TL;DR
- 01The 2024 framing was: hand-design eight primitives and 41 typed classes; ship as a single closed schema.
- 02The 2026 framing is harder: hand-design the kernel of eight; let everything below grow per case, validated against the kernel.
- 03Closed ontologies fail by maximalism. Schema-free graphs fail by minimalism. Kernel-with-extensions sits between, with a four-stage validator.
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.
Why the 2024 move was right
A closed schema made disputes typable; made commitments distinct from claims; made positions distinct from interests. It is the reason every TACITUS product reads and writes the same graph.
Why 2026 had to evolve
- ▸A Commitment in HR mediation should not have the same shape as a Commitment in a peace process.
- ▸Pretending otherwise is the failure mode of every closed-ontology project that came before us (SNOMED, FIBO, LKIF have decades of accumulated hierarchical errors).
- ▸A schema-free agentic graph dodges that work and pays for it later, in cross-corpus incoherence (Buehler 2025 documents this).
The middle path
The kernel constrains. The extensions specialise. The provenance log records what each case taught the system. The forkable contract keeps it honest. Every extension passes a four-stage validator: schema → kernel invariant → extension invariant → provenance binding. Rejected extensions go to a queryable quarantine; they are not silently dropped.
SOURCES
