The Hierarchy of Verifiability
Verifiability is not a single feature. It is a property that emerges from a hierarchy of five distinct conditions. Fail at the base, and the entire reasoning chain collapses.
The nested order — five layers, in this sequence
Five properties must hold at once — and they must hold in this specific order. You cannot see the inner concept until the outer layer is intact. Strip away the foundation and every layer above it collapses.
The walkthrough — one story, five failures
Five failures, one patient, one system. Each failure is what happens when the layer below is missing.
A 61-year-old patient with Type 2 diabetes is admitted with acute chest pain. The hospital's AI clinical decision support system is asked to recommend a management pathway. Everything the system needs to answer correctly has been ingested into the knowledge graph — every guideline, every lab result, every patient history note. The question is what the system does with it.
Consistency
FoundationDoes it give the same answer every time?
The attending physician runs the query on admission. The system recommends aggressive anticoagulation. Six hours later, a colleague reruns the identical query — same patient, same data — and gets a conservative management recommendation. The retrieval engine pulled different clinical guideline documents on each run because its vector similarity scores are probabilistic.
Two clinicians now hold contradictory recommendations from the same system for the same patient. Neither knows the other ran it. The patient receives a management decision based on whichever clinician acts first. The inconsistency surfaces only in a post-incident review.
This is not one of five properties — it is the prerequisite for all of them. If retrieval shifts between runs, tracing is meaningless. You cannot trace a different path than the one that produced the answer. Consistency is where verifiability either begins — or ends.
Traceability
Can you show exactly where the answer came from?
Domain alignment
Did it understand the field — or just pattern-match words?
Completeness
Did retrieval surface every applicable context — or stop short?
Signal clarity
Was what it used actually relevant — or was there noise?
The intersection — where all five hold
When all five properties hold: every clinician who runs this query gets the same recommendation. The system cites the specific ACC/AHA guideline section, the HIT contraindication node in the clinical ontology, and the renal function result — all retrieved because the traversal was complete. The context passed to the model contains only the three nodes directly applicable to this patient's presentation. The attending physician can walk the family, the regulator, and the court through each step of the reasoning, source by source. That is accountable clinical AI.
Where verifiable AI lives — and why most systems don't
Verifiability is not for every use case. It's the requirement for regulated, high-accountability industries — and most AI deployed in those industries today fails the test.
Most regulated industries today operate in the danger zone — high-accountability decisions made by probabilistic systems. Verifiable AI is the architectural answer to that gap.