Every claim traces back to a specific public record — not a plausible-sounding generation.
The underlying records are reconciled and connected before any model touches them, not summarized on the fly from raw text.
A model's training data goes stale. Public records need to reflect what's true now, not what was true at some past cutoff.
Knowing what counts as a comparable sale, or which entity fields indicate common ownership, requires domain-specific logic a general model doesn't have on its own.
The same question, asked twice, should return the same sourced answer — not two differently-worded guesses.
A conclusion without a visible reason behind it isn't something a person can act on with confidence.
The failure mode isn't that AI produces wrong answers on purpose. It's that a fluent, confident-sounding answer and a correct, sourced one can look identical from the outside — until someone checks. A model asked "is this landlord's portfolio clean?" can generate something reasonable-sounding either way. What makes an answer trustworthy is whether it's actually grounded in a specific violation record, licensing status, and eviction filing that someone can go verify.
That's why AI sits inside the ALKARTIS pipeline as one component — connecting, summarizing, and explaining already-reconciled records — rather than as a replacement for the pipeline itself.