account policyholder member client insured counterparty prospect household user party How many words does your company have for “customer”? Try it. Count. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 01 / 27
The meeting you’ve all been in SECURITY the same request LEGAL the same request PROCUREMENT the same request ARCHITECTURE the same request Four times. Four vocabularies. Four documents. Four queues. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 02 / 27
Why good processes take two quarters Security review waiting… Legal review waiting… Procurement waiting… Architecture waiting… Final approval waiting… “Two quarters later.” Nobody is lazy. The structure serializes. The waiting compounds. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 03 / 27
“Do we already have one of these?” 2019 2021 last week Three tools doing one job — because nobody could check. Retirement never happens, because nothing tracks what serves what. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 04 / 27
Onboarding is the same everywhere NEW EMPLOYEE NEW TECHNOLOGY NEW VENDOR NEW CLIENT Same five questions. Different door. 15% expertise 85% overlap we re-do by hand, every time The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 05 / 27
The 85% is a language problem one fact “Acme processes payments” SECURITY System of record? Data classification Access owner LEGAL Contracting party Liability clause Jurisdiction PROCUREMENT Supplier ID Spend category Renewal date ARCHITECTURE Component Integration pattern Lifecycle state Four forms. Four vocabularies. Zero shared dictionary. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 06 / 27
Twenty-five translations — or ten. A B C D E 1 2 3 4 5 TEAMS SYSTEMS point to point — 5 × 5 = 25 translations every pair keeps its own phrasebook, by hand, forever A B C D E 1 2 3 4 5 TEAMS SYSTEMS the shared model through one shared model — 5 + 5 = 10 mappings add a sixth system: one new mapping — not five The language tax is quadratic. The shared map makes it linear. That is the whole business case, in arithmetic. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 07 / 27
A shared map of meaning Customer Product Purpose Risk Policy Outcome It’s called ontology — no, not oncology. A rigorously agreed map of what our words mean and how things relate. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 08 / 27
A glossary defines words. An ontology defines a world. GLOSSARY definitions of words TAXONOMY words in a hierarchy ONTOLOGY kinds, relationships, rules TOP-LEVEL ONTOLOGY one bedrock under every vocabulary KNOWLEDGE GRAPH your facts, bound to the meaning Data becomes information with context. Information becomes knowledge through relationships. The richer the relationships, the more a machine can safely do with them. ORIGINS Not an invention — a discipline. Aristotle catalogued the kinds of being 2,300 years ago. Formal ontology became mathematics in the 20th century. In 2021 the ISO standardized top-level ontologies (ISO/IEC 21838). Twenty years of life-science ontologies already run on exactly this bedrock. Boring, proven plumbing. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 09 / 27
Every vocabulary bottoms out in the same dozen kinds. A THING “the customer” persists — wholly there today and tomorrow A HAPPENING “the onboarding” unfolds — exists only across time A ROLE “policyholder” a part a thing plays — not a kind of person A QUALITY “the risk rating” always a quality OF something AN AGREEMENT “the contract” exists only while it binds its parties Your twelve words for “customer” disagree about the label — never about the kind. Agree once at the bottom, and you stop arguing at the top. That is the whole trick. ORIGINS Top-level ontologies are small on purpose — a few dozen kinds. One (BFO, ISO/IEC 21838-2) anchors hundreds of science and defense ontologies; another (UFO) underpins enterprise conceptual modeling. They earned trust the boring way: twenty years of production use. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 10 / 27
A deal with five parties is one fact — not ten lines. a pairwise map — which line is the deal? buyer seller guarantor regulator the asset one hyperedge — the deal, kept whole Reviews, agreements, programs — enterprise facts hold many things at once. A map that only draws pairs shreds them — and every shred needs a human to reassemble the meaning. ORIGINS Claude Berge formalized hypergraphs in the 1960s: edges that may join any number of vertices. First combinatorics; then database design, network analysis, computational biology — anywhere relationships refuse to come in pairs. Graphs of graphs. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 11 / 27
Three ways a machine can know. ONE FAMILY: INDUCTION — LEARNED, IMPLICIT THE OTHER: DEDUCTION — AUTHORED, EXPLICIT MACHINE LEARNING induction — it estimates REPRESENTS a fitted function — weights ANSWERS a likelihood: “what is likely?” GUARANTEE error bounds — on data like it trained on WHEN WRONG silent — drifts as the world shifts …! GENERATIVE AI induction — it samples REPRESENTS the joint distribution of language ANSWERS a sample: “what is plausible?” GUARANTEE none per output — fluency is not reference WHEN WRONG confident — confabulation HYPERGRAPH AI deduction — it derives REPRESENTS explicit facts + axioms — kinds, rules ANSWERS an entailment — with its proof GUARANTEE sound — every answer traces to premises WHEN WRONG explicit — “cannot be derived” Generative AI is still machine learning — induction at scale. Confabulation is sampling, working perfectly, with no world to check against. So pair the families: the model proposes; the graph commits — and the proof rides along. ORIGINS AI has always had two lineages: statistical learning (perceptrons, 1958 → deep learning) and symbolic knowledge (logic, semantic networks, description logics). For decades they took turns being fashionable. The current moment is the marriage: fluency from one lineage, truth-maintenance from the other. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 12 / 27
“Generative” means two opposite things. GENERATIVE AI GENERATIVE PROGRAMMING PRODUCES FROM a learned distribution a formal specification MECHANISM sampling — the plausible derivation — the entailed SAME INPUT TWICE different outputs identical bytes TRUST LIVES IN eyeballing every output certifying the generator once THE AUDIT QUESTION “does this look right?” “does the diff show drift?” One improvises; the other compiles. Put the improviser at the interface — and the compiler everywhere trust must accumulate. ORIGINS The word arrived twice. Czarnecki & Eisenecker (1990s) coined generative programming: deriving software from specifications — deterministic, like a compiler. “Generative AI” (2020s) names the other thing: sampling new instances from a learned distribution. Same adjective, opposite guarantees. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 13 / 27
Bottom-up meets top-down at a gate — not in a workshop. TOP-DOWN the dozen kinds — principled, consistent… and empty of your business THING HAPPENING ROLE QUALITY AGREEMENT ↑ admitted — kind declared, identity set, evidence attached THE SIEVE a machine-checked gate — pass with proof, or triaged with a reason customer acct holder risk review “the process” policyholder insured pty risk rating the contract ✕ duplicate of customer ✕ ten different happenings ✕ duplicate of policyholder triage — not dropped clustered on the graph, escalated to a human decision BOTTOM-UP a language model harvests candidates from your documents and systems — broad, cheap, in your words… and merely plausible Bottom-up gives coverage, in your words. Top-down gives identity and consistency. The middle is not a committee — it is a build step. Reconciliation becomes certification. ORIGINS Ontology engineering’s classic failure: hand-authored maps were too slow; text-mined ones too messy. “Meeting in the middle” meant a workshop that never ended. Two things changed: language models made harvesting cheap, and formal validators made admission checkable. The gate replaced the workshop. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 14 / 27
Stop hand-copying meaning. Project it. the shared model the security form the dashboard the review checklist code & API an AI’s answer each one: a generated view · evidence flows back in The four forms from earlier stop being four documents to keep in sync. They become four views of one model. Change the model — every view follows. The work remains; the copying, and the waiting, disappear. “The entire history of software engineering is that of the rise in levels of abstraction.” — Grady Booch ORIGINS Generative programming (Czarnecki & Eisenecker, 1990s): software that derives software from a higher-level description. Model-driven architecture made it a discipline; compilers made it ordinary. When meaning is formal enough, a form, a dashboard and an answer are compilations — repeatable, auditable, regenerable. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 15 / 27
Your Lens Is Yours. That's Not A Feature — It's A Principle. One Fact LEGAL SECURITY FINANCE ARCHITECTURE Same fact. Your tint. We never force our words on you — we learn yours. built in untool-lens — 11 lenses, live The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 16 / 27
A capability map is the index of “what we do” What we do Customer Onboarding Claims Handling Portfolio Review Identity Verification Risk Review Purpose Alignment Plain words. A hierarchy people, tools and AI read the same way. This is how “do we already have one of these?” gets answered at intake — before buying. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 17 / 27
Maturity is an honest ladder Aspirational. Empty on purpose. crawl walk … run A full top rung means the ladder is marketing. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 18 / 27
AI needs ground truth “Certainly! The answer is… whatever you like.” floating — confidently making it up your certified facts standing on the map — answering from what you certified The map keeps the AI honest. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 19 / 27
Parallel gates on one shared model BEFORE two quarters AFTER the shared model security — own lens, same book legal — own lens, same book procurement — own lens, same book architecture — own lens, same book weeks The work is still real. The waiting is what disappears. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 20 / 27
Knowledge compounds instead of evaporating every engagement lands in the shared model “the deck that lived in email” Every translation, every lesson, every mapping — kept. The next project starts smarter than the last one ended. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 21 / 27
Every architecture is a philosophy. Most were never chosen. THE RIGID GRID fixed categories, enforced forever essentialism — frozen the day it shipped reality outgrows the schema; every change is a migration INHERITED THE PURE CLOUD no kinds — only correlations connectionism — everything flows no identity, no “why”; drift is invisible INHERITED THE PILE every record its own truth relativism, at enterprise scale reconciliation debt, quadratic, forever INHERITED THE TRIAD kinds with identity, revised by evidence essentialist AND inclusive — world, mind and knowledge all in the map; change is an edit, not a migration CHOSEN Rigidity, drift and reconciliation debt are not technical accidents — they are philosophical positions, inherited unexamined. Choose the philosophy on purpose — all the way down to the silicon. ORIGINS Architectures inherit metaphysics. The relational schema froze Aristotle’s fixed categories; the connectionist platform revived Hume (only correlations); the document store chose Protagoras (each record its own measure). The triad stance is older and newer at once: kinds that persist, minds that see, evidence that revises. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 22 / 27
The triangle underneath everything. THE KNOWLEDGE what we agree we know THE WORLD what actually happens — events · evidence THE PERSON who sees, and means harvest & evidence up · accountability down your lens — sovereign, mapped, never overwritten experience A glossary nobody uses: knowledge without person. A confident hallucination: knowledge without world. A re-explanation meeting: person without shared knowledge. Everything in this deck is three corners, kept honestly connected. ORIGINS Karl Popper (1972) split reality into three worlds: physical things, minds, and objective knowledge — theories, models, maps. Chinese philosophy drew the same triangle as Ch’i, Hs’in and Li; Indian thought as Maya, Atman, Brahman. Triad philosophy (Ye, 2019) merges them — down to a three-valued logic: true, false, and “cannot be derived.” The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 23 / 27
Why This Onboards Fast LISTEN in your language MAP onto layers already built MODEL drafted, then certified PARALLEL side-by-side, thin review KEEP your lens. yours. The map moves fast because the scaffolding already exists. What’s kept, you keep. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 24 / 27
The Map Is Not The Territory THE MAP ≈ THE TERRITORY Maps go stale — someone has to keep re-walking the territory. A map drawn at the wrong zoom hides exactly what mattered. Mistaking the map for the territory is how confident nonsense happens. Drawing it carefully the first time is real cartography, not homework. The hard part is drawing it well the first time. Do it once, well — instead of redrawing it forever, badly. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 25 / 27
The only 8 words you need capability a plain-words name for something the company can do lens your own language, kept, laid over the shared map ontology the rigorously agreed map of what our words mean top-level ontology the dozen bedrock kinds every vocabulary shares hypergraph a map whose facts hold many things at once, whole maturity level an honest answer to “how ready, really?” ground truth the certified facts everyone — and every AI — reads kept map your lens — portable, yours, kept for good Screenshot this one. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 26 / 27
Where would YOU start? 1 Which process of yours is onboarding-shaped? 2 Where do two teams use different words for one thing? 3 What would your capability map show three-of? Pick one. Talk to the person next to you for 60 seconds. The Language Problem — 101 · accordant.work powered by untool™ · © 2026 Nicholas P. Clarke · Enterprise Arts · untool™ 27 / 27

The Language Problem

101 · 19 min · narrated by George & Lily