01
The answer
Value, unit, currency and scale, period and basis (as reported or restated).
Proof
Don’t take our word for it.
Take your documents’.
The strongest evidence we can give you is a test on your own documents, agreed before you commit. Here’s how that works, the third-party result we lead with, and what every answer hands you to check.
01
Acceptance-tested on your own documents
- Bring your answer key. Bring 50–200 questions your team has already answered, with the documents behind them. Your final models, closed-deal files and reviewed reports are the answer key.
- Agree the pass criteria first. Figures correct within an agreed tolerance; every citation checked against its page; known conflicts flagged; time per question against your current process.
- Run it on your documents. In your environment or ours.
- Get the results as measured. Misses included.
- Pilots: Acceptance-tested on your own documents before you pay the balance. Fixed fee, capped.
- Enterprise: Acceptance test on your own documents, agreed in writing before you sign.
- Early Adopter: Your acceptance plan is agreed in week one, and six months of development support are aimed at passing it.
Fixed fees are capped. If a pilot misses its target, you get the full results report and the test harness, and you decide whether to extend.
02
A third-party-run result
#1NetArena MALT on AgentBeats
100% / 100%Correctness and safety over 2,500 queries
0.056 sFastest of the perfect runs
NetArena tests agents on Kubernetes network-policy repair in a live microservices environment. Each task injects a connectivity fault, and the agent has to fix it without opening traffic that should stay blocked. One Jump’s agent completed 2,500 dynamically generated tasks with 100% correctness and 100% safety and holds the #1 position, the fastest of the perfect runs. Other entries have also reached 100%; the leaderboard ranks perfect runs by speed.
Record page Technical report Leaderboard Agent profile Result commit
Recorded September 21, 2026 · standing checked October 6, 2026.
03
What every answer hands you to check
02
Sources
Document, page and table for the answer and for every input.
03
Values used
Each input figure, with its own source.
04
Calculation trace
The steps from inputs to answer, re-runnable.
05
Conflicts
Where documents, periods or versions disagree. Both sources are shown, along with which one was used and why.
06
Version and request ID
For your audit trail and support tickets.
04
Representative work
Anonymized
Aerial public safety
Detection networks developed inside a drone-flight simulator, then prepared for deployment on real aircraft for earlier wildfire detection.
Anonymized
Healthcare efficacy modeling
A predictive model combining pathogen kinetics, room geometry and exposure time, indicating about 80% lower cumulative exposure for continuous disinfection than spot-cleaning.
Anonymized
Quantitative research systems
A nightly research loop that tests one change at a time against the production version and logs every keep-or-revert decision.
Client work is anonymized. Illustrative scenarios elsewhere on the site are labeled as such.
05
How the acceptance test is scored
These are the measures we agree with you before the run. You set the tolerance and the question set; we report against them.
| Metric | Definition |
|---|---|
| Figure accuracy | Within the agreed tolerance |
| Citation validity | The cited page contains the value |
| Conflict recall | Known discrepancies flagged ÷ known |
| Time per question | Against the current process |
Results are reported as measured, misses included.
Plan your acceptance test.
Bring the questions your team has already answered. We agree how to score them before you commit.
Prefer to see the test plan first? Start with your acceptance-test plan