dec — the decision-context layer¶
dec represents decisions as holons: units that are at once a self-contained,
accountable deliberation and a part of a larger decision holarchy. It sits above a
data graph (e.g. a FHIR-derived graph) and captures what data models structurally cannot.
Namespace: https://w3id.org/iladub/dec#.
What FHIR (and most data models) cannot capture¶
- The deliberation space — the options considered and why one was rejected, not just the recorded outcome.
- Situated meaning — the frame that makes a fact decision-relevant (the same lab value means different things in different decision contexts).
- A holarchy of authority — local decision ⊂ protocol ⊂ guideline ⊂ policy.
- The deontic/commitment layer — obligations, expectations, validity conditions.
- Decision lineage — what evidence supported a decision, and what upstream change
should reopen it (
dec:revisitIf/onChange).
Core shape¶
A dec:DecisionHolon (⊑ prov:Activity) records: consideredEvidence,
constrainedBy, an optionSpace (≥2 — the no-change option counts), exactly one
chosen option (which must be in the option space — SPARQL-enforced), decidedBy (an
agent, human or automated), a rationale, optional confidence ∈ [0,1], produced
(the resulting node), governedBy (policy), partOf (holarchy), withinScope
(autonomy scope).
Why it matters¶
For agentic AI safety: an agent reasoning inside a holon knows its autonomy scope, the constraints reframing its options, the option space it must justify against, and how its decision composes upward. That is the difference between emitting a plausible output and making an accountable, re-evaluable decision.
It is also the most domain-neutral, transferable part of the project — the same holon models a clinical therapy switch or a trading hedge decision unchanged. That portability is what makes dec cleanly separable — a component designed to be upstreamed to HGA once the substrate supports strict decidability.