A citation is not verification.
Explanation, citation, and traceability are not synonyms — they are three different verification regimes for three different classes of decision. The Verification Matrix maps which one your use case actually needs.
Three regimes, not one.
Most AI deployments conflate the three. They ship a citation and call it verification. Or they demand traceability where an explanation would have done the job — and burn six months of engineering for a consumer-grade FAQ bot. The first move is to separate the regimes.
Explanation
"Why did you say that?"
A coherent account of the model's reasoning. The user judges the answer on its face.
The answer is recoverable if wrong. Nobody is going to audit you for it.
Citation
"Where did that come from?"
A pointer to a specific source — a URL, a doc section, a record ID. The user or auditor can navigate to it.
Tells you the breadcrumb. Does not tell you it was the current version, the authorized policy, or a complete retrieval.
Traceability
"Was this the authorized, current, complete source at the moment of decision?"
A chain of custody: versioned source, completeness signal, governed retrieval path, deterministic provenance.
Expensive to build. Only worth it when the decision is high-stakes and the source is too complex for a human reviewer to verify by hand.
The Verification Matrix.
The first lens. Place your use case on two axes — how much a wrong answer costs, and how complex the underlying source is — and the right regime falls out.
Understanding when citations suffice — and when they don't.
Key Insight
Citations were designed for a navigable web of simple sources. They carry low-stakes work, and they carry high-stakes work when an expert reviewer is in the loop. They break when the source is complex enough that no human reviewer can read it all — and the decision is high-stakes enough that a missed dependency is a regulatory event. That corner is the only one that actually requires verification.
Which regime does your industry need?
The first matrix tells you when citations break. This one tells you who lives in which corner. Decision stakes on one axis, audit burden on the other — three regimes mapped across four quadrants.
Same thesis, different cut: stakes against the regulator, not the corpus.
How to read this
The diagonal carries the regime gradient — explanation in the permissive corner, traceability in the high-stakes / high-audit corner. The two off-diagonal quadrants both run on citations, but for different reasons: one to satisfy a paper trail, the other to carry trust between humans. If your industry sits in the top-right, a citation is a breadcrumb. You need the chain.
Citations are breadcrumbs. Build the chain of custody.
If your use case sits in the high-stakes / complex-source quadrant, retrofitting citations on top of an agentic pipeline will not get you there. The verification primitive has to be designed in.
- Versioned sources with deterministic retrieval paths
- Completeness signals, not just citation lists
- Provenance the regulator can replay, not the model can hallucinate