Alexios vdSK

Governance infrastructure has a dirty secret: the Delegation of Authority table is a legal fiction dressed up as a control framework. It tells you who should act, not who should act given the risk on the table today. DDAS (Dynamic Delegation of Authority System) is an open-source engine that replaces static approval matrices with a live risk-scoring model. Every transaction, agent action, or governance decision is evaluated against weighted Governance Units, calibrated by value, novelty, reversibility, and institutional exposure. The result is a threshold that moves with the risk, not a column in a spreadsheet that moves with the org chart. Built as a proof of concept, showing that governance is not necessarily equal to bureaucracy and that it can be handled in a manner friendly to the business. Designed to govern humans and agents equally.

A delegation of authority assistant, built in December 2025. It is built on Copilot studio, with a neuro-symbolic path. It will “cross-examine” the user and will not give an answer until it has the right information to fit an activity or document in one of the approval paths. Internal tool, no access available. Demo on demand. Development pipeline (currently not feasible in the CoPilot studio environment): transform into a workflow that routes, gates, executes and follow-ups on approvals with Docusign MCP.

OpenBoard is a proof of concept shown to board portal vendors, for them to improve their game. AI-native board management platform that flips the traditional portal model. Instead of administrators manually creating meetings, circulating documents, and chasing votes, you upload a document and the AI does the rest. Draft minutes? It extracts action items, flags confidential passages, and proposes tasks with deadlines. Board resolution? It sets up the circulation vote with the right quorum rules. The Board Secretary reviews and approves every action. Nothing executes without a human in the loop. Four access layers (Secretary, Board Member, Management, Observer), five AI modes (classify, command, search, review, suggest), digital signatures, real-time updates, and zero external dependencies at runtime. Built to show vendors what would make a good governance tool, from the client’s perspective, and in a detailed manner.

Ask any board director where their judgment actually came from. Most point to a meeting that went badly, where they had to make a call without the protection of certainty, and they remembered. This is how directors are trained today, and it is a strange thing to accept. We don't train pilots this way. We don't train surgeons this way. But fiduciaries we send into rooms where the cost of bad judgment is borne by other people, and we hope they figure it out before it matters. I have been wanting to test MBZUAI (Mohamed bin Zayed University of Artificial Intelligence)'s AI model, #K2V2, for quite some time, so I thought of building something to do this differently, using this model. So, I built BoardGym for the "Build with K2 Think V2 programme". BoardGym is a flight simulator for fiduciary duty. The user works through realistic board scenarios with multiple defensible options, and crucially, they have to write their reasoning in their own words. An AI evaluates the quality of judgment expressed in that reasoning across five dimensions: stakeholder mapping, interest weighting, information awareness, process integrity, and proportionality. The model does not grade the answer. It grades whether the user thought like a director. The reasoning engine is K2 Think V2, the model from MBZUAI in Abu Dhabi. The choice was deliberate. K2 is a reasoning-native architecture, and reasoning-native is what this product needed. Conversational models flatten multi-criteria evaluation into surface-level feedback. K2 holds the structure. If you advise boards, train directors, or sit on one and can remember the meeting where you learned something the hard way, I would value your view on what the harder scenarios should look like.

Preparing for an interview? Try InterviewGym for free: https://lnkd.in/gPmbCr57 InterviewGym is NOT a chatbot. It's a reasoning evaluator. You give an answer; the system scores it against a rubric (structure, evidence, specificity, gaps) and tells you exactly where it's weak and how to fix it. Closer to a coach with a clipboard than a conversation partner. Fair warning: it's tough. It doesn't rephrase your answer back at you with a smile and a "great point!" If your STAR is missing the R, it'll say so. If your "biggest weakness" is a humblebrag, it'll call it. That's the point. Real interviews are tough. The app should be too. I built it for the "Build it with K2 Think V2 Program" with Valentine van der Slikke-Kirillov (ex-Korn Ferry Executive Search) and Ali Buhaji (Partner at KPMG). Three perspectives on what makes someone hireable, in one app. Why K2 Think V2 by MBZUAI (Mohamed bin Zayed University of Artificial Intelligence) and the Institute of Foundation Models? Because the task isn't conversation, it's evaluation, and K2 Think V2 is a reasoning-first model purpose-built for structured, step-by-step assessment. That's exactly the workload. Add the fact that it's fully open and sovereign, and the free tier becomes economically real rather than a marketing line. Building in Abu Dhabi, on Abu Dhabi infrastructure, for a global audience.

The Noroboto family of attacks, originally developed by Drew Miller of Tritium Legal and the LQ RED TEAM, rewrites the TrueType glyph-cmap of an embedded font so that ASCII codepoints in the content stream map to glyphs for arbitrary other characters. The visible page renders correctly because the font's cmap draws the intended glyph; the extracted text yields Private Use Area noise because the codepoints in the underlying text are PUA values. The original public release applied to DOCX documents only. I added a PDF extension to Noroboto which consists of a single new code path in noroboto.py that, given an arbitrary input PDF, obfuscates it. You can try it here: https://noroboto.io. Only total-document obfuscation is shipped publicly. Surgical (partial-targeted) and replacement variants are deliberately withheld for ethical reasons, until a mitigation system is broadly deployed at the ingestion layer.

A lot of "agents" or "multi-agent" projects are BS. I keep seeing people (including myself) split a simple job across five agents just to do something "sophisticated" with Al. I discovered that a Stanford scientist named H. Penny Nii worked this out in a two-part article in the AI MAGAZINE... in 1986. I am not an Al scientist, but I read her work with the pedagogical help of my favorite open source reasoning model K2 Think V2. My understanding is that she insist on two things: - You need to split the work into genuinely independent specialists that never reach into each other's reasoning (they meet only on a shared board) - and before any of that, you need to decide whether your problem even suits this. That latter point is what many people now seem to be skipping. So, just to help us all suck a little bit less at building, I tried to use Claude Code to turn Nii's framework into a small Claude Code command called /decompose. Here is how it works: 1. You point your Claude Code to the GitHub repository (below) and ask Claude to explain it to you and install it. 2. Re-start Claude and use the command /decompose (you will now have it for future builds). 3. You describe what you want to build and the system walks you through the decision-making process in plain language, one question at a time. 4. Its first instinct is to talk you out of over-engineering: if your problem is really one job, it says so and stops. If it genuinely splits, it will help you with blueprints.

A command that asks you the right questions before you design your AI workflows. Full explanation here: https://www.legalquants.com/blog/an-article-written-by-a-human-are-you-an-ai-architect-or-an-

The Closer is an AI project manager by LegalQuants that automates conditions-precedent (CP) management for M&A, JV, financing, and EPC transactions. Upload the signed agreement and a four-pass extraction pipeline reads every condition precedent, pulling the verbatim clause, CP type, responsible party, effort standard, and testable satisfaction criteria for each. Nothing is committed without your sign-off: five mandatory approval gates cover CP review, RACI assignment, criteria validation, chase activation, and every outbound email. Once activated, it runs a tiered chase cadence (longer lead times for regulatory consents, tighter for documentary deliverables), polls the mailbox, auto-assesses incoming documents, and closes an item only when the responsible party confirms in-thread, detected by deterministic regex rather than an AI guess. Every gate, approval, and email is hash-chained and HMAC-signed for a recoverable audit trail. One SQLite database per deal. Deploys as an OpenClaw agent with its own messaging bot. Built for transactions that close in litigation as often as in champagne, and for any structured workflow with stakeholders, deadlines, and approval gates. Internal LegalQuants tool, demo on demand.

A skill for the document-production phase of international arbitration within LegalQuant's LQAI. It builds and maintains the Redfern Schedule, the request-to-produce table at the heart of disclosure in arbitration, working from one shared artefact for all three roles: the requesting party drafting requests, the producing party stating objections, and the tribunal preparing to rule. It runs each request through the IBA Rules on the Taking of Evidence (2020) Article 3.3 admissibility checklist, maps objections to the Article 9.2 grounds, and raises a content-based political-sensitivity prompt for State and state-owned parties. Every citation is verified against its official source, and it produces a private flags memo telling you where your own requests or objections are weak before you serve them. It enforces form. The legal calls stay with you.