Loading TACITUS
NOTES BY TACITUS
Working notes on the grammar of conflict, the engine, and the work itself.
Short, technical, citation-bearing essays from TACITUS Research. Ontology design, engine internals (AGON · KAIROS · DIALECTICA), mediator and policy-analyst cognition, OSINT methodology, neurosymbolic architecture, benchmarks, and what it means to ship experimental tools for high-stakes reasoning. Every note has a TL;DR, sources, and a specific problem it answers.
CATEGORY
TAG
ALL · 17
newest first
·7 minEngine
AGON — the conflict-vision crate inside DIALECTICA.
AGON is the half of the neurosymbolic stack that recognises what kind of thing a span of text is — actor, claim, interest, constraint, leverage, commitment, event, narrative — under a strict typed contract. Rust. MIT. Hot path.
agonrustextraction
Read
·7 minEngine
KAIROS — temporal vision and the discipline of two clocks.
KAIROS implements Allen-13 interval logic, bi-temporal commitment status, and as-of / drift / retrospective queries. It is what makes the LLM-side claim that something happened on Thursday survive contact with reality.
kairosrustbi-temporal
Read
·8 minCognition
How mediators think — and why the kernel was designed downstream of that.
Mediators have a small set of cognitive moves: separate position from interest, distinguish claim from commitment, surface BATNAs and red lines, track narrative drift. Most of these moves map cleanly to ACO primitives. The kernel was designed downstream of this practice, not upstream.
mediationcognitionprimitives
Read
·6 minArchitecture
Seven graphs, one engine — why DIALECTICA partitions state.
Different parts of the system have incompatible mutability contracts. A single graph satisfying all of them satisfies none of them well. Here is why we partition — and why the partition is the engineering thesis.
dialecticaarchitectureengine
Read
·8 minCognition
How policy analysts think — and where the work breaks under generic AI.
Policy work decomposes into stakeholders, constraints, trade-offs, and reception. The part that breaks under generic AI is the constraint-stress test: which constraint binds first, which actor responds first, under which option. The typed graph gives a deterministic answer.
policyoptions-memoconstraint-stress
Read
·8 minSources
Media source analysis in 2026 — citation is not the bottleneck anymore.
Models can summarise everything. They cannot tell you which source is load-bearing. The structural moves: cite at the span level, attribute every claim to a source class, model reliability as a typed property, treat contradiction as data.
osintsourcesreliability
Read
·7 minArchitecture
Task atomization — the operating system underneath a neurosymbolic agent.
A neurosymbolic system is only as good as its task decomposition. The neural component handles fluent recognition; the symbolic component handles typed structure and validation. Atomization is what tells the orchestrator which is which.
neurosymbolicagentsorchestration
Read
·6 minMethods
Bi-temporal honesty — when valid-time and transaction-time disagree.
Every node and edge in the kernel carries two clocks: when something was true in the world and when the system learned about it. Most software collapses them. Conflict work needs both, separately.
bi-temporaldesigncommitment
Read
·6 minPositioning
The knowledge layer is not a chatbot.
A chatbot is a query interface. A knowledge layer is the institution's typed reasoning, persisted, contestable, and inheritable. Confusing the two is how policy and political teams end up paying for fluent prose with no provenance.
knowledge-layerpositioningarchitecture
Read
·6 minPositioning
Inheritance is the missing primitive of institutional knowledge work.
Institutional reasoning fails at handoff. Generic AI stores text; the knowledge layer stores typed reasoning that survives the analyst. The kernel is fixed, the extensions are logged, the provenance is mandatory.
inheritanceinstitutionsdesign
Read
·6 minOntology
The ontology is dynamic. The kernel is fixed.
Closed ontologies fail by maximalism (SNOMED). Schema-free agentic graphs fail by minimalism (Buehler 2025). The kernel-with-extensions pattern sits between, with a four-stage validator that keeps extensions honest.
ontologydesigndynamic
Read
·5 minOntology
Why we hold disagreement instead of resolving it.
Medical ontologies assume an authoritative voice. Legal ontologies eventually do. Policy and political work has none — the file is the disagreement. The system's job is to shape it, not to dissolve it.
contestabilitydesignethics
Read
·4 minPositioning
Why "experimental" is not a hedge.
We label our products experimental because they are. Honesty about which surfaces are stable and which are placeholders. Here is what we mean by it and what we ask of early users.
trustpositioning
Read
·7 minLLM
Why LLMs break on conflict.
Three structural failures — time, causality, provenance — each a property of transformer architecture, not a tuning problem. RAG fixes none of them at the reasoning step.
llmarchitecturefailure-modes
Read
·6 minCognition
Positions, interests, and what the graph can see.
Fisher and Ury split positions from interests in 1981. Forty years later, almost no software implements the distinction. Here is what happens when you do — and why the kernel tiers interest inference rather than guessing.
primitivesmediation
Read
·5 minOntology
What we mean by commitment — and why it is its own primitive.
Claims are what a party says. Commitments are what they are on the hook for. Conflating these is how careful negotiations become careless ones. The kernel keeps them distinct at the type level.
primitivesontology
Read
·4 minOntology
The ontology is not a schema.
A schema describes how to store data. An ontology describes how to argue about the world. Conflating the two produces yet another JSON standard and no shared grammar.
ontologydesign
ReadSTAY IN THE LOOP
New notes ship as the work does.
Subscribe via RSS or JSON Feed. Or send us what we missed — short notes about short notes are welcome.