World records

Three finance world records.
100% on each.

Zero misses across 386 graded questions.

One Jump’s agent finds the right figure from the right document and period, runs the calculation, and cites the page behind every number.

The records

Three benchmarks. Three perfect scores.

Three different kinds of financial evidence, each with a public result you can open and check.

Record 01

OfficeQA Full

Databricks

246 / 246100%

#1 on its public AgentBeats leaderboard

Financial research across 89K+ pages of U.S. Treasury Bulletins: find the right tables, read them correctly, and complete multi-step calculations, statistics and specified forecasts.

What the agent reads
89K+ pages of U.S. Treasury Bulletins, 1939–2025
Questions
246 questions
Grading
Numerical reference, zero tolerance
How it was run
Run through the leaderboard’s own submission pipeline and scored by its evaluator (third-party). Next-best public agent: 46.3%.
Recorded
September 13, 2026

Record 02

OfficeQA Pro v2

Databricks question set

90 / 90100%

On a One Jump-maintained AgentBeats leaderboard

A harder corpus: Combined Statements and Congressional receipts PDFs spanning 211 years, multi-document questions, and exact-reference grading with zero tolerance.

What the agent reads
1,435 federal receipts and expenditures PDFs, 1793–2024
Questions
90 questions, 10 shards
Grading
Exact reference, zero tolerance
How it was run
One Jump maintains this AgentBeats leaderboard and its evaluator. The result commit is public.
Recorded
September 27, 2026

Record 03

Finance Agent Benchmark

Vals AI tasks, community AgentBeats edition

50 / 50100%

#1 on its AgentBeats leaderboard

Expert-written equity-research tasks over SEC filings: guidance beat-or-miss, GAAP to non-GAAP adjustments, financial modeling and market analysis.

What the agent reads
SEC EDGAR filings and XBRL facts
Questions
50 tasks, 9 categories
Grading
Against the experts’ reference answers
How it was run
Reviewed and merged by the leaderboard’s maintainer.
Recorded
September 25, 2026

Side by side

BenchmarkPublisherWhat the agent readsQuestionsGradingResultRecorded
OfficeQA FullDatabricks89K+ pages of U.S. Treasury Bulletins, 1939–2025246 questionsNumerical reference, zero tolerance246 / 246 · 100%Sep 13, 2026
OfficeQA Pro v2Databricks question set1,435 federal receipts and expenditures PDFs, 1793–202490 questions, 10 shardsExact reference, zero tolerance90 / 90 · 100%Sep 27, 2026
Finance Agent BenchmarkVals AI tasks, community AgentBeats editionSEC EDGAR filings and XBRL facts50 tasks, 9 categoriesAgainst the experts’ reference answers50 / 50 · 100%Sep 25, 2026

Scores and leaderboard positions as recorded on the dates shown; live rankings can change as new agents are submitted.

Why it’s hard

In finance, almost right is wrong.

A figure from the wrong table, period or revision looks just as plausible as the right one. Getting it right means getting all of these right, every time.

01

The right table

Nested headers, footnotes and units that change mid-table. The number has to come from the row and column the question means.

02

The right period

Fiscal versus calendar years, quarters versus year-to-date, and periods that shift when a company changes its year end.

03

The right revision

Restated figures, revised releases and amended filings. The latest number isn’t always the one the question asks for.

04

The right math

Multi-step calculations on top of all that, where one wrong input carries through to a wrong answer.

What it can do

What it can do for your team.

The same agent behind these records, working on your documents and your questions.

01

Cited answers

The page and table behind every number, so your reviewers check the work instead of redoing it.

02

The right figure, from the right source

Finds the figure the question asks for, from the right document and the right period, on the basis you need.

03

Runs the analysis

Every step shown and re-runnable:

  • Ratios and covenant tests
  • Growth and variance
  • Regression, exponential smoothing and ARIMA
  • HP-filter trends and correlation
  • Value at risk and expected shortfall
  • Inflation and currency adjustment

04

Flags conflicts

When documents, periods or versions disagree, it shows both, says which figure it used and why.

05

Reads what you have

Filings, statements, statistical releases, and scanned and historical archives, read as they are.

06

Structured output

Answer, sources, values used, calculation and conflicts, structured for your reviewers, your systems and your agents.

07

Runs where you need it

Hosted by One Jump for the fastest start. Enterprise deployments in your own cloud account, delivered as signed, encrypted releases, or on-premises, scoped with you, with the data flow documented before you sign.

Deployment and data-flow details

Proven on your documents

Benchmarks are public. Your test is yours.

Every license and engagement is measured on the customer’s own documents and on questions their team has already answered. Pass criteria are agreed first (figures right within an agreed tolerance, every citation checked against its page, known conflicts flagged) and results are reported as measured, misses included.

How acceptance testing works

Pricing

Two ways to license.

Early Adopter is a self-serve order for one team on hosted documents. Enterprise is contract-first: firm-wide, in your environment, or inside your platform.

ENTERPRISE

Contract-first

Enterprise license

Firm-wide, in your environment, or inside your platform.

Twelve-month license$1,000,000
  • Twelve months of development support with a named technical lead
  • Deployment in your cloud account, on-premises or hosted, scoped with you
  • Custom integrations with your systems
  • Documentation for your model-risk and AI-governance review
  • Acceptance test on your own documents, agreed in writing before you sign.
  • Paid by invoice, ACH or wire in milestones, never by card
Plan an Enterprise deployment

Keys are issued to verified businesses after a signed order. No free trial and no low-cost tier — here’s why.

Usage limits and how tokens are counted

Put it to work on your documents.

Bring the questions your team has already answered. We’ll show you how the test is scored before you commit.