The Agent Is Not the Control Plane | Claude vs Codex vs Kimi on a Real AI Audit

The Agent Is Not the Control Plane | Claude vs Codex vs Kimi on a Real AI Audit

I gave three advanced AI systems the same task: audit the formulas, calculations and worked examples in my forthcoming professional book on FX options. Claude, Codex and Kimi. Seven hundred and sixty-four typeset pages, roughly four hundred and seventy equations, two hundred and thirty-five worked examples. This was not a writing competition. The question was whether the book's technical claims could survive scrutiny — and what that experiment reveals about AI governance, agentic AI and the future of coordinated work. In this episode: * Why a technical book is not a document but a network of dependent representations * Why "review this chapter" is the wrong instruction to give an AI model * Claude as conceptual reader, Codex as implementer, Kimi as auditor — and why one model won * The crashes that revealed the architecture: the session is not the system * The principle that became non-negotiable: the model explains, tools calculate * Governance as infrastructure — and the Post-Project World in miniature Based on the OrbaOS field note "The Agent Is Not the Control Plane". Read the full field note: https://orbaos.com [https://orbaos.com]Orchestra, persistent semantic memory for agents: https://orchestra.berta.one [https://orchestra.berta.one]The book, FX Options Volatility Trading: https://rondanini.com/fx-options [https://rondanini.com/fx-options] The Post-Project World is hosted by Luigi Pascal Rondanini, founder of Rondanini Publishing and OrbaOS, with thirty-eight years in FX markets. New episodes explore how AI coordination, governance and persistent infrastructure are replacing the traditional project as the unit of work. Keywords: AI governance, agentic AI, AI agents, Claude, Codex, Kimi, AI orchestration, LLM evaluation, AI audit, deterministic tools, multi-agent systems, control plane, OrbaOS, post-project world, FX options, financial publishing, AI in finance, machine-mediated work.

OrbaLM: Coordination Capital for the Treasury Desk

What happens when a fluent AI provides a perfect-looking table of FX Greeks that is mathematically impossible? In this technical deep dive, we explore OrbaLM, a governed specialist runtime designed for the high-stakes environment of treasury and FX desks. Moving beyond the "commodity" nature of base foundation models, OrbaLM acts as a control plane that wraps your firm’s approved language model within its own security boundary. We break down the core philosophy—the model explains; governed tools calculate—and explain how this architecture eliminates "invented capital" and hallucinations by using deterministic tools for arithmetic while leaving the AI to provide context and explanation. Key topics covered in this 30-minute briefing: * The Coordination Capital Bridge: How OrbaLM operationalises alignment between authority, state, and constraints on a dealing floor. * Transparency with Boundaries: A look at the read-only public delivery board and why showing status—not internals—is the key to professional trust. * The "We Don't Only Preach" Proof: How the team uses its own OrbaOS tools, including ScenarioForge and Berta Seal, to maintain the product. * Excel as the Practical Heart: Integration with multi-tab treasury workbooks for FX forwards, bonds, and interest rate swaps. Founding Beta Terms: Details on the September shipping window and the unique offer where the runtime remains free for the user after the three-month pilot. * Inquiry: Orbalm@orbaos.com * Delivery Board: orbalm.orbaos.com [https://www.google.com/url?sa=E&q=https%3A%2F%2Forbalm.orbaos.com] * Fleet Portal: berta.one [https://www.google.com/url?sa=E&q=https%3A%2F%2Fberta.one] Access & Inquiry: Founding pilot shipping starts in September 2026. Access is strictly by inquiry using a company email address.

OrbaLang — Write Rules Once, Emit Them Everywhere

Every regulated institution has rules. Most cannot prove that their systems enforce those rules consistently. A payment threshold approved by a committee in 2019 is implemented in Python by one team, Java by another, and a COBOL batch job on the mainframe that nobody wants to touch. Each drifted from the original intent. When audit asks which line of code enforced that limit on the 15th of March, the room goes quiet. That gap — between what the institution says its rules are and what its systems actually enforce — is a governance failure. OrbaLang addresses it. OrbaLang is a transpilation-first business-logic language. You write business rules once in a small, readable syntax. You validate them. Then you emit them — to Python, TypeScript, Java, Go, SQL, COBOL, and other targets. Not a proprietary runtime. Not a central rules server. Source code your teams deploy wherever they already operate. The same rule. The same logic. The same audit trail. Across every target. Each OrbaLang file is one of two things: a function (typed inputs, deterministic steps, typed return — what is the value?) or a rule (typed inputs, Boolean condition, required governance action, audit metadata — what must we do, and under what control?). That distinction separates a business-logic language from a governance-logic language. In this episode, Luigi Pascal Rondanini introduces OrbaLang through a treasury desk use case: daily limits, high-risk escalation paths, invoice approval thresholds, blocked jurisdictions. Each encoded as a rule or function. Emitted to Python for the microservices team, TypeScript for the web tier, COBOL for the settlement engine. The same when clause. The same then action. The same audit block. Not three interpretations that drift over quarters. The episode covers why OrbaLang differs structurally from Drools, OPA, DMN, and Python-alone approaches. Drools gives you a runtime and a vendor dependency. OPA gives you policy as data. OrbaLang gives you policy as auditable language with multi-target code generation, including legacy validators that some programmes still require. Python alone is infinitely flexible, which is the problem. Flexibility without governance vocabulary becomes ten styles in ten repositories. OrbaLang connects directly to the Coordination Capital Doctrine (Chapter 15) and the OrbaOS governance formalisation architecture. The Coordination Capital Ratio tells you how much coordination exists and what it costs. The Structural Floor tells you how much is mandatory. OrbaLang tells you what the rules actually say — in every system, in one language, with one audit trail. Measurement without enforceable logic is observation. Logic without measurement is assertion. Together, they create a governance system that can explain itself. OrbaLang is MIT licensed and lives in the open. Playground available at lang.orbaos.com. Part of the OrbaOS ecosystem alongside OrbaOS Instruments, ScenarioForge, and the Coordination Capital Doctrine. Hosted by Luigi Pascal Rondanini, author of The Coordination Capital Doctrine and founder of OrbaOS. Keywords: OrbaLang, transpilation, business logic, governance logic, rules engine, policy as code, write once emit everywhere, treasury transformation, payment threshold, sanctions, compliance, audit trail, audit metadata, governance chain, regulated financial institution, COBOL, legacy systems, mainframe, Python, TypeScript, Java, Semantic IR, code generation, transpiler, decision tables, governance vocabulary, OrbaOS, coordination capital, coordination capital ratio, structural floor, governance formalisation, CFO governance, audit committee, risk management, Drools, OPA, DMN, policy language, version control, traceability, enterprise governance, institutional governance, Rondanini Publishing, Post-Project World Topics/Categories: Business, Technology, Management

OrbaOS Deep Dive: Proving FRTB & GloBE for the Post-Project World

In this extended 50-minute technical session, we explore why institutional regtech must move toward falsifiable lineage. For bank CROs and Group Tax Directors, the challenge is no longer just the calculation—it is the provenance. We break down the OrbaOS doctrine ("Admitted Before Calculated") and its application across two of the most complex regulatory regimes today: the Fundamental Review of the Trading Book (FRTB) and OECD Pillar Two (GloBE). Key Technical Topics: * The Merkle DAG Architecture: Why every capital or tax figure must be a node in a hash tree to prevent "silent drift". * FRTB Standardised & Internal Models: Handling sensitivities, default risk, and liquidity horizons with regulator-template exports. * GloBE Article 5 Walks: Navigating the jurisdictional top-up process from book profit to GIR XML export. * QDMTT / IIR / UTPR Routing: How the engine uses OECD-published facts to determine exactly which jurisdiction collects top-up tax. * Self-Hosted Governance: Deploying Postgres + FastAPI + React within your own network to ensure data residency and security. * For FRTB: frtb.orbaos.com [https://www.google.com/url?sa=E&q=https%3A%2F%2Ffrtb.orbaos.com] * For GloBE: globe.orbaos.com [https://www.google.com/url?sa=E&q=https%3A%2F%2Fglobe.orbaos.com] "Your regulator can replay our exports offline with a single Python file. No backend access required." Keywords: FRTB, OECD Pillar Two, GloBE Model Rules, RegTech, Market Risk, Basel 3.1, Merkle DAG, QDMTT, IIR, UTPR, GIR XML, Bank Capital, Tax Governance, PRA SS21/15, EBA COREP, APRA APS 116. Call to Action: Institutional risk and tax leaders can request access to a private demo tenant to walk the doctrine end-to-end on synthetic data. Podcast Episode Chapter Timestamps * 00:00 – 05:00 | Opening: Institutional Infrastructure in the Post-Project World: Framing the shift from temporary consulting "projects" to permanent, living infrastructure. Why OrbaOS is built for a world where "one-and-done" compliance is no longer enough. * 05:00 – 12:00 | The Reconciliation Crisis: Why FRTB and Pillar Two Programs Fail: A deep dive into the operational failure modes where Risk, Finance, and Audit cannot agree on input rows. Discussing why "most FRTB programmes don’t fail on the math — they fail on reconciliation". * 12:00 – 22:00 | The Doctrine: Admissibility Before Calculation: Technical breakdown of the Merkle DAG and the hash-echo commit flow. Explaining why "admitted before calculated" is the core governing principle of the engine. * 22:00 – 32:00 | FRTB Specialist Segment: SA, IMA, and Regulator-Ready Exports: Walkthrough of the SBM, DRC, and RRAO modules, alongside the Internal Models (IMA) pipeline including Expected Shortfall and RFET. How exports map to PRA, EBA, and APRA templates. * 32:00 – 42:00 | Pillar Two Specialism: Article Walks and Top-Up Routing: Detailed navigation of the Article 3.2 and 4.1 walks, Article 9 safe harbors, and the deterministic routing of QDMTT, IIR, and UTPR based on OECD Central Record facts. * 42:00 – 47:00 | Operational Governance: The Warn-Gate-Block Path: How to deploy the system without breaking legacy paths. Moving from observational logging to strict four-eyes control and eventually "fail-closed" production. * 47:00 – 50:00 | The Verification Replay & Demo Tenant Experience: How to use the standalone verify_lineage.py for offline audit replay and an invitation to the 45-minute private demo tenant for experiential proof.

Does ChatGPT Know Your Company Exists? Introducing Arbor | Bonus Episode.

Search engines rank pages. AI assistants assemble answers. In this bonus episode of The Post Project World, Luigi Pascal Rondanini, using some AI actors, introduces Arbor, a platform designed to measure how companies appear within the answer layer of generative AI. As buyers increasingly ask ChatGPT, Claude, Gemini and other AI systems for recommendations, explanations and comparisons, traditional search rankings no longer tell the whole story. An organisation may perform well on Google while remaining absent, misrepresented or incorrectly described in AI-generated answers. This episode explores how Arbor helps organisations examine: • visibility across major AI models • citation frequency and source attribution • factual accuracy and hallucination risks • competitive positioning within generated answers • gaps in authority, evidence and discoverability • changes in AI visibility over time Rather than relying on opaque visibility scores or marketing claims, Arbor uses an auditable, evidence-led methodology. Each result can be traced back to the prompt, model, response, citation and date on which it was observed. The discussion also considers why answer-engine visibility matters to communications teams, SEO specialists, PR professionals, digital strategists and enterprise leaders—and how organisations can move from being merely present online to becoming trusted sources that AI systems recognise and cite. Arbor is available through cloud-based and self-hosted deployment options for organisations requiring greater control over their data and monitoring processes. Learn more at arbor.berta.one. A bonus episode of The Post Project World, produced by Luigi Pascal Rondanini for Rondanini Publishing Ltd.