Evidence should travel with the answer

Most of the demos that made large language models famous share a quiet assumption: that the cost of being wrong is a retry. You ask for a poem, it's mediocre, you ask again. The feedback loop is instant and the blast radius is one browser tab. Under those conditions, fluency is a reasonable proxy for quality, and the interface can be a blank box.

The work I care about breaks that assumption. In a legal review, or a preflight go/no-go call, the cost of a wrong answer is asymmetric and delayed. You don't find out at the moment of the mistake; you find out weeks later, in a dispute or an incident report, when the decision has already compounded. There is no cheap retry. The whole game is catching the error before you act on it.

The question stops being “is this answer right?” and becomes “can a person check this answer before it becomes a decision?”

Those are different problems, and they want a different kind of output. A paragraph of confident prose is nearly impossible to check: to verify it you'd have to redo the work. What you actually need is for the answer to arrive with its evidence attached — the specific rule it applied, the fact it read, the source span it pulled from, and how sure it is. Not appended as a polite bibliography at the end, but structurally, so that pulling the evidence out would break the answer.

Grounding is a property of the pipeline, not the prose

This is where retrieval earns its keep. A generation-only system invents text and then, if you ask nicely, invents a citation to match. A retrieval system does the opposite: it finds the passages first, and the answer is assembled from them. When you build it that way, the citation isn't decoration — it's the thing the sentence was made of. You can highlight the exact span that supports each claim because that span is a real object in the system, not a plausible-looking string.

You can see the mechanics of this in the retrieval visualizer I built for this site: a question moves through hybrid retrieval and reranking, and each sentence of the answer points back at the passage it came from. When coverage is thin — when the retrieved passages don't actually support the claim — the system's confidence should fall, and it should say so.

What this costs

Honesty about the tradeoff: this is slower and less magical. Grounded systems refuse more often. They surface uncertainty instead of hiding it behind confident phrasing. They make you build ingestion, retrieval, extraction, and review as one connected thing rather than wrapping a model in a chat UI over a weekend. The demo is less impressive and the product is more trustworthy, and in high-stakes work that is the correct trade every time.

It also changes what "better" means. A generation-only system gets better by sounding more authoritative. A grounded one gets better by being more checkable — tighter spans, cleaner traces, better calibrated confidence, fewer claims that outrun their evidence. Those are the metrics I actually watch, because they're the ones a skeptical professional will watch when they decide whether to trust the tool with something that matters.

Where I've landed

Every system I'm building points at the same principle from a different angle. In Littman, it's a legal decision that carries the rule, fact, and citation behind it. In FlightReady, it's a risk score that decomposes into per-factor contributions and a tail you can inspect rather than a single opaque number. The unifying idea is small and, once you've been burned by a confident wrong answer, hard to un-see:

The hard part was never generating an answer. It's showing which rule, which fact, and which precedent produced it.

Build the second thing and the first mostly takes care of itself. Build only the first and you've made something that demos well and can't be trusted with anything you'd be sad to get wrong.