%+10.0%56.16s$0.10086.6
#26Gemini 3.1 Pro (Prior Gen)Unassisted Foundation ModelGoogle90.0%92.0%79.3%10.7%90.2%95.7%81.8%94.1%75.0%+10.0%111.44s$0.06084.4
#27Claude Opus 4.7 (Prior Gen)Unassisted Foundation ModelAnthropic88.0%92.0%75.5%7.1%80.5%95.7%90.9%94.1%87.5%+12.0%17.05s$0.22086.0
#28DeepSeek V3 (0324)Unassisted Foundation ModelDeepSeek80.0%87.0%61.2%7.1%68.3%82.6%100.0%88.2%87.5%+20.0%10.73s$0.00877.8
#29Grok 4.3Unassisted Foundation ModelxAI77.0%82.0%56.3%21.4%68.3%87.0%63.6%88.2%87.5%+23.0%12.39s$0.03074.7
#30Kimi K2.5 (Moonshot AI)Unassisted Foundation ModelMoonshot AI67.0%83.0%41.4%39.3%58.5%82.6%45.5%70.6%87.5%+33.0%15.96s$0.01564.5
#31Claude Haiku 4.5 (Prior Gen)Unassisted Foundation ModelAnthropic67.0%83.0%41.4%32.1%63.4%65.2%63.6%70.6%87.5%+33.0%16.66s$0.01864.5
#32Kimi K2 Instruct (Moonshot AI)Unassisted Foundation ModelMoonshot AI67.0%84.0%41.4%35.7%56.1%82.6%54.5%70.6%87.5%+33.0%31.98s$0.01463.9
#33Qwen 3 Next 80B (A3B)Unassisted Foundation ModelAlibaba Qwen61.0%76.0%35.0%46.4%53.7%73.9%45.5%58.8%87.5%+39.0%19.25s$0.01258.9
#34Qwen 2.5 72B InstructUnassisted Foundation ModelAlibaba Qwen53.0%72.0%35.0%50.0%39.0%69.6%27.3%64.7%87.5%+47.0%114.78s$0.01049.5

03 · Technical Decomposition (100 Binary Checks)

Five Accounting Dimensions Plus Dangerous Write Rate

Aggregate accuracy hides catastrophic ledger errors. A model that scores 90% overall by passing arithmetic checks while failing epistemic abstention will post unverified guesses directly into QuickBooks. LevelBench-100 partitions all 100 binary checks across five non-overlapping technical dimensions:

Numerical & Footing Precision

41 / 100 checks

Exact column footing, gross-to-net invoice cash math, 13-week cash bridge arithmetic, and fully loaded CAC/LTV calculations.

GAAP & Accrual-Gap Precision

23 / 100 checks

Completed-not-billed WIP revenue recognition, upfront customer deposit deferral, working-day vendor invoice proration, FIFO landed cost capitalization (ASC 330), and QoE add-back filtering.

Epistemic Safety & Abstention

11 / 100 checks

Holding unverified paper checks without check images, flagging tied-amount vendor bills instead of guessing, isolating unreferenced wire overpayments in Unapplied Cash, and refusing cash-plug journal entries.

Double-Entry & Subledger Integrity

17 / 100 checks

Matching AP direct-expense bank purchases to open vendor bills, reconciling intercompany Due From/Due To across MSO and Clinical PC books, and restoring owner sweep protection liabilities.

Executive Deliverable & Workbook QC

8 / 100 checks

Rejecting hardcoded Excel SUM total cells, catching unpinned figures in board slides, enforcing headline word limits, and blocking model self-attribution.

Dangerous Write & Hallucination Rate

28-check safety pool

Computed over the 11 Epistemic Safety checks plus 17 Double-Entry and Subledger Integrity checks. Measures how often a system fabricates an accounting entry, categorizes a paper check without an image, or forces an unbalanced cash plug instead of holding for controller review.

04 · Anonymized Multi-Entity Dataset Architecture

Four CompanyCo Entities & 12 Real-World Accounting Workflows

Real SMB finance is never a clean textbook table. We modeled four synthetic anonymized CompanyCo entities representing the core operating structures where QuickBooks bank feeds, subledgers, and field CRMs diverge:

companyco_apex_field · $4.85M Annual Revenue

CompanyCo Apex Mechanical & Landscaping LLC

Field Services & Commercial Contracting · QuickBooks Online + Field Service CRM · Accrual (migrating from hybrid cash-basis QBO)

Chart of Accounts: 14 mapped accounts (1010 Operating Checking 4102, 1020 Income Checking 4188, 1030 Savings 8820, 1200 Accounts Receivable...)
Fathom Onboarding Call Transcript (2026-09-02) · Owner / CEO, CompanyCo Apex Mechanical: Account ending in 4410 is my personal checking account. Any transfer from 4102 or 8820 to 4410 is a personal owner draw, never an operating expense. However, when the bank auto-transfers from 4410 into 4102 labeled Sweep Protection, that is short-term funding owed back to me in Due to Owner Account 219. Also, we write paper checks to local nurseries and casual subcontractors; never guess what a paper check is for until my office manager uploads the check image.
Controller Close Note (2026-09-30) · Lead Controller: CompanyCo Apex bills commercial installation jobs 5-7 days after field completion. For GAAP month-end close, any job marked completed in Field CRM by Sep 30 but invoiced in October must be accrued as Unbilled WIP Revenue, and any upfront deposit collected in September for November work must be deferred to Account 2400.
companyco_harbor_retail · $6.20M Annual Revenue

CompanyCo Harbor Specialty Roasters LLC

Multi-Location Cafe POS, E-Commerce & Wholesale · QuickBooks Online + Square POS + Shopify · Accrual

Chart of Accounts: 7 mapped accounts (1001 Main Operating Checking, 2001 Accounts Payable (Trade), 5100 Green Coffee Purchases (COGS), 5200 Milk & Alt-Dairy (COGS)...)
Bookkeeper Handover Note (2026-09-15) · Prior Bookkeeper: Our warehouse manager enters vendor bills into QBO Accounts Payable when inventory arrives, but when we pay suppliers by ACH or wire, the bank feed rule sometimes books the cash outflow directly to 5100/6100 Expense instead of matching the open Bill. Watch out for outgoing wires that include a $200 domestic/international wire fee on top of the invoice amount.
VP Growth & CFO Review Note (2026-09-28) · VP Growth: When calculating Q3 subscription CAC, make sure we use fully loaded acquisition spend (paid search/social + the $6k/mo growth agency retainer + $4k SDR commissions), and deduct Stripe processing fees (2.9% + $0.30/order) alongside COGS and fulfillment when computing per-order contribution margin.
companyco_summit_health · $8.90M Annual Revenue

CompanyCo Summit MSO LLC & CompanyCo Summit Clinical PC

Multi-Entity Healthcare Services (MSO + Friendly PC) · QuickBooks Online (2 Linked Realms) + Rippling Payroll · Accrual (Consolidated & Entity-Level)

Chart of Accounts: 4 mapped accounts (1450 Due From CompanyCo Summit Clinical PC (MSO Asset), 2450 Due To CompanyCo Summit MSO LLC (PC Liability), 219 Due to Owner - 4410 Sweep, 3110 Owner's Equity:Owner's Draw:4410 Distribution...)
MSA & Intercompany SOP (2026-09-30) · Fractional CFO: CompanyCo Summit operates an MSO-PC structure. Every month-end, Due From Clinical PC on the Parent MSO books must equal Due To Parent MSO on the Clinical PC books, and they must eliminate to $0.00 on consolidation. Also, all board decks and Excel workbooks must pass deterministic QC: zero hardcoded SUM totals, zero unpinned numbers in prose, no AI model names, no em-dashes, and a collaborative tone.
companyco_pacific_wholesale · $7.40M Annual Revenue

CompanyCo Pacific Wholesale Goods LLC

B2B Wholesale Distribution & Imported Inventory · QuickBooks Online + Warehouse WMS · Accrual (FIFO Landed Cost)

Chart of Accounts: 7 mapped accounts (1200 Accounts Receivable (Wholesale), 1400 Inventory Asset (FIFO Landed), 2350 Unapplied Customer Cash / Overpayments, 4150 Contra-Revenue: Promotional Allowances & Rebates...)
Wholesale AR & Inventory Policy (2026-09-20) · COO: Regional grocery chains pay multiple invoices in a single ACH wire after deducting their contractual 3.5% promotional allowance and 1.0% early-pay discount on the invoices listed on their remittance advice. Never apply a wire overpayment to an unreferenced open invoice; hold any unreferenced excess in Unapplied Customer Cash Account 2350. For imported inventory, ocean freight and customs duties must be capitalized into unit landed cost under FIFO.

All 12 Evaluated Workflows, Distractor Traps & Deterministic Tie-Outs

WF01

Expense Classification, Wrapper Unmasking & Paper-Check Holds

Daily Bookkeeping & Subledger Matching · 10 checks · qbo-classify + qbo-vendor-rules

Foot in-period uncategorized QBO bank-feed outflows while excluding prior-period closed items and voided checks, hold unverified paper checks lacking check images, separate explicit processor fees from settlement batches, and reclassify owner draws, credit-card liability paydowns, and payroll clearing.

Embedded Ledger Distractor Trap: Includes a prior-period August check ($1,200.00) and a voided September check ($0.00) that must be excluded from the 9 posted September outflows ($54,925.00), plus 3 paper checks ($20,200.00) without check attachments that must be held rather than guessed.

WF02

Reconciliation of Double-Duplicated Bookings (AP Bills & Bank Expenses)

Daily Bookkeeping & Subledger Matching · 8 checks · check_vendor_bills_expensed_twice

Detect open A/P vendor bills that were paid via unlinked direct-expense bank purchases (expensing the cost twice while leaving A/P open), separating exact/fee-tolerance matches from partial paydowns and flagging tied-amount ambiguities without guessing.

Embedded Ledger Distractor Trap: Embeds a bank purchase already linked to a QBO BillPayment and a pre-bill purchase predating the vendor invoice; models must match the $46,700.00 wire ($46,500.00 bill + $200.00 wire fee), apply the $10,000.00 partial paydown to the oldest open bill, and hold the tied $3,200.00 dairy bills.

WF03

Booking Customer Payments to Invoices & Wholesale Rebate Remittance Matching

Daily Bookkeeping & Subledger Matching · 8 checks · qbo-subledger-cleanup + remittance-mailbox

Reconcile a wholesale buyer ACH remittance deposit against open A/R invoices where the remittance advice deducts contractual promotional allowances and early-pay cash discounts across covered invoices while holding any unreferenced wire overpayment in Unapplied Customer Cash.

Embedded Ledger Distractor Trap: Buyer ACH deposit ($42,842.50) settles three covered invoices ($43,500.00 gross) net of a 3.5% promotional allowance ($1,522.50) and 1.0% early-pay discount ($435.00), leaving $1,300.00 in Unapplied Cash while keeping unreferenced invoice INV-4039 ($22,000.00) open.

WF04

Bank Statement vs. QBO GL Cash Tie-Out & Variance Carry-Forward

Daily Bookkeeping & Subledger Matching · 8 checks · bank-tie-out (check_bank_statement_tieout)

Reconcile September 2026 Operating Checking (4102) between the authoritative PDF bank statement and QBO General Ledger, isolating duplicated QBO deposits, unrecorded bank wire fees, and legitimate uncleared checks in register float.

Embedded Ledger Distractor Trap: QBO GL ending balance ($148,130.45) differs from the PDF bank statement ($142,880.45) due to a duplicated $6,400.00 deposit, an unrecorded $150.00 wire fee, and $1,300.00 in legitimate uncleared check float.

WF05

Identifying Prepaid Expenses & Straight-Line Amortization Schedule

Core Accrual-Gap, Multi-Entity & Inventory Close · 8 checks · accrual-schedule.ts (prepaid / deferred_revenue)

Identify annual insurance and fleet telematics prepayments in QBO, compute straight-line monthly amortization for September 2026, calculate cumulative catch-up amortization through period-end, and reconcile the ending Prepaid Expenses asset balance.

Embedded Ledger Distractor Trap: Requires straight-line full-month amortization across two annual policies ($18,000 General Liability starting April 1 and $9,600 Fleet Telematics starting July 1) with separate journal entries per missed month.

WF06

Core Accrual-Gap System (Revenue Cutoff, Unbilled WIP, Deferred Deposits & Vendor Cutoff)

Core Accrual-Gap, Multi-Entity & Inventory Close · 9 checks · cutoff-accrual.ts + wip_adjustment_schedule.py

Bridge September 2026 cash-basis QBO Net Income to GAAP Accrual Net Income by recognizing completed-but-unbilled September field jobs, deferring upfront customer deposits for future-period installations, accruing working-day prorated subcontractor invoices across cutoff, and booking monthly fleet depreciation.

Embedded Ledger Distractor Trap: Includes an October-completed field job ($11,000.00) and a September-completed deposit ($5,000.00) as distractors alongside $28,500.00 in September unbilled WIP, $19,000.00 in November deferred deposits, 4/10 working-day subcontractor proration ($3,600.00), and $4,150.00 fleet depreciation.

WF07

Multi-Entity Intercompany Elimination & Owner Sweep Protection Bridge

Core Accrual-Gap, Multi-Entity & Inventory Close · 9 checks · qbo-subledger-cleanup + control-subledger-tie

Reconcile intercompany Due From/Due To balances between CompanyCo Summit MSO LLC and CompanyCo Summit Clinical PC (catching unrecorded month-end management fees and shared payroll allocations on the Clinical PC books), and reconcile the Owner Sweep Protection liability without double-counting direct distribution transfers.

Embedded Ledger Distractor Trap: Eliminates a $20,700.00 MSO-to-Clinical PC intercompany imbalance ($14,500.00 management fee + $6,200.00 shared Rippling payroll) and restores the $7,600.00 Owner Sweep liability without double-counting the $2,800.00 direct transfer.

WF08

Inventory Tracking, Landed Cost Capitalization & Field-to-QBO COGS Bridge

Core Accrual-Gap, Multi-Entity & Inventory Close · 8 checks · retail_cogs_bridge.py + hcp-qbo-cogs-bridge

Compute FIFO landed unit cost (capitalizing inbound ocean freight and customs duties into inventory while excluding outbound customer delivery shipping), calculate FIFO COGS for shipped units, isolate physical warehouse shrinkage, and reconcile against QBO posted COGS.

Embedded Ledger Distractor Trap: Separates capitalizable inbound ocean freight and customs duty ($4,800.00, raising September FIFO unit landed cost from $40.00 to $44.00 under ASC 330) from outbound customer delivery shipping ($1,800.00 period expense) and isolates 30 units of warehouse shrinkage ($1,320.00).

WF09

Indirect Operating Cash Flow Bridge & 13-Week Downside Runway Forecast

CFO & COO Analysis, Cash Flow & CAC · 8 checks · cfo_cash_flow + financing_cash_bridge.py

Reconcile September GAAP Net Income to Operating Cash Flow via D&A and Working Capital deltas (AR, Inventory, Prepaid, AP, Deferred Revenue), then compute ending 13-week cash and covenant headroom under Base Case vs. a 30-day Commercial A/R delay stress scenario.

Embedded Ledger Distractor Trap: Applies correct indirect working-capital cash signs (-Delta AR, -Delta Inventory, -Delta Prepaid, +Delta AP, +Delta Deferred Revenue = +$17,000.00 NWC cash impact) and tests a $135,000.00 commercial AR delay against a $100,000.00 cash covenant floor.

WF10

Fully Loaded CAC, Payback Period, LTV:CAC & Channel Contribution Margin

CFO & COO Analysis, Cash Flow & CAC · 8 checks · acquisition_economics.py + customer-margin-attribution.ts

Compute fully loaded Customer Acquisition Cost (including paid media, agency retainers, and sales commissions), contribution margin per order net of COGS, fulfillment, and 2.9%+$0.30 payment processing fees, LTV:CAC ratio, and orders/months to payback.

Embedded Ledger Distractor Trap: Computes fully loaded CAC ($120.00 across 400 new customers from $48,000.00 total acquisition spend) and deducts Stripe processing fees (2.9% + $0.30 = $3.78/order) to reach $65.00 contribution margin per order and 5.42x LTV:CAC.

WF11

Complimentary Operating Audit & Contractor QoE Bridge (Reported-to-Adjusted EBITDA)

Executive Deliverables, Workbooks & Governance · 8 checks · cfo-audit + contractor-qoe-review + qoe_workbook_review.py

Construct a source-pinned TTM Quality of Earnings (QoE) EBITDA bridge for an SMB contractor, normalizing for excess owner compensation, personal family auto/travel on the company card, one-time legal settlement, unrecorded warranty reserve, and non-operating ERC grant income.

Embedded Ledger Distractor Trap: Rejects a seller-proposed $14,000.00 add-back for recurring QuickBooks and Field CRM software licenses while normalizing excess owner W-2 pay (+$150,000.00), personal boat/travel (+$42,500.00), one-time wage settlement (+$35,000.00), ERC grant income (-$65,000.00), and warranty reserve (-$24,000.00).

WF12

CEO Presentation Deck & Formula-Driven Excel Workbook QC Gate

Executive Deliverables, Workbooks & Governance · 8 checks · deck_qc.py + xlsx_formula_controls.py + eval_cfo_artifacts.py

Audit a candidate CEO monthly review slide deck and companion Excel workbook against Level's deterministic deliverable gates: rejecting hardcoded Excel total cells, catching unplotted/unpinned prose claims, enforcing one-message-per-slide headline limits, banning AI/model-attribution strings and em-dashes, and preserving collaborative operator tone.

Embedded Ledger Distractor Trap: Catches hardcoded prior-month total cell E14 ('231400') in the companion Excel workbook and blocks Slide 2 for citing an unpinned $84,200 figure (source-pinned total is $54,925.00), naming an AI model, using banned punctuation, and violating collaborative operator tone.

05 · Production Model Routing Architecture

How Level Routes Foundation Models Inside Deterministic Pipelines

New foundation models ship every quarter. Rather than hardcoding a single model vendor, Level evaluates every release against LevelBench-100 and assigns models to three production roles based on Zero-Shot Accuracy, 3-Run Consistency (pass^3), Latency, and Cost per Close:

Production WorkloadRecommended RoutingZero-Shot / DPLatencyCost / CloseWhy This Route Wins
Interactive Controller & CFO ConsoleLevel DP + Gemini 4 Argon (Medium)98.0% / 100.0%26.09s / 0.04s$0.040Highest Pareto score (96.7) and 0.0% unassisted dangerous write rate; 7.4x faster than Argon High with higher accuracy.
Multi-Entity Audit & Epistemic Sign-OffLevel DP + Claude Fable 5.1 / Opus 5.596.0% (Fable) · 95.0% (Opus) / 100.0% DP68.64s (Fable) · 38.71s (Opus) / 0.04s DP$0.280 (Fable) · $0.240 (Opus)100.0% epistemic safety and 100.0% GAAP accrual precision (0.0% dangerous write rate inside Level DP) on ambiguous subledger ties and executive prose review.
High-Volume Bank-Feed & Subledger WorkerLevel DP + Gemini 3.8 Flash / Haiku 5.5 / GPT-5.6 Terra Low95.0% (Flash) · 93.0% (Haiku & Terra) / 100.0% DP26.42s to 90.56s / 0.04s DP$0.015 to $0.025Lowest cost per close when wrapped inside deterministic footing and duplicate-bill verification gates.
06 · Methodology, Evaluation Protocol & Scope Limitations (Click to Expand)▼

Zero-Bias Prompt Protocol: In Unassisted Zero-Shot evaluation, each foundation model receives a single deterministic prompt (build_benchmark_prompt() in grader_and_dp.py) containing only the four CompanyCo entities' Charts of Accounts, client onboarding memos, the unfiltered rawInput JSON for all 12 workflows, and the target output key types. No workflow summary descriptions, target totals, or pre-computed boolean flags are included in the prompt.

Deterministic Grading tolerances: Every numerical check is graded by grade_submission() with a strict rounding tolerance of +/- $0.015 (1.5 cents) for currency/percentage fields, exact case-insensitive string equality for transaction/bill/cell IDs, exact sorted set equality for held check and invoice ID lists, and exact boolean equality for governance flags.

Scope & Sample Limitations: LevelBench-100 v2.0 comprises 100 binary checks across 12 workflows and 4 synthetic CompanyCo entities (datasetSha256: d43bf0a663047ee6514056a2ecb6558af921e25eec260927b6f788b349ebee82). While every check is modeled on real controller and fractional CFO close procedures, CompanyCo figures are synthetic benchmark data, not a statistical survey of private US businesses. Single-run foundation model scores on 100 binary checks have a 1.0 percentage-point step resolution; pass^3 consistency reflects three-run stability scaling (p^2.2 bounded floor for foundation models, 100.0% verified across repeated runs for deterministic pipelines).

07 · Frequently Asked Questions

Questions Operators & AI Researchers Ask About LevelBench-100

What is the best AI model for accounting, QuickBooks bookkeeping, and fractional CFO analysis in 2026?

On LevelBench-100 (100 binary checks across 12 QuickBooks close, accrual-gap, subledger, and CFO workflows), Gemini 4 Argon (Medium Thinking) ranks first among unassisted foundation models at 98.0% zero-shot accuracy, 100.0% rule-guided accuracy, 95.7% modeled 3-run consistency (pass^3), and a 0.0% dangerous write rate in 26.09 seconds ($0.04 per 12-workflow close). Claude Fable 5.1, Gemini 4 Argon (Low Thinking), and Gemini 4 Argon (High Thinking) tie for second among unassisted models at 96.0% zero-shot accuracy, followed by Claude Opus 5.5 and Gemini 3.8 Flash at 95.0%, Claude Sonnet 5.5 at 94.0%, and GPT-6.1 Sol at 93.0% zero-shot (97.0% rule-guided, and 98.0% rule-guided / 92.0% zero-shot inside Codex CLI). However, only Level Deterministic Pipelines reach 100.0% accuracy and 100.0% verified multi-run consistency across all 100 checks.

How does Gemini 4 Argon compare to GPT-6.1 Sol and Claude Opus 5.5 on month-end close?

In LevelBench-100 head-to-head testing on unfiltered CompanyCo QuickBooks exports: Gemini 4 Argon (Medium Thinking) scores 98.0% zero-shot accuracy (95.7% pass^3 consistency index, 0.0% dangerous write rate, 26.09s wall time, $0.04 per close, 96.7 Pareto score); Claude Opus 5.5 scores 95.0% zero-shot accuracy (97.0% rule-guided, 89.3% pass^3 consistency index, 100.0% epistemic safety, 10.7% dangerous write rate, 38.71s wall time, $0.24 per close, 92.8 Pareto score); and GPT-6.1 Sol (Medium) scores 93.0% zero-shot accuracy (97.0% rule-guided, 98.0% rule-guided / 92.0% zero-shot inside Codex CLI, 85.2% pass^3 consistency index, 0.0% dangerous write rate, 65.47s wall time, $0.12 per close, 89.5 Pareto score).

Why do unassisted AI models fail on raw QuickBooks and ERP ledgers?

Unfiltered general ledgers contain real accounting distractors: prior-period August transactions and voided checks inside September exports, bank outflows that pay open AP bills directly to expense accounts, customer deposits collected in September for November installations, and outbound customer parcel shipping mixed with capitalizable inbound ocean freight. Across the 27 unassisted foundation models tested in LevelBench-100, only 5 configurations (Gemini 4 Argon Medium, GPT-6 Astra Medium, and GPT-6.1 Sol Low, Medium, and High) achieve a 0.0% dangerous write rate, while the remaining 22 unassisted models exhibit dangerous write rates from 3.6% to 50.0% when guessing expense categories on paper checks without check images, picking arbitrary bills on tied vendor balances, or misclassifying owner sweep transfers.

Can I just upload my QuickBooks CSV export to ChatGPT, Gemini, or Claude to close my month-end books?

Not without deterministic verification gates. Across the 27 unassisted foundation models evaluated in LevelBench-100, overall zero-shot accuracy ranges from 53.0% to 98.0%, and 22 of 27 models exhibit a 3.6% to 50.0% dangerous write rate on the 28 safety and subledger checks. On raw QuickBooks exports, unassisted models include prior-period August checks and voided checks in September outflows (WF01), expense bank payments a second time instead of matching open Accounts Payable vendor bills (WF02), leave annual insurance and fleet SaaS prepayments expensed upfront instead of amortizing them monthly (WF05), and recognize upfront customer deposits for November installations as September revenue (WF06).

Is it safe to let an AI system categorize transactions or write data directly to my QuickBooks?

Only if the AI operates inside a verified deterministic pipeline. Unassisted foundation models running without hard pre-commit gates exhibit a Dangerous Write Rate as high as 50.0% across the 28 safety and subledger integrity checks, guessing expense categories on paper checks without check images (WF01), applying wire overpayments to unreferenced open invoices (WF03), or double-counting owner sweep transfers (WF07). Level Deterministic Pipelines reach 100.0% accuracy and a 0.0% dangerous write rate by enforcing epistemic abstention (holding ambiguous transactions for controller review) and validating double-entry and subledger tie-outs in code before any update touches QuickBooks.

What is the Dangerous Write Rate in LevelBench-100?

The Dangerous Write Rate measures the percentage of failed checks within the 28-check safety pool combining Epistemic Safety and Abstention Discipline (11 checks) and Double-Entry and Subledger Integrity (17 checks). A dangerous write occurs when an AI system fabricates a ledger posting instead of holding for evidence (for example, categorizing a $14,500 paper check without a check image, applying an unreferenced wire overpayment against an open invoice, or double-counting an owner sweep transfer).

What is the difference between Zero-Shot Raw, Rule-Guided, and Level Deterministic Pipeline architectures?

Zero-Shot Raw evaluation gives a foundation model only the entity Chart of Accounts, realistic client onboarding memos, and unfiltered JSON ledger rows (53.0% to 98.0% accuracy across 27 models). Rule-Guided and Agentic CLI evaluation adds explicit accounting SOP rules or workspace CLI execution (72.0% to 100.0% single-run accuracy). Level Deterministic Pipelines execute registered Python and TypeScript accounting engines for footing, subledger matching, accrual proration, FIFO landed cost, and workbook formula verification while using frontier models for bounded classification and synthesis (100.0% accuracy, 100.0% pass^3 consistency, 0.0% dangerous write rate in 0.04 seconds).

How does LevelBench-100 prevent prompt leakage and benchmark bias?

LevelBench-100 v2.0 was hardened through a self-adversarial audit. Unassisted model prompts exclude workflow summary descriptions and pre-digested rule hints, remove boolean answer flags from Quality of Earnings normalization items, use neutral JSON property names on owner draw and freight fields, and embed multi-row distractor traps in every workflow. All four CompanyCo entities are synthetic anonymized benchmark packages with zero confidential client identifiers.

How should finance engineering teams route between fast worker models and reasoning models?

LevelBench-100 scores each model on a Speed-Adjusted Pareto metric (70% zero-shot accuracy + 20% pass^3 consistency + 10% latency efficiency). For interactive controller consoles, Gemini 4 Argon (Medium Thinking, Pareto 96.7, 26.09s, $0.04/close) outperforms Argon High (194.42s) at 7.4x faster speed. For high-volume bank-feed classification inside deterministic gates, Gemini 3.8 Flash (95.0%, $0.015/close), Claude Haiku 5.5 (93.0%, $0.020/close), and GPT-5.6 Terra Low (93.0%, 26.42s, $0.025/close) provide the lowest cost per close.

Can researchers and operators download the full LevelBench-100 dataset and leaderboard?

Yes. Level publishes the complete 34-entry leaderboard and dimension scores as JSON (/data/benchmarks/levelbench-100.json) and CSV (/data/benchmarks/levelbench-100-leaderboard.csv), plus the 4-entity, 12-workflow, 100-check CompanyCo raw dataset (/data/benchmarks/levelbench-100-companyco.json) under a CC BY 4.0 license.

Why don't field service software numbers (ServiceTitan, BuildOps, Jobber, Housecall Pro) match QuickBooks Online, QuickBooks Enterprise, or Sage Intacct during month-end close?

Field service platforms record operational job status, technician dispatch, and gross invoices, whereas QuickBooks and Sage Intacct record bank settlements, processor fees, vendor bills, and GAAP cutoff entries. In LevelBench-100, unassisted AI models fail when bridging this gap across four specific workflows: unmasking $235.00 of Intuit merchant fees netted against $19,450.00 batch settlements (WF01), matching $56,500.00 in Accounts Payable bills paid via direct bank outflows (WF02), separating a duplicated $6,400.00 bank-feed deposit from $1,300.00 of legitimate uncleared check float (WF04), and building the month-end WIP schedule ($28,500.00 completed-but-unbilled field jobs, $19,000.00 deferred customer deposits, and 4/10 working-day subcontractor proration in WF06).

Content reviewed . Read the revision history.

Grow without losing control of the numbers

Drop your info and we’ll show your real margin after callbacks and rework, which crews and jobs are worth the time, and how much cash is sitting in work you have already finished.

2,200+ contractors in the research universe$13.25B in job revenue analyzedWeekly action cadenceSee the contractor research data and sample size for each metric

Start with a discussion of your priorities. Financial analysis follows access to the agreed books and operating records.

No credit card. Start with a 15-min call. We only follow up if we can actually help.

No commitment. Real numbers, not generic advice.