Posts

LLM Enterprise Concept All is revolving around security No organization internal system can “reach out” across the safe boundary LLM has to be organization owned and completely hosted inside its safe perimeters Safe Perimeter has only one gate Single point of entry and exit Organization owned LLM Same attributes as any other infrastructural part on the “inside” (RDBMS, AD, ESB) Same features and capabilities as the “front tier” LLMs Also trained on the Enterprise Private Data Aleph Alpha: The Right Model Aleph Alpha, founded in Heidelberg in 2019, built Luminous — a full LLMOwn architecture, training, and own inference. Same category as OpenAI and Anthropic: model creators. ...

DBJ Taxonomy defines the language for coordinates of the infospace. Content nobody can locate is content nobody finds. Taxonomy defines the locations on the map the business holds firmly in its hand. Taxonomy is the language used for programming infospace navigation instruments The .taxonomy file connects the abstract taxonomy to actual content. It declares a taxonomy location using the DBJ Core Taxonomy key-words. Any kind of content: repositories of documents, images, specifications. ...

IMPORTANT Durable execution is a crash insurance for a long-running software. It is not a business process, and adopting it does not mean your organization “has a workflow.” What it actually solves A program holds its state in memory, and memory does not survive the program. When a process dies halfway through a job, everything it knew is gone: which step it reached, which steps already completed, whether the payment went out once or twice. Regardless of the job. ...

IMPORTANT When someone says “we have a workflow for that,” they mean one of (at least) three things. Ask which. The answer tells you whether you are looking at a business decision, a description of one, or a piece of plumbing. One word, three jobs Start with something older than software. A workflow was a paper routing slip. A form moves desk to desk: approve, stamp, forward to accounting. Each desk does one step. The slip records where it is and what has already happened to it. If a clerk quits mid-task, someone finds the slip in the in-tray and picks it up, because the state is written on it. ...

C++ DBJ Error Concept — Revisited Back in 2019 I wrote C++ DBJ Error Concept, at last, proposing VALSTAT — Value & Status — as an answer to a gap ISO C++ has never closed: there is no standard C++ concept for “how did this call actually go.” Seven years on, the standard still doesn’t have one. The idea might be worth restating. (Certainly std::expected seems not to be widely adopted. My own assertion.) ...

Someone described DBJ Taxonomy this way: “a taxonomy that maps to Zachman-style layers (Strategic/Conceptual → Logical → Physical → Implementations).” Sounds credible. Sadly, it isn’t true. That image is the simplified four-layer version quoted in the claim. It’s not what Zachman actually defined — his framework is a 5×6 matrix (Scope/Business/System/Technology/Component crossed with What/How/Where/Who/When/Why), not a single top-to-bottom stack. The linear “Strategic → Logical → Physical → Implementations” progression is a loose paraphrase, which is also why the mapping claim doesn’t hold up. ...

See also: Agentic Agile Manifesto, Scrum Is a Category Error for Agentic Teams. The other day I stumbled upon good article (again) on the BCG (Boston Consulting Group) site: BPT vs Agile https://www.bcg.com/publications/2024/why-companies-get-agile-right-wrong My summary so far BCG’s finding: only about half of companies claiming high agile maturity actually achieved transformation goals with real business impact — while in the same time 47% adopt the practices but get none of the benefit. Naturally those accused of “operating under the illusion of agility”, will be (very) weary of adopting “yet another ideology” on top of or instead of agile. ...

An open-source event-driven control plane for platform engineering — and what it gets right about operational workflows. Platform teams write a lot of glue: bash scripts, cron jobs, Slack bots, wiki pages titled “runbook.” That stitches CI, alerts, and release trains together — until it doesn’t, and then the “glue abyss” opens, nobody knows which piece failed. SuperPlane is an open-source attempt to fix that at the source: a control plane that sits across your existing toolchain — GitHub, PagerDuty, Datadog, Slack, and 40+ others — and coordinates actions between them without replacing any of them. ...

IMPORTANT One Liner Summary: skill.md is not guaranteed to be used DBJ Observations and Comments Observation: runtime infrastructure is not deployment infrastructure Agreed, so what? In the era of the general lack of experienced engineers that has to be said. Plainly. Observation: Skill is key problem is, it is keeping the whole agent scaffolding non-deterministic It’s not Skill, it’s the mechanism that is supposed to use the Skill. Fuzzy natural-language matching against a description, decided by the model at invocation time rather than part of a fixed dispatch. same mechanism is what makes deferred-tool loading (ToolSearch), subagent selection. Also the ordinary tool choice (Grep vs Read vs Agent) is then non-deterministic too. Skills are just the most visible facet of the LLM non-determinism because they’re named and listed explicitly. But skills are used non deterministically. The classical-software analogue is late binding / reflection-based plugin dispatch Skill mechanism is trading a fixed call graph for runtime flexibility, and paying for it in determinism lacking. Repeating. It is not “Skill” that is the problem, it is that harness resolves most capability binding (skills, tools, subagents, memory recall) via probabilistic matching instead of a deterministic dispatch table And that’s a structural feature of the whole LLM architecture, not a flaw isolated to the Skills. There is no deterministic table dispatch. It is as simple as that IMPORTANT The Message ...