RESEARCH LAB
TACITUS Research builds the layer that turns fragmented institutional reading — cables, reports, transcripts, statutes, briefs — into typed, source-bound, time-ordered context that the next analyst can inherit and the next model can reason against. A kernel ontology of eight primitives. Dynamic extensions learned per case. Ontology-Augmented Generation as the inference pattern. Contestability as architecture. Conflict reasoning is the first specialization, not the whole pitch.
RESEARCH PROGRAM
The technical bet is simple: frontier models need explicit structure in domains where uncertainty, incentives, timing, and provenance determine the answer. Generic chat does not survive the handoff. The knowledge layer turns each piece of institutional reading into typed graph state — and lets the next analyst, the next desk, and the next model inherit it instead of reconstructing it.
Generic AI gives you a chatbot on top of your documents. The knowledge layer gives you a typed, source-bound, time-ordered graph of the institution’s reasoning — actors, claims, interests, constraints, leverage, commitments, events, narratives — before any model writes a sentence. The graph is the artifact. The chat is one interface to it.
Institutional knowledge work breaks at handoff. The analyst leaves; the file becomes opaque. Standard tools store text. The knowledge layer stores typed reasoning: every claim provenance-bound, every commitment bi-temporal, every extension validated against a kernel. The next analyst inherits the graph, not the rumor of it.
Eight universal primitives sit at the core. They do not change. What changes are the per-case subclasses: a Commitment in HR mediation, a Commitment in a multilateral peace process, and a Commitment in a regulatory pilot are all Commitments — and they have entirely different shapes. The graph learns those extensions and validates them against the kernel.
The inference pattern that runs over the layer: typed (constrained by the kernel), grounded (every claim cites a span), contested (counter-claims are first-class objects), and temporally honest (every fact carries valid time and transaction time). The methodology paper is in flight; the reference implementation is in DIALECTICA.
Most ontologies — medical, legal, telecom — assume an authoritative voice. Policy and political work has none. The graph holds the disagreement structurally: claim and anti-claim, evidence and counter-evidence, commitment and broken commitment. The system structures the dispute; humans decide what to do with it.
Conflict reasoning is the first specialization, because it stresses everything: time, causality, provenance, commitment tracking, interest/position separation, narrative drift. TCGC measures it. Policy options memos, stakeholder analysis, regulatory contestation, and ADR work get their own specializations as they mature.
FIELD SIGNALS WE BUILD AGAINST
The 2025–2026 KG-construction literature has converged on hybrid schema-based + schema-free systems (Bian, Oct 2025). The pure-static and pure-dynamic positions are both losing. The kernel-with-extensions design is where the field is going.
OG-RAG (Sharma et al., EMNLP 2025) showed that grounding RAG in domain ontologies lifts factual recall by 55% and answer correctness by 40% over flat RAG. The TACITUS knowledge layer extends that result with bi-temporal grounding, contestability, and cross-domain extensions.
Anthropic’s 2025 work on context engineering and Letta’s Context-Bench (2025) name the same failure mode: agents accumulate self-generated context and accuracy decays in ways that look fluent. Inheritance — typed, validated, and kernel-bound — is how an institution avoids paying that bill twice.
REFERENCE POINTS
Ontology-Grounded Retrieval-Augmented Generation. +55% recall, +40% answer correctness over flat RAG.
LLM-Empowered KG Construction · arXiv 2510.20345Survey of LLM-driven KG construction: schema-based vs schema-free, dynamic memory for agentic systems.
GraphRAG · arXiv 2404.16130Graph-based retrieval for global sensemaking across large corpora.
NIST AI RMFTrustworthy AI framing: validity, transparency, accountability, governance.
The formal argument for an AI knowledge layer for policy and political work: typed context creation, inheritable institutional reasoning, Ontology-Augmented Generation, contestability as architecture.
Read the paperEight primitives — actor, claim, interest, constraint, leverage, commitment, event, narrative — that compose the universal grammar of institutional reasoning. Per-case extensions specialize the kernel into conflict, policy, mediation, regulatory work. MIT-licensed. Forkable.
Read the grammarOAG: the inference pattern of the knowledge layer. Typed by the kernel, grounded in source spans, contested by design, temporally honest. Comparison table against RAG, GraphRAG, OG-RAG, and Think-on-Graph.
Read the definitionThe first specialization measured. The TACITUS Conflict Grammar Corpus: 14 task types in v0.1, three more in v0.2 for dynamic-ontology behavior. The thing that makes claims about the layer falsifiable on conflict.
Read the methodologyShort essays on the knowledge layer, inheritance as architecture, dynamic ontology, OAG, contestability, and what it means to ship experimental tools for high-stakes institutional work.
Browse notesOPEN SCIENCE
The eight-primitive kernel is published, versioned, forkable. An ontology kernel that nobody can read or fork is not a kernel; it is a private database.
Every dynamic extension learned by the system is logged, typed, and reviewable. Schema drift is a feature with a paper trail, not a black box.
TCGC v0.1 is public; v0.2 will include dynamic-ontology task types. Other specializations get their own benchmarks as they mature. Results without a benchmark are opinion.
Each research surface flags what is validated, what is experimental, and what remains unresolved. Pretending to certainty in a field this young would be dishonest.
THE EDITORIAL SIDE
Concordia Discors is the editorial home for the thinking behind TACITUS: long-form essays on conflict, structure, institutions, and the grammar of disagreement. If the Vision paper is the thesis, Concordia Discors is the notebook.
“Conflict is not the absence of order; it is the shape order takes when more than one agent has a goal.”