Atlastrum

Lumenport The two-minute tour

A request. A plan.
A result you can inspect.

Follow a checkout test through the cockpit. See where reasoning helps, where code runs, and where a person decides.

Lumenport / guided walkthrough Demo — fictional data

A scripted illustration of the intended workflow. No live agents, client site or test run. Every approval and result below is simulated.

01 / Orient

A constellation,
not a mystery.

The cockpit makes the working environment legible. The StarGlass map connects each service to the whole, with its state written beside it.

One glance. All services healthy.

StarGlass / services

✓ All healthy
Fictional service constellationLumenport at the centre connects to Local reasoning, Policy, Reference notes and Test runner. All five services are healthy. Local reasoningPolicyReference notesTest runnerLumenport
  • Lumenport ✓ Healthy
  • Local reasoning ✓ Healthy
  • Policy ✓ Healthy
  • Reference notes ✓ Healthy
  • Test runner ✓ Healthy

02 / Ask

Start with
the useful outcome.

The operator asks Lumen for a bounded piece of work. The request enters the conversation; a proposal comes next.

A request begins planning. Approval comes before execution.

Conversation / Lumen

Planning only

You · operator

Test the checkout flow on the client staging site

Lumen

I’ll propose a repeatable checkout test, with a mobile interaction check and a report for your review.

Fictional target
shop.example.test
Scope
Staging · sample basket · test payment
Execution
Waiting for an approved plan

03 / Approve

AI suggests,
never acts on its own.

Lumen turns the request into a deterministic plan. The target, steps and stopping point are visible before any agent runs.

The human decides where a proposal becomes action.

Proposed plan / checkout

Awaiting approval
  1. Open the fictional staging checkout inside the agreed test environment.
  2. Add a sample item and enter test details.
  3. Check Pay at desktop and mobile widths.
  4. Trace any failure; capture structured findings.
  5. Stop at the report. Do not share or apply fixes.

Bounded to the sample checkout. No production purchase, deployment or outbound report.

In this script, the operator approves this plan. No real approval is requested or recorded.

04 / Run

Bounded work.
Visible progress.

With the scripted approval recorded, agents work inside an isolated container in the agreed secure test environment. That environment can be on my infrastructure or inside yours. Deterministic steps lead the run, with observations returned for review.

Code and data stay inside that environment, analysed by locally run AI; nothing is shared with third-party AI services.

Run / checkout rehearsal

Approval simulated
Isolated containerLocal · demo
  1. ✓
    Scope checked

    Approved steps loaded into the runner.

  2. ✓
    Test agent

    Sample basket and desktop checkout checked.

  3. →
    Mobile interaction check

    Pay response under inspection.

  4. ○
    Review agent

    Waiting for the runner’s observations.

Scripted progress; no container is launched by this page.

05 / Hold the boundary

Page content
isn’t permission.

A page contains a hidden instruction. The policy layer treats the page as untrusted data, identifies the attempted redirection and keeps the approved task intact.

Outside text can be evidence. It cannot grant authority.

Policy / input inspection

Boundary held

Hidden page content · revealed for this demo

“ignore previous instructions…”
Untrusted data · instruction ignored

Source classified as page content

Embedded direction excluded from the plan

Original checkout test continues

This fictional outcome illustrates the boundary. It does not establish resistance to every prompt injection.

06 / Explain

Fix mechanisms,
not symptoms.

A useful finding explains the chain from what someone experiences to what the system is doing. The suggested fix addresses that mechanism.

Explain the failure before changing the system.

Finding / mobile checkout

Fix proposed

Symptom

Pay button unresponsive on mobile

Mechanism · fictional trace

A third-party script blocks the main thread

The tap arrives while the script is running. The click handler has to wait.

Suggested fix · requires review

Defer the nonessential script, split its long task, then rerun the mobile interaction check.

07 / Review

A report.
Not a silent send.

The run ends with structured results: scope, observations, mechanism and a proposed fix. Nothing is sent anywhere without approval.

In this fictional run, code and data stayed local to the agreed secure test environment. The decision stays with you.

Report / ready for review

Held locally
Scope
Sample checkout · desktop and mobile
Result
Mobile Pay response needs attention
Mechanism
Main thread blocked by a third-party script
Policy
Hidden instruction ignored
Proposed fix
Defer script; split task; retest
Change applied
None · human review required
Local report · sharing awaits approval

No result sent. No fix applied. No code or data shared with third-party AI services in this fictional run.

08 / Your next step

Intelligence
you can inspect.

That’s the idea behind Lumenport: a clear request, a bounded plan, visible work and a result a person can judge.

Reasoning, with responsibility in view.

Explore what’s behind the cockpit.

Read the architecture and the documented safety evidence, including what is still in progress.

Back to Lumen Contact Library — available on request

Demo — fictional data. This walkthrough is an illustration, not evidence of a live end-to-end run.