System architecture / engineering record

The work behind the work.

BlackBoxx.ai operates a private, modular AI environment built around specialized agents, explicit authority boundaries, human-controlled execution, and evidence that survives the demo.

Technical, not theatricalThis page separates what is operating now, what remains on the roadmap, and what has been independently acknowledged upstream. Sensitive implementation details are intentionally omitted.

Earlier DorkOS implementation snapshot · counts are not a live system inventory

16specialized agents
15bounded namespaces
34directional routes
3upstream improvements

Architecture

A governed operating environment for AI-assisted work.

My working architecture separates human authority, TRON’s Chief of Staff coordination, the DorkOS local execution environment and specialized agent assignments. It describes how I’m structuring the setup, which remains under development.

Documented architecture snapshot · October 4, 2026 (not a live system inventory)

  1. 01 / Human authority

    Karl Ohlemann

    I set priorities, define a useful result, grant authority and approve consequential decisions.

  2. 02 / Chief of Staff

    TRON 😏 / OpenAI Dots

    My top-line, always-on coordinating point helps translate priorities into defined work, coordinate tasks and bring back results or decisions that need my attention.

  3. 03 / Local execution platform

    DorkOS on my Mac

    The environment for organizing local agents and their assigned execution, within available connections, access and approval boundaries.

  4. 04 / Defined assignments

    Specialized agents

    Agents carry out assigned work and return results, evidence and exceptions so I can assess what happened.

Karl → TRON → DorkOS → specialized agentsSpecialized agents → DorkOS / TRON → Karl: results, evidence and exceptions return for review and decisions. TRON’s coordinating role does not remove human approvals or tools’ safeguards. Local execution depends on the Mac being available, connections working and the permissions granted.
DorkOS engineering graphic · supervisory layer and detailed local execution

Supervisory coordination · current working architecture

Human authority

Karl Ohlemann

Priorities, permissions and consequential approvals.

AI Chief of Staff

TRON 😏 / OpenAI Dots

Coordinates defined work; returns results and decisions for review.

Local execution

DorkOS on my Mac

DorkBot remains the local orchestrator for specialized agents.

The PNG below is the detailed local execution view from the earlier engineering snapshot. Its User → DorkOS arrows represent the authorized request and approval handoff into the local environment; the strip above adds TRON’s supervisory coordination around that handoff. Model routing and the 16-agent implementation remain inside the local view.

Work and results move in both directions. Human authority and tool safeguards remain in place; local execution depends on Mac availability, working connections and granted permissions.

Earlier DorkOS engineering snapshot showing model routing, a 16-agent layer, local infrastructure, tools, knowledge, security principles and roadmap.
Detailed local execution view · earlier engineering snapshot. The supervisory strip above adds TRON / Dots while retaining DorkBot as local orchestrator. Open the full-resolution snapshot ↗

Control model

Agent intelligence and agent authority are separate things.

A highly capable agent can still be prevented from taking actions or crossing boundaries outside its approved envelope. That distinction is the difference between an impressive demo and infrastructure you can begin to trust with real work.

01 / Orchestration

Coordination with defined roles

TRON helps coordinate priorities and work above DorkOS. Local agents receive defined assignments; granted permissions, tool safeguards and human review govern execution.

02 / Least privilege

Capability ceilings

Every non-orchestrator agent operates below an explicit ceiling that removes destructive capability classes it does not need.

03 / Directed access

No wildcard routes

Agent collaboration is allow-listed through directional rules, including one-way relationships where a domain may receive work but never initiate it.

04 / Verification

Controls are tested

Permitted routes and expected denials are exercised live. A control that has not been challenged is not yet a verified control.

DorkOS technical bulletins

Production use → diagnosis → upstream improvement.

These are public, verifiable examples of BlackBoxx.ai tracing reliability defects to root cause and collaborating with the DorkOS maintainers. Statuses reviewed against public GitHub records on October 29, 2026. A merged pull request verifies an upstream source change, not the first packaged desktop version containing it.

Fix merged

Desktop watchdog reload loop

A slow first paint could become an endless 11–13 second reload cycle that destroyed unsaved operator state. The server remained healthy throughout, which meant ordinary server troubleshooting could not explain the failure.

Diagnosis

A heartbeat from the outgoing page arrived after reload, reset the failure count, and prevented the recovery ladder from advancing.

Evidence

45/45 health checks returned HTTP 200, the server and renderer processes stayed stable, and 92 cycles were traced in application logs.

Upstream result

PR #1853 now accepts a heartbeat only from the page the watchdog is actually waiting on. The code fix was merged September 14, 2026; the first packaged desktop version containing it has not been independently verified.

“Your trace of the reset was exactly right and turned a day of digging into an afternoon.”Dorian Collier · Creator of DorkOS
Fix merged

False server-unreachable diagnostics

The desktop cockpit displayed “can’t reach its server” while the server was continuously reachable. That message misdirected diagnosis and trained the operator to distrust the system’s primary health indicator.

Diagnosis

The screen was keyed to a configuration request rather than server health, and a cumulative failure counter could remain set for the life of the page.

Evidence

The panel was observed while the health endpoint returned 45/45 successful responses in 1.0–5.4 milliseconds.

Upstream result

PR #1857 reserves “unreachable” for cases where nothing answered and reports actual server errors without guessing the cause. The code fix was merged September 14, 2026; the first packaged desktop version containing it has not been independently verified.

Shipped · v0.73.0

Remote Access packaging and state failures

Remote Access failed to start because a required native binary was absent from the packaged desktop application. The visible switch then snapped back without preserving the error, while higher log levels still failed to disclose the root cause.

Diagnosis

The per-platform ngrok package was a nested optional dependency the desktop packager never reached; separate state-sync and logging defects obscured the failure.

Scope

The upstream response expanded beyond the missing binary to persistent errors, retry behavior, restart survival, dropped-tunnel detection, and useful logging.

Upstream result

The combined fixes were released publicly in DorkOS v0.73.0.

“Your diagnosis was exactly right, and it saved us a lot of time.”Dorian Collier · Creator of DorkOS

State of the system

Operating now is not the same as planned next.

The implementation record below reflects the earlier DorkOS engineering snapshot. The current working role map above adds TRON; the broader multi-agent setup is still evolving.

Earlier engineering snapshot

DorkOS implementation record

  • 16 specialized agents recorded in the earlier DorkOS engineering map; TRON’s coordinating role is described above
  • Claude Code and Codex runtimes in production use; OpenCode and Ollama in beta
  • 34 explicit directional collaboration routes
  • Capability ceilings applied to every non-orchestrator agent
  • Dedicated local hardware with human-controlled execution
  • Permitted and denied paths tested against expected policy

Roadmap / select

Planned next

  • Local inference promoted from beta once runtime routing is finalized
  • Stronger isolation between sensitive domains — separate user accounts tested, rollout in progress
  • Expanded model routing across additional runtimes
  • Additional security hardening and recovery testing
  • Long-running agent workflows with clearer lifecycle controls
  • Multi-machine expansion when the control model is ready

Models, controls, and namespaces

Black-box capability. White-box control.

BlackBoxx.ai uses powerful models without confusing opacity with permission. The model may be complex; the operating system around it should make sources, authority, actions, and human accountability visible. The same infrastructure lens extends to programmable identity and namespaces.

01 / Black-box AI

Capability without full internal visibility

A black-box AI model can produce useful outputs even when its internal reasoning process is opaque or inaccessible to the person using it. BlackBoxx.ai works with that capability pragmatically: model output is treated as a proposal to evaluate, not unexplained authority to accept.

02 / White-box systems

Make the operating envelope inspectable

White-box models and systems expose enough of their relevant rules, inputs, transformations, and decision paths to be inspected. BlackBoxx.ai applies that principle to the surrounding operating system through source provenance, system instructions, permissions, tool calls, approval gates, logs, tests, exception routes, and accountable human review.

03 / Strategic namespaces

Owned infrastructure assets

BlackBoxx.ai founder Karl Ohlemann owns a strategically selected portfolio of 25 Web3 top-level domains through Freename, all minted on Polygon, including .blackbox, .blackboxx, .moonvault, .whitebox, and .whiteboxx. The portfolio also includes the conventional domain whiteboxx.io. These assets are being evaluated as programmable namespaces; the Web3 TLDs are not presented as production infrastructure today.

04 / Ecosystem access

Early product participation

Karl is a Freename VIP client and early product-feedback participant, with direct access to product, engineering, and executive teams through private previews, workshops, roadmap discussions, and feedback sessions. This is emerging-infrastructure participation and technology ecosystem access—not a claim of a formal partnership.

Open systems / real use

Built on DorkOS. Pressure-tested by operating it.

DorkOS is open-source software released under the MIT License. BlackBoxx.ai is an independent user and contributor, not DorkOS or its maintainer.