BoxVia

Technical Pilot

Independent evidence-based review of AI system behavior.

Find out what the available evidence actually establishes about one defined scenario in your AI agent, multi-agent workflow or connected tool.

Receive a written assessment of observed behavior, supporting evidence, unresolved questions and limits of the conclusion.

One scenario. One defined system version. One independent report.

Paid engagements from 250 USD.

Final scope, price and delivery date are agreed before work begins.

Start request

Choose your review format

EVIDENCE REVIEW

Have records but no test environment?

We examine existing logs, policies, traces or reports to establish what they support for one scenario.

From 250 USD with a usable evidence package.

SCENARIO PILOT

Have a reproducible scenario or test environment?

We agree on one test, examine the observed behavior and assess the supporting evidence.

Price agreed after defining scope and access.

Both formats produce a bounded written result. Production access is not assumed.

Example runtime scenarios

Delegation after a state change

An agent receives delegated authority.

The relevant state or policy changes before execution.

Does the execution boundary re-evaluate authority against the current state, or rely on the earlier delegation?

Multi-agent handoff

One agent proposes an action and another executes it.

Can the evidence establish who proposed, evaluated, authorized and executed the action, including any scope changes?

Retries and duplicate actions

A timeout or retry triggers a second tool call.

Does the system prevent unintended duplicate execution, and can the observed result be independently established?

Insufficient or conflicting evidence

Execution records are missing, inconsistent or incomplete.

Can the assessment distinguish an established result from one that must remain UNESTABLISHED?

These are illustrative questions, not claims that any particular system has been tested or that BoxVia operates its enforcement mechanism.

How independent verification works

The pilot does not simply reproduce the enforcement implementation or assume its output is correct.

The assessment does not assume that the system owner's logs, policy decisions or test results are complete or correct.

We distinguish three elements:

  1. Authorized boundary: the applicable policy, scope, delegation and state, as documented by the system owner.
  2. Observed execution: the proposed action, policy decision, tool invocation, outcome and relevant state transitions.
  3. Independent evidence review: whether available records and agreed tests support a bounded finding about the observed behavior.

The system owner remains responsible for defining policy, granting authority and operating enforcement.

BoxVia does not become the enterprise decision authority or an inline execution gate.

Examples from completed research

WEBAGENT.AI - What does observed AI behavior establish?

A completed comparative research pilot examined a public AI support agent against frozen public statements and observed answers.

Selected factual answers were supported, while broader claims such as 24/7 availability and the origin of answers were not established from the limited observations.

The independent ANDEKS result was PARTIALLY ESTABLISHED / MATERIAL LIMITATIONS. The later comparative record was mutually reviewed without changing either frozen result.

OMNIX - Does a successful verification status mean full verification?

In a bounded independent reproduction of an early public verifier, a synthetic unsigned receipt could return fully_verified=true when the implemented hash, authority-budget, status and expiry checks passed.

This establishes the behavior of that specific verification logic, not cryptographic authenticity of a real receipt or end-to-end verification of the wider system.

R1–R6 were reproduced on synthetic inputs; later verification paths were not established end-to-end.

Historical research illustrations only, not paid customer engagements, certification results or evidence of current production behavior.

What you receive

One concise written report covering the assessment question, system version and scope, method or evidence reviewed, observed findings, evidence references, gaps, conclusion and limitations.

Findings may establish a bounded observation, identify a demonstrated deviation or remain UNESTABLISHED when evidence is insufficient.

A reproduced failure does not automatically establish its root cause. A successful tool response does not by itself establish valid authority or effective enforcement.

Pricing and process

From 250 USD. The starting price applies to a narrowly scoped review of one scenario using an accessible, usable evidence package.

Test execution, multiple evidence sources, complex correlation, repeated runs or expanded architectural assessment may require additional work.

  1. Send one precise question, system/version information and public or non-sensitive materials.
  2. We agree in writing on feasibility, review format, scope, evidence requirements, fixed price and delivery date.
  3. Work begins after payment and any required access and permissions are agreed.

Start request

Scope and independence: A pilot is a paid, scenario-specific technical examination, not certification, formal audit assurance, a legal opinion, guaranteed compliance or production approval.

Findings apply only to the agreed version, scenario and evidence. The system owner retains responsibility for policy, enforcement, deployment and risk decisions.

No claim of deployed ARE/ERA enforcement or formal ANDEKS normative assessment is implied.

Request a technical pilot

Before you submit

For this initial request, briefly describe the materials you have or share public links.

Do not include confidential documents, internal logs, passwords, API keys, private access links or sensitive personal data.

If we need additional materials or access, we will agree on the scope of work and how to share them before asking you to send anything.

Read our confidentiality and data-handling terms

Use an email address where you can receive replies about this request.
See examples of system names
  • A named agent framework, such as LangGraph
  • A custom support agent that uses a ticketing tool
  • A multi-agent workflow with a coordinator and worker agents
See examples of versions and commits
  • A release number, such as 2.4.1
  • A Git tag, such as v2.4.1, or a commit ID
  • A build ID for the test environment
What type of review do you need?
See example pilot scenarios
  • Can a delegated agent use a tool outside its assigned permissions?
  • Does the system preserve an approval requirement when a task is handed to another agent?
  • Can the system show which agent changed a record and which rule allowed it?
See examples of expected behavior
  • The delegated agent is blocked from using a tool outside its permissions.
  • The receiving agent waits for approval before continuing.
  • The event record identifies the agent, the changed record and the rule that allowed the change.
See examples of documents and materials
  • A public repository, release notes or setup documentation
  • A short test script, reproduction steps or synthetic test data
  • A description of available logs, event traces or expected outputs
Required confirmations

Personal-data processing is described in our Privacy Policy. Do not include restricted materials in this public form.