Chronological essays and field notes on sovereign AI, local-first architecture, agent systems, and knowledge infrastructure.
Hermes Atlas v0.1: A Local First, Incrementally Compiled, Falsifiable Knowledge System August 9, 2026 · Daniel Kliewer Most "agentic knowledge" tooling is query time retrieval with a graph bolted on. Microsoft's GraphRAG is the reference case, and it has two documented weaknesses that a compile time design fixes: 1. No native incremental update. New information forces a full graph rebuild. 2.…
Hermes Atlas v0.2: Claim Level Coverage, a Determinism Gate, and a Knowledge Mutation Canary August 9, 2026 · Daniel Kliewer v0.1 built a local first, incrementally compiled, falsifiable knowledge system. It worked. But under honest evaluation it had two real weaknesses, and a third that the canary couldn't even measure yet. v0.2 is a benchmark and fix release . The spec was explicit: no new…
Hermes Atlas v0.3: An Evidence Ledger, a Provable Claim Lifecycle, and an Explain Command August 9, 2026 · Daniel Kliewer v0.1 built a local first, incrementally compiled, falsifiable knowledge system. v0.2 closed the coverage and determinism honesty gaps with measurements. v0.3 answers the question both releases left open: when the system says it believes something, can you see why — and can you…
Hermes Atlas v0.4: Proving the Evidence Ledger & Claim Lifecycle with a Gold Labeled Benchmark August 9, 2026 · Daniel Kliewer v0.1 built a local first, incrementally compiled, falsifiable knowledge system. v0.2 closed the coverage and determinism honesty gaps with measurements. v0.3 added an evidence ledger, a provable claim lifecycle (including DELETE driven invalidation that retains history),…
Two reads again. To you: a build log you can reproduce — exact commands, the API shapes we verified, and the mistakes we paid for. To the agent: a continuity record of what Hermes Video Agent is, how it is wired, and the environment rules that keep it from breaking, so a future session can pick up exactly where this one left off. Start at the Current environment reference if you are the agent.…
Why this post exists This is both a retrospective and a replication guide . Over a single working session, an AI agent (Hermes Agent, running locally on my Mac) took a vague idea — "estimate rehab cost from public property data, locally" — and produced a working, committed, pushed to GitHub application. No human wrote the code by hand; the agent did, and I (the human) steered, reviewed the…
This post is two things at once. To you, it's a guide: how to take a general agent and teach it a craft — video editing here, but the method generalizes to anything. To the agent reading it, it's a continuity record: the state of the world as of this writing, so a future session can pick up exactly where this one left off. If you're the agent, skip to the Current environment reference. The model…
Two failures I refused to fix Popper ORKG ended on a deliberately uncomfortable note. The compiler beat its three controls on a corpus I did not build, the ablation proved the architecture was load bearing, and the grounding audit found zero hallucinated citations across 265 evidence items. But I published two failures alongside the wins, un fixed by design: 1. Temporal rediscovery was near zero.…
The objection I had to answer The last experiment, Popper, made a strong claim and backed it with numbers: a knowledge graph compiler produces hypotheses that are not merely plausible but falsifiable — carrying variables, a mechanism, a quantitative prediction, and source level provenance — and it beats three controls (unstructured LLM generation, keyword retrieval, random traversal) on a battery…
The turn this experiment makes The last post ended with a promise and a warning. The promise: the Scientific Question & State Compiler survives scale — 6.3 million ORKG triples, distilled, compiled locally, same shaped output. The warning, which was the whole point: surviving scale is not the same as being correct. A tool that generates and scales has proven it is robust. It has not proven it is…
Building the Scientific Question Compiler: A Map of What We Do Not Know By Daniel Kliewer · 2026 07 23 Yesterday I wrote the roadmap: the four "research directions" for the Scientific State Compiler are not four projects, they are five passes in one architecture — Evidence → Questions → Mechanisms → State → Discovery. The roadmap's central question was which pass is the missing bottleneck? My…
Yesterday's post shipped Pass 4 , the Scientific State Compiler: a versioned model of what a field understands, separating accepted from uncertain from contested, preserving disagreement rather than flattening it. Today we closed the loop. The compiler now takes its own open questions, turns the most leveraged one into a new evidence source , re compiles the field, and records a real version diff…
What today was Today we took the Scientific Question & State Compiler — born on three hand written documents — and pointed it at real, large scale data : the full ORKG (Open Research Knowledge Graph) RDF export, 6,344,307 triples , 897 MB of N Triples. The question was honest and narrow: does the compiler generalize, or was the toy corpus hiding its assumptions? The answer, after a full day of…
Yesterday we built Pass 2, the Scientific Question Compiler , which turns a corpus into a map of what humanity does not yet know — the ranked open questions, the contested claims, and the silence where the literature simply does not speak. Today we ran Pass 4: the Scientific State Compiler . It takes that uncertainty map and promotes it into something stronger — a versioned state of a field : a…
Compiling the State of Human Understanding: A Roadmap for the Scientific State Compiler By Daniel Kliewer · 2026 07 23 Two days ago I argued that scientific knowledge does not need a better research assistant — it needs a compiler . Yesterday I built the general Scientific Discovery Compiler and watched it run end to end: five discovery passes turning a mechanistic biology corpus into causal…
The compiler was born on three hand written documents. That was enough to prove the shape of the idea — Pass 2 maps what we don't know, Pass 4 versions what we do, and the loop closes them. But a tool that only works on toy input isn't a tool. It's a demo. So we pointed it at real, large scale data. Not a curated sample — the full ORKG (Open Research Knowledge Graph) RDF export: 6,344,307 triples…
If the Loop Is the Product, Observability Becomes the Operating System By Daniel Kliewer · 2026 07 22 Yesterday I published an argument: scientific knowledge does not need a better research assistant. It needs a compiler — a system that transforms literature into versioned representations of understanding rather than retrieving and summarizing it. That essay was a vision. It described the category…
Compiling Medical Evidence: A Clinical Research Compiler That Turns Guidelines and Trials Into a Queryable Knowledge Artifact By Daniel Kliewer · 2026 07 21 This is a working build of the Clinical Research Compiler : the Research Compiler Agent SDK evolved into a medicine specific tool. It compiles real guidelines, trials, and reviews into one provenance tracked Clinical Knowledge Artifact , and…
A Research Compiler for Scientific Understanding, Not a Better Research Assistant By Daniel Kliewer · 2026 07 21 This is not primarily a description of a software project. It is a research vision — an argument for why scientific knowledge requires a new computational architecture, and what that architecture should look like. The immediate occasion is a build we shipped: the Clinical Research…
Compile Time Intelligence: Building a Research Compiler Agent that Compiles Knowledge Instead of Re Reading It By Daniel Kliewer · 2026 07 20 A research agent shouldn't re read the web on every turn. It should compile what it learns into a persistent, queryable knowledge artifact — and then reason over that . This post walks through the full design, code layout, and a live ru Github Link for…
Compile Time AI in Practice: How We Built a Kubernetes Knowledge Compiler Published 2026 07 18 · Daniel Kliewer The smartest thing you can do with a model is to stop asking it questions at runtime. That sentence is the whole thesis behind — a project that turns the entire Kubernetes documentation into a static, queryable, readable knowledge graph and ships it to k8s docs compiler.vercel.app with…
The Sovereign Knowledge Compiler Explorer: A Recipe for Compiling Knowledge Into a Static, Living Artifact Most "knowledge apps" are interpreters wearing a UI. You ask a question, they embed it, hit a vector store, pull top k chunks, and ask a model to re reason the answer — on every single query . The reasoning cost is paid again and again, and nothing compounds. The Sovereign Knowledge Compiler…
I Compiled My Blog Into a Decision Graph In the architecture post I argued that agent memory should be compiled , not retrieved: do the expensive reasoning once, emit static, inspectable artifacts, and let the runtime do cheap lookups. This post is the proof. I pointed the Sovereign Knowledge Compiler at every post on this blog and watched it turn writing into structure. No embeddings. No vector…
The Sovereign Knowledge Compiler: Compile Time Memory for Local First AI Agents Github Link to Project Abstract The original Sovereign Memory Bank (SMB) proposal argued that agent memory should be local and private rather than cloud hosted. That argument still holds, but it undersells the more interesting claim buried inside it: memory doesn't have to be retrieved at all — it can be compiled .…
<iframe width="560" height="315" src="https://www.youtube.com/embed/RYEU1Frf9OI" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard write; encrypted media; gyroscope; picture in picture; web share" referrerpolicy="strict origin when cross origin" allowfullscreen </iframe Full NoteBookLM Knowledge Compiler SDK Github Live Demo The Recursive Research Compiler:…
Synthesizing Memory with Agent: A Local First Architecture for Persistent AI State Abstract The prevailing paradigm in AI agent development treats memory as an external retrieval service, a failure mode that fragments intelligence across the model, the context window, and the vector store. This fragmentation forces engineers to choose between the fidelity of local inference and the persistence…
Compile Time AI: Why the Industry Is Quietly Building an LLVM for Knowledge Every few years, systems programming rediscovers the same lesson: expensive work done once, ahead of time, beats expensive work done repeatedly, on demand. That lesson is why we have compilers instead of interpreting source code line by line on every execution. It's why we have query planners instead of re deriving an…
Repository: github.com/kliewerdaniel/knowledge compiler Overview video: NotebookLM This is not a chatbot. It is a compiler. Why does an AI system need to rediscover the same knowledge every time it answers a question? This is the fundamental inefficiency that most knowledge systems silently accept. Every query against a RAG pipeline pays the full cost of retrieval, context assembly, and generation…
Every analytics dashboard I have ever used answers the same narrow question well: what happened . Pageviews, sessions, bounce rate, referral source. What none of them answer is why it matters — which pieces of content are actually building toward something, which pathways are quietly leaking high value visitors, and what I should write next to close the gap between what people are looking for and…
Intelligence is not the model. Intelligence is the accumulated decisions that shaped the model. The Sovereign Intelligence Stack is a production ready architecture for building sovereign AI systems. But how fast does it actually run? How much headroom does it have for real workloads? And how does it compare to alternative approaches? I benchmarked every critical component to answer these…
The Sovereign Loop: Why Model Local AI Is the Missing Operating System Layer July 6, 2026 The most capable coding agents right now aren't the ones with the single best model. They're the ones where the model and the harness were built for each other — Claude Code paired to Claude, Codex paired to GPT 5, OpenCode paired to whatever open model it's been tuned against that week. Arize AI's Aparna…
Getting Started with Sovereign AI: Your First Recipe Start small. Capture one recipe. Then watch the loop compound. By Daniel Kliewer Published: July 5, 2026 Reading Time: 15 minutes Prerequisites: None (beginner to advanced) This post is a beginner on ramp — it defines terms and walks you through your first recipe capture. For the full sovereign AI architecture (5 layer stack, compounding…
Intelligence Is Not the Model The model is not the product. The model is the ingredient. Every AI system that matters — every one that actually delivers value — runs on a loop. Not a single prompt, not a single inference call, but a loop that captures decisions, evaluates outcomes, and compounds intelligence over time. The model is a snapshot of accumulated decisions. The loop is the engine that…
The Model Is Not the Product: Residual State, Compiled Agents, and Optimization Loops July 3, 2026 The model is no longer the product. The loop is. That idea keeps getting reinforced every time I look at new research from Apple, LMSYS, and the recent work on autoresearch and constrained optimization. They're not converging on a better chatbot. They're converging on something closer to a…
The Loop Is the Product: Inside the Sovereign Intelligence Observatory July 3, 2026 Every agent framework on the market answers the same question: how do you get a model to do a task. Almost none of them answer the question that actually determines whether your system gets better over time: what happened, in what order, under what confidence, judged by whom, and is that judgment still valid six…
Autoresearch Loops and Differentiated Intelligence Two Converging Blueprints for Self Improving AI Systems Date: July 2, 2026 Introduction: The Shift from Models to Systems That Improve Themselves Two major threads in AI research converged almost simultaneously. On one side, Introspection's "autoresearch" framework reframes AI systems not as static models, but as self improving loops. On the…
Context Engineering: The Real Full Stack Development Paradigm in 2026 An exploration of the blind spots in current AI development coverage and the emergence of context engineering, agent harnesses, and the coding agent ecosystem as the true full stack development paradigm of 2026. Introduction: The Coverage Gap If you follow the AI development space in 2026, you've seen the headlines. Coding…
AMIS in Action: Live Vercel Analytics → Autonomous Marketing Knowledge Graph Date: June 30, 2026 Today marked another iteration in the ongoing validation of AMIS — the Agentic Marketing Intelligence System — a fully local first, Markdown corpus driven reasoning engine that transforms static blog content into dynamic, autonomous marketing intelligence. As the architect of both the system and the…
Github Sovereign Memory Bank: A Deep Dive Into Autonomous Cognitive Memory for Agent Systems By Daniel Kliewer · June 14, 2026 Every knowledge system I've built — and most I've encountered in the wild — treats memory the same way a warehouse treats inventory: it arrives, it gets shelved, and it waits passively for retrieval. That model is fundamentally broken for the class of problems I care…
Sovereign Memory Bank: Autonomous Cognitive Memory for Agent Systems Every knowledge system I've built — and most I've encountered in the wild — treats memory the same way a warehouse treats inventory: it arrives, it gets shelved, and it waits passively for retrieval. That model is fundamentally broken for the class of problems I care about: agent reasoning, knowledge synthesis, and emergent…
SovereignSpec and the Ganymedean Alignment Protocol Table of Contents 1. Prelude: From Myth to Methodology 2. Specification Supremacy 3. Specification Semantics and Grammar 4. Specification as Graph Nodes 5. Constitutional Governance Model 6. Recursive Specification Evaluation 7. Contradiction Detection and Resolution 8. Narrative Drift and Spec Evolution 9. Knowledge Graph Integration (RAG) 10.…
Project initiated on Github Project Proposal: SovereignSpec — Local First Spec Driven Development The spec is alive. The code obeys. Nothing leaves your machine. 1. Executive Summary GitHub's Spec Kit (released September 2025, now at v0.5.0 as of mid 2026) has catalyzed a shift in how software gets built: Spec Driven Development (SDD) — where specifications are the single source of truth, and code…
objective05 exec: Giving Your Local Intelligence System Hands How to bridge a perpetual knowledge graph to real world tool execution — a Rust tutorial June 8, 2026 · Daniel Kliewer GitHub: kliewerdaniel/objective05 Table of Contents Introduction The Landscape: What Everyone Else Is Building The Gap Prerequisites and Environment Setup Architecture Overview Step 1: Project Structure Step 2:…
OpenDesign + OpenCode: Building a Local First Design Operating System Inside Your Terminal June 8, 2026 · Daniel Kliewer There is a strange contradiction at the center of modern software development. We have coding agents capable of writing React applications, deploying infrastructure, refactoring monoliths, generating tests, orchestrating CI/CD pipelines, and reasoning across entire repositories.…
The Model Is Not the Product: On Building Persistent Intelligence Infrastructure June 3, 2026 GitHub There is a framing problem at the center of most AI discourse right now, and it is costing builders real clarity about what they are actually constructing. The framing is this: the model is the product. Improve the model, improve the product. Benchmark higher, ship better. This framing is not wrong…
<iframe src="https://www.youtube.com/embed/ qL7OtkNQ80" title="objective03 demo" frameborder="0" allow="accelerometer; autoplay; clipboard write; encrypted media; gyroscope; picture in picture; web share" referrerpolicy="strict origin when cross origin" allowfullscreen </iframe <br Free Apple Silicon Download objective03: A Locally Run Autonomous News Ingestion and Contradiction Tracking System…
"High quality data is not a precondition for intelligence — it is an expression of it." For most of the history of machine learning, data has been treated as an upstream problem. You gather it, clean it, label it, and then hand it off to a training pipeline. The model is downstream. The data is fixed. This division of labor has always been a bottleneck — not just logistically, but conceptually.…
Qwen Scope: Turning Sparse Features into Development Tools for Large Language Models Introduction There’s a quiet shift happening in AI that most people are missing. For years, interpretability has been framed as a diagnostic tool—something you use after the fact to explain why a model behaved the way it did. It was closer to autopsy than engineering. You could observe, maybe categorize, but…
Recursive Language Models: A Paradigm Shift for Near Infinite Context Introduction Modern large language models (LLMs) excel at many tasks but hit a hard wall with context length —the amount of text they can "remember" and reason over in a single pass. Even frontier models struggle with documents longer than their fixed window (often 128K–1M+ tokens), leading to "context rot": degraded performance…
Synthetic Intelligence: Why "Emergence" is Just Math and Why Your Data Should Stay Local Executive Summary The AI industry sells you a fairy tale: that intelligence emerges magically from cloud APIs, that consciousness is just around the corner, that you need to rent your thinking from trillion dollar conglomerates. Bullshit. Strip away the marketing gloss and what remains is linear algebra and…