the world of dizzi

The ten

Nine applications, one contract — the core collects it

Every one of them runs on its own port, owns its own database and is designed to be separable. The nine around the core are built to answer the same mandatory endpoints.

The test counts below were measured on 26 and 27 July 2026, in the code extract that was public from 23 July. That extract is no longer published — since 6 August the public folder holds a curated set of documents, so these counts can no longer be reproduced from outside. They are kept, dated, rather than dropped.

The ten at a glance
Application What it does Port Tests
Dizzi-Core Orchestrator, registry, MCP hub, health watch. :8200 344 tests
Dizz Money Finance, double-entry ledger, German tax filing. :8210 462 tests
Dizz Communication Unified inbox, contacts, appointments. :8218 191 tests
Dizz Creating Media workbench over one GPU job queue. :8214 core not in this snapshot
Dizz Memory Markdown vault, full-text search, retrieval. :8212 118 tests
Dizz Management Agent administration, autonomy ladder. :8213 132 tests
Dizz News Sector digest, dedup, watchlists. :8216 77 tests
Dizz Healthy Metrics, training, wearables. :8217 54 tests
Dizz Admin Vault, projects, business, studies, deadlines. :8222 190 tests
Dizz Trading Autonomous bot fleet, hard human gate. :8137 471 tests

What each one does, the port it answers on, and the test functions counted in it. Where an application is thinner than its description, the row says so where the number would be. Add the nine up — Dizz Creating carries none, because its core was never in the extract — and you get 2,039, less than the network total of 2,667: the numbers page says where the rest sat →

Dizzi-Core

:8200 344 test functionsHow the test functions were counted, and when

Ten applications that each know how to talk to an assistant would mean ten connections to manage, ten places to authenticate, ten places for drift to creep in.

  • AI orchestrator over local models, driven by JSON schema grammar rather than hopeful prompting — temperature 0, fail-safe throughout.
  • Panel registry at /api/panels: every app announces its panels, so the core knows the whole interface without being told twice.
  • MCP hub — all app tool gateways pulled to one point, read-only and bearer-gated; writing and high-security tools stay closed to the outside.
  • Health watch over ten targets, plus the master instance of single sign-on and the layered memory.
The decision, and what it changed

The decision was to make the core headless and put everything shared in it: one orchestrator, one registry, one tool bus, one identity. The result is that adding an application costs a contract, not an integration.

apps/core ↗

The core never acts outward. It routes, relays and logs. Every effect on the outside world belongs to an application, and every one of those sits behind a human approval step.

Dizz Money

:8210 462 test functionsHow the test functions were counted, and when

Personal finance software either sees your bank or it sees nothing.

  • Import, categorisation, budgets, cash flow, foreign exchange.
  • A ledger care-core that keeps the books consistent rather than merely pretty.
  • German tax filing over the official transport path, with its own written contract for the submission format.
  • Bank sync over FinTS — read-only, specified as such in its own contract document.
The decision, and what it changed

The decision here was read-only by construction: the bank connection over FinTS can import and reconcile, and it cannot move a cent. Everything that does have an effect — a tax submission, a transfer — is a separate, gated act by a human. The result is the second-largest application test suite in the network, after the trading fleet — because money is where being wrong costs the most.

apps/money ↗·docs/64 — bank sync contract ↗

Why read-only is a feature. A finance app that cannot initiate a payment cannot be tricked into initiating one. The convenience I gave up is real; so is the class of attack I gave up with it.

Dizz Communication

:8218 191 test functionsHow the test functions were counted, and when

Messages arrive in six places and a person has to be the router.

  • Unified inbox across e-mail providers and messenger bridges.
  • Smart contacts, per-contact chat history, appointment detection out of message text.
  • Classified at the highest sensitivity level — which means AI routing is local only, by default and by contract.
The decision, and what it changed

The decision was one inbox across e-mail and messengers, with contacts as the organising unit rather than threads — so a conversation follows a person, not a protocol. The result is that scheduling, follow-ups and history stop being a function of which app someone happened to use.

apps/communication ↗

Sensitivity is a field, not an intention. Each app declares its class in its manifest, and the class decides where the AI may run. High and highest never leave the machine.

Dizz Creating

:8214 core not in this snapshot

Image, video and music generation each want the whole GPU.

  • One shared GPU job queue for image, video and audio work.
  • Asset management with AI labelling of generated content.
  • A workflow graph library covering image, video and audio pipelines.
The decision, and what it changed

The decision was a single job queue in front of one external generation service, so the network schedules the card instead of fighting over it, and every generated asset is labelled as generated at the moment it is written. The result is a media workbench that behaves like a service rather than a pile of scripts.

apps/creating ↗

Said plainly: the core implementation of this app was never published, and since 6 August 2026 no application’s code is. What the extract carries for this app is its own README. The repository README says the same thing. A gap you can see is worth more than a gap you find later.

Dizz Memory

:8212 118 test functionsHow the test functions were counted, and when

A knowledge store that lives in a proprietary database is a knowledge store you can lose.

  • Markdown vault plus SQLite full-text search and retrieval.
  • Folders and labels, area-level categorisation, import from existing material.
  • An MCP namespace of its own, so the assistant can search the vault without a second connection.
The decision, and what it changed

The decision was a markdown vault on disk as the truth, with a full-text index beside it for speed — so the files stay readable with any editor, and the index is always rebuildable. The result is a long-term memory across every area of life that outlives the software that made it.

apps/memory ↗

Files first. The index is a cache. If it burns down, nothing is lost — that is the whole point of choosing plain text.

Dizz Management

:8213 132 test functionsHow the test functions were counted, and when

Once assistants start doing work unattended, the question stops being “can it?” and becomes “how far may it go, and who says stop?”

  • Autonomy ladder T0–T3 with a hard human-in-the-loop step and an emergency stop, specified in its own contract.
  • A network inbox for what the agents want to tell the human.
  • Editorial planning per area, channels, drafts and schedules — publishing itself stays gated.
The decision, and what it changed

The decision was a written autonomy ladder from T0 to T3 with an emergency stop, so every agent has a declared ceiling instead of an implied one. The result is agent administration that can be audited: what ran, at which level, approved by whom.

apps/management ↗·docs/63 — agent direction contract ↗

An autonomy level is a promise you can check. It is written in a contract, enforced in code, and recorded in the audit log — three places that have to agree.

Dizz News

:8216 77 test functionsHow the test functions were counted, and when

A news feed optimised for engagement is optimised against you.

  • Sector-structured digest from a vetted source list.
  • Cross-source deduplication — one event, one entry.
  • Living watchlists that carry topics forward across editions.
The decision, and what it changed

The decision was a fixed set of vetted sources and a digest structured by sector rather than by recency, with cross-source deduplication so one event does not read as eight. The result is a briefing you finish, instead of a stream you escape.

apps/news ↗

No ranking model. The sources are chosen once, by a human, and written down. That is a weaker system than a recommender and a much more predictable one.

Dizz Healthy

:8217 54 test functionsHow the test functions were counted, and when

Health data is the most sensitive class in the network, and the temptation to send it to a large cloud model for a clever summary is exactly where these systems go wrong.

  • Measurements, supplements, injuries, training, appointments.
  • Deterministic trends, with local AI hints clearly separated from them.
  • Standard health data vocabularies, so the data can leave in a form other software understands.
  • Wearable and mobile connection points are prepared and documented, not claimed as built.
The decision, and what it changed

The decision was strict: this app’s AI routing is local only, and the trends it shows are deterministic calculations, not model output. The result is a health overview that gives hints and never a diagnosis.

apps/healthy ↗

Prepared is not built. The written laws require that dormant connection points are documented as dormant — so nobody, including the author six months later, mistakes a socket for a feature.

Dizz Admin

:8222 190 test functionsHow the test functions were counted, and when

Personal bureaucracy does not fit into projects or into folders; it fits into areas of life that persist for years.

  • Vault, projects, business paperwork, studies, deadlines.
  • Area-first: the canonical area-type model lives here and is shared by contract.
  • Feeds the area cockpit that Money, Management and Memory report into.
The decision, and what it changed

The decision was to make the area a first-class type with its own canon, shared across the network, so finance, knowledge and social activity can all report into the same structure. The result is the home of the area-type model — and the reason cross-app views line up at all.

apps/admin ↗·docs/49 — area-type model ↗

One naming note, in the open: this application was called leading until 20 June 2026. Some documents inside its folder still carry the old name and the old port. The canonical answer is the repository README and the system map: admin, port 8222.

Dizz Trading

:8137 471 test functionsHow the test functions were counted, and when documents not published

An autonomous trading fleet is the sharpest possible test of the network’s central rule, because here the machine is fast and the human is slow.

  • A bot fleet with a learning layer and a strategy research catalogue.
  • Kelly sizing with its own written specification.
  • A hard human-in-the-loop gate; paper trading is free; real money is a separate, highest-tier act.
  • Defence in depth: exchange keys never carry withdrawal rights.
The decision, and what it changed

The decision was that the fleet may propose without limit: a real transfer requires the highest security tier. Since 4 August 2026 a small part of the fleet trades with my own money, behind that gate. The result is a system whose learning layer can be aggressive precisely because its hands are tied.

This one is not in the extract at all. Thirty-odd dated documents sit next to this app — night runs, system checks, deep audits, research protocols and honest post-mortems from 8 June 2026 onwards. They were public until 6 August 2026 and were taken out with the rest of the fleet’s papers: sizing rules and decision mechanics of something that now moves real money do not belong on an open page. It is the one application here with no source link, and this is why.

Nine applications speak one written contract; the core is the hub that collects it. That contract is what makes an application separable — and separable is the precondition for selling one on its own.

Why that matters