Product

Polygon: acceptance testing of AI systems

The test bench answers two questions in a single run: how many concurrent users the AI assistant can handle and whether its responses degrade under load. The result is a report with a PASS/FAIL verdict based on pre-agreed criteria.

Hero image for the page Polygon — AI systems acceptance testing stand

What is this product

The "Testbed" is developed by Invent LLC, a company within the RESTART Group. The testbed is deployed within the customer's environment and evaluates the AI system along two axes simultaneously: whether it can handle the specified load and whether responses remain correct as the load increases. Both axes are measured using the same traffic within a single test run.

The main output is not a web page with charts, but a test report in PDF and DOCX formats: containing the test program and methodology, test conditions, results for each axis, pass/fail verdicts for each acceptance criterion, and a signature block. The document is intended to be included in the delivery acceptance certificate package.

What one can feel

RAG assistants

Corporate knowledge bases and document assistants are evaluated both on response time and whether the responses are grounded in sources as the number of queries increases.

Support chatbots

Customer-facing and internal bots ahead of the seasonal peak: how many concurrent requests can they handle, and do they start making up facts?

LLM API and inference services

Own inference pipeline: throughput, queue formation time, behavior upon node failure.

Assistants in business processes

Scenarios where the model's response continues into the process—there, an error is costlier and is verified separately.

How are acceptance criteria defined

The test profile is agreed upon prior to execution and becomes the single source of truth: what is recorded in it is quoted verbatim in the protocol. Typically, four groups of parameters are captured.

What are we settingParameter examplesWho is involved in the approval
Load profileSingle requests, one-time burst, stepwise ramp-up, background load with pauses, random arrivals within a time windowAI Product Owner, Site Reliability Engineer
Validation datasetA set of real questions and reference answers, sample size for quality assessmentProcess owner, subject matter experts
Acceptance thresholdsSuccess rate, P95 response time, allowable number of failures, minimum quality score, number of users without queueAcceptance committee, test customer
Run boundariesEnvironment, system version, execution window, access to infrastructure metricsIT Operations, Security Service

The infrastructure metrics section is optional: even without access to containers and GPUs, the test stand still loads the system via HTTP and delivers a verdict based on the remaining criteria.

What does the customer get?

  • Test protocol in PDF and DOCX formats based on an editable template: title page, program and methodology, conditions, load and quality results, verdict for each criterion, and a signature block.
  • Response for each acceptance criterion in the format "threshold — actual — PASS or FAIL", without general statements such as "there are comments".
  • Example applications: question, system response, evaluation, and textual explanation of the verdict — the evaluation logic can be manually rechecked.
  • Behavior under load data: queue formation stage, concurrency level at which rejections begin, queue and inference engine memory state, if metrics access is provided.

How it differs from audit and laboratory

The group's three proposals are easily confused—they address different questions and are applied at different stages.

FormatWhat question does it answer?What's the output
Secure AI auditWhat risks does the AI loop architecture entail, and how are data and access managed within it?Risk register, architecture review, remediation roadmap
Information Security LaboratoryHow a specific security control or product behaves within the customer's architecturePilot protocol, compatibility matrix, implementation requirements
PolygonDoes the AI system pass acceptance testing according to agreed-upon load and quality criteria?Test protocol with PASS/FAIL verdict for the acceptance certificate

Bench tests are not certification and do not confirm compliance with regulatory requirements—they confirm fulfillment of the criteria agreed upon by the acceptance parties.

How to get started

1

Application and call

We discuss the system type, available traffic of real queries, and access to infrastructure metrics.

2

Test Profile

We'll align the load profile, reference dataset, and acceptance thresholds with your system.

3

Run in your environment

The test bench stresses the target and simultaneously evaluates the response sample using a local judge.

4

Protocol

You receive a document with a verdict for each criterion, ready to be included in the set with the report.

Request a demo or pilot

We will show the product on your environment and offer a pilot format with a clear output result.

Request format
AI assistant
Hello! I am an AI assistant at RESTART. I’ll help you find the right section of the site, answer questions about services, licenses, partnerships, contacts, or formulate an appeal to the sales department.