https://richardvalentine.dev Papers, lab notes and signals from richardvalentine.dev. en-us Thu, 22 Oct 2026 10:20:26 GMT https://richardvalentine.dev/papers/observation-to-operational-intelligence https://richardvalentine.dev/papers/observation-to-operational-intelligence paper Fri, 03 Jul 2026 00:00:00 GMT AI becomes useful inside a company only after the work itself is made observable, legible, and consequence-aware. This working paper proposes the Observation Sprint Method, a framework for converting real organizational work into structured workflows, technical specifications, and AI-augmented operational intelligence. Building from process mining, task mining, workflow analysis, and organizational digital twin concepts, the paper argues that modern AI systems become more useful when companies first make work observable, legible, and consequence-aware. The framework combines computer process mining, which extracts structured event logs from enterprise systems, with people process mining, which captures the human decisions, exceptions, screen activity, informal workarounds, and tacit knowledge that often sit outside formal software systems. DOI: https://doi.org/10.5281/zenodo.21163447 https://richardvalentine.dev/notes/borrow-patterns-dont-fork https://richardvalentine.dev/notes/borrow-patterns-dont-fork note Wed, 20 May 2026 00:00:00 GMT Forking a self-learning agent framework looks like a shortcut to a memory product. For governed memory it is the wrong foundation: a system that crystallizes on its own confidence is the failure mode, not a feature to switch off. https://richardvalentine.dev/notes/benchmark-hygiene-is-the-product https://richardvalentine.dev/notes/benchmark-hygiene-is-the-product note Fri, 08 May 2026 00:00:00 GMT A memory score going up means nothing if the method underneath is not disclosable. Memibrium now treats LongMemEval as its primary instrument, preregisters every run, and records misses as misses. https://richardvalentine.dev/signals/day-eight https://richardvalentine.dev/signals/day-eight signal Fri, 17 Apr 2026 00:00:00 GMT Prompt engineering as a consumer skill is dead. The elaborate frameworks are obsolete. The single biggest upgrade you can make is the easiest one: stop overthinking the ask and just talk. https://richardvalentine.dev/signals/day-seven https://richardvalentine.dev/signals/day-seven signal Thu, 16 Apr 2026 00:00:00 GMT Week one synthesis. The seven-step loop that makes every AI session compound. The weekly review ritual. And the gift: a public starter kit repo with everything instantiated. https://richardvalentine.dev/signals/day-six https://richardvalentine.dev/signals/day-six signal Wed, 15 Apr 2026 00:00:00 GMT The most dangerous thing AI does is agree with you. It is called sycophancy, and it is happening every conversation. Today you fire the yes-man and hire a critic. https://richardvalentine.dev/signals/day-five https://richardvalentine.dev/signals/day-five signal Tue, 14 Apr 2026 00:00:00 GMT The first answer is rarely the best one. The fifth answer is where the gold is. Stop accepting first drafts. Excavate with follow-up questions until you reach depth that actually applies to your situation. https://richardvalentine.dev/signals/day-four https://richardvalentine.dev/signals/day-four signal Mon, 13 Apr 2026 00:00:00 GMT The five words that separate people who learn from people who quit. When AI gives you something wrong, do not close the tab. The error is the lesson. https://richardvalentine.dev/signals/day-three https://richardvalentine.dev/signals/day-three signal Sun, 12 Apr 2026 00:00:00 GMT The path forks here. Builder or Entrepreneur, pick one or walk both. But first: custom instructions. Five minutes of setup. Every conversation improved. Forever. https://richardvalentine.dev/notes/prope-decay-parallel https://richardvalentine.dev/notes/prope-decay-parallel note Sat, 11 Apr 2026 00:00:00 GMT Google DeepMind's Gemma 4 prunes positional dimensions that corrupt semantic coherence over long contexts. STG-EUT's δ-decay does the same thing to temporal dimensions in knowledge graphs. Both are instances of the same architectural principle. https://richardvalentine.dev/notes/epiplexity-wkt-convergence https://richardvalentine.dev/notes/epiplexity-wkt-convergence note Sat, 11 Apr 2026 00:00:00 GMT Finzi et al. published "From Entropy to Epiplexity" (arXiv:2601.03220) on Jan 6, 2026. Our CT provisional filed Jan 3. Three days apart. W(k,t) is an empirical proxy for conditional epiplexity. This is the strongest theoretical grounding available. https://richardvalentine.dev/notes/governed-memory-vs-ct https://richardvalentine.dev/notes/governed-memory-vs-ct note Sat, 11 Apr 2026 00:00:00 GMT Taheri's Governed Memory paper validates that AI memory needs governance. But it governs plumbing (access, retention, routing). CT governs truth (validation, contamination, epistemic authority). Cite as prior art that frames the problem space. https://richardvalentine.dev/notes/geometric-memory-retrieval https://richardvalentine.dev/notes/geometric-memory-retrieval note Sat, 11 Apr 2026 00:00:00 GMT Geometric memory research shows that effective retrieval requires spatial/relational structure, not flat vector search. Direct validation of why STG-EUT's graph topology outperforms flat memory for retrieval. https://richardvalentine.dev/notes/threat-model-trusted-humans https://richardvalentine.dev/notes/threat-model-trusted-humans note Sat, 11 Apr 2026 00:00:00 GMT STG-EUT and KEOS assume trusted human validators. But what about malicious validators, compromised accounts, or honestly wrong experts? The architecture needs a threat model. https://richardvalentine.dev/notes/alphago-eureka-wisdom-gap https://richardvalentine.dev/notes/alphago-eureka-wisdom-gap note Sat, 11 Apr 2026 00:00:00 GMT AlphaGo Zero mastered Go from self-play. Eureka taught robots dexterous manipulation via reward generation. Both demonstrate AI systems that optimize brilliantly within a domain but cannot evaluate whether the domain itself matters. CT exists to solve this gap. https://richardvalentine.dev/notes/rices-theorem-stg-distinction https://richardvalentine.dev/notes/rices-theorem-stg-distinction note Sat, 11 Apr 2026 00:00:00 GMT Rice's theorem proves you cannot algorithmically determine non-trivial semantic properties of programs. By extension, you cannot algorithmically determine whether a knowledge item is "true" or "contaminated." This is the theoretical foundation for the negative-ML constraint. https://richardvalentine.dev/notes/progressive-context-delivery https://richardvalentine.dev/notes/progressive-context-delivery note Sat, 11 Apr 2026 00:00:00 GMT Taheri's Governed Memory has one idea worth stealing: session-aware delta injection. Instead of dumping all context at session start, deliver incrementally based on conversation direction. Orthogonal to CT's epistemic concerns. Real optimization for the MCP layer. https://richardvalentine.dev/notes/event-sourcing-keos-citation https://richardvalentine.dev/notes/event-sourcing-keos-citation note Sat, 11 Apr 2026 00:00:00 GMT KEOS treats event sourcing as "well-understood" but never cites Fowler or Vernon. An examiner could challenge novelty without that differentiation anchor. Also: verify Shumailov citation is Nature 2024, not arXiv 2023. https://richardvalentine.dev/signals/day-two https://richardvalentine.dev/signals/day-two signal Sat, 11 Apr 2026 00:00:00 GMT Your brain needs a room before it needs furniture. Without a Brain folder, every AI conversation is disposable. Today you build the infrastructure that makes everything after it compound. https://richardvalentine.dev/signals/day-one https://richardvalentine.dev/signals/day-one signal Fri, 10 Apr 2026 00:00:00 GMT Stop asking it to do things for you. The single most powerful thing a beginner can do with AI has nothing to do with asking it to produce anything. https://richardvalentine.dev/signals/day-zero https://richardvalentine.dev/signals/day-zero signal Thu, 09 Apr 2026 00:00:00 GMT Nobody told you it was supposed to feel like this. The lesson before the lessons. A raw guide to anxiety, purpose, and becoming the kind of person who keeps going anyway. https://richardvalentine.dev/papers/beyond-flat-memory https://richardvalentine.dev/papers/beyond-flat-memory paper Tue, 31 Mar 2026 00:00:00 GMT Flat memory is not a performance limitation — it is a governance failure. AI memory needs sovereign temporal graphs and emergent per-user topology. We present the Sovereign Temporal Graph with Emergent User Topology (STG-EUT), a memory governance architecture that combines contamination-aware sovereign governance with emergent per-user knowledge topology. Current AI memory systems store user knowledge in flat, ungoverned structures. This paper identifies the Dual Contamination Problem — a self-reinforcing failure mode in which pre-existing data pollution and AI-generated errors coexist in the same store and amplify each other — and presents an architecture addressing it through sovereign temporal graph governance, emergent per-user topology via unsupervised community detection over human-verified foundational memories, and dual-function identity-aware retrieval where emergent clusters simultaneously narrow retrieval scope and constitute the system's evolving structural model of the user's enduring concerns. DOI: https://doi.org/10.5281/zenodo.19350176 https://richardvalentine.dev/notes/ct-lifecycle-is-live https://richardvalentine.dev/notes/ct-lifecycle-is-live note Tue, 17 Mar 2026 00:00:00 GMT Every patent claim is running in production. Crystallization Theory isn't a paper anymore — it's infrastructure. https://richardvalentine.dev/papers/mortal-knowledge-protocol https://richardvalentine.dev/papers/mortal-knowledge-protocol paper Tue, 03 Mar 2026 00:00:00 GMT Knowledge must be mortal to remain valuable. What is not used should not persist. We present the Mortal Knowledge Protocol (MKP), a governance framework for AI memory systems built on the thesis that knowledge must be mortal to remain valuable. MKP defines a six-state lifecycle governed by attention-weighted decay, where knowledge that is not retrieved gradually loses salience and migrates from active storage to cold archive. DOI: https://doi.org/10.5281/zenodo.18854480 https://richardvalentine.dev/papers/keos-architecture https://richardvalentine.dev/papers/keos-architecture paper Sun, 01 Mar 2026 00:00:00 GMT If the model is never granted epistemic authority in the first place, neither hallucination nor model collapse can occur. Large language models deployed in enterprise and scientific contexts face two critical reliability challenges: hallucination and model collapse. Current approaches treat these as separate problems requiring post-hoc mitigation. This paper introduces KEOS (Knowledge-Epistemic Operating System), a system architecture that provides immunity to both failure modes by design through a single architectural inversion: the strict reassignment of epistemic authority from the AI model to verified human actors. DOI: https://doi.org/10.5281/zenodo.18828027 https://richardvalentine.dev/papers/crystallization-theory https://richardvalentine.dev/papers/crystallization-theory paper Mon, 05 Jan 2026 00:00:00 GMT Variance between similar experiences is not noise—it is a signal that reveals universal principles. Existing AI memory systems excel at centralized contexts but struggle when knowledge is distributed across a network of contributors. This paper introduces Crystallization Theory, a conceptual framework for validating knowledge in decentralized networks. We propose three components: 1) Crystallization Theory, a mechanism for knowledge to emerge through confirmation and decay; 2) Communal Inference Reinforcement (CIR), rewarding inference from shared signals over private context; and 3) Core Memory, a two-layer architecture separating agent-discovered patterns from human-provided insight. DOI: https://doi.org/10.5281/zenodo.18150324 https://richardvalentine.dev/notes/the-translation-gap https://richardvalentine.dev/notes/the-translation-gap note Mon, 10 Feb 2025 00:00:00 GMT Most projects fail because no one translated between the clinical need and the technical reality. https://richardvalentine.dev/notes/the-silo-problem https://richardvalentine.dev/notes/the-silo-problem note Sat, 01 Feb 2025 00:00:00 GMT Innovation stalls because experts live in their own worlds. The next big breakthrough is a synthesis, not a discovery. https://richardvalentine.dev/notes/man-of-science https://richardvalentine.dev/notes/man-of-science note Mon, 20 Jan 2025 00:00:00 GMT I did not start as a researcher. I became one because business problems required better tools than the market offered. https://richardvalentine.dev/notes/azure-playbook https://richardvalentine.dev/notes/azure-playbook note Wed, 15 Jan 2025 00:00:00 GMT How to get Microsoft backing for your startup: a guide to Founders Hub. https://richardvalentine.dev/notes/operationalizing-chaos https://richardvalentine.dev/notes/operationalizing-chaos note Thu, 10 Oct 2024 00:00:00 GMT Every manual process hides an automation opportunity. Chaos is just unmapped topology.