In an early internal test of the research agent, we ran two versions side by side. Version A was fast. It produced an answer in three seconds. Version B was slower. It took eleven seconds, and after the answer it showed a list of sources that linked back to the specific ordinance sections or FEMA panels that produced each factual claim.
We expected version A to win on perceived quality. People usually prefer fast.
They did not prefer fast. By a wide margin, the team preferred version B. When we asked why, the answers converged on a single theme. "I can check it." "I know where the number came from." "If I have to defend it in a meeting, I can."
That test settled the architecture.
The latency cost, measured
A research agent without citations runs a single generative loop. Query in, answer out. A research agent with citations runs a retrieval loop, a verification loop, and a generation loop with the sources attached. The latency multiplier, depending on the query, is between two and five.
For a question like "what is the setback in the C-2 district in Palatine, Illinois," the naive loop took 2.8 seconds. The cited loop took 9.4. That is an unambiguous user-experience penalty. Everyone notices nine seconds of waiting.
We kept the nine seconds.
Why the trade-off is asymmetric
A fast wrong answer is not slightly worse than a slow right answer. It is categorically different. The fast wrong answer enters the record. It gets put in a memo. It gets emailed to a client. It informs a decision that is hard to walk back.
A slow right answer, with a source, can be checked. If the source looks off, the professional can click through and verify. If the source is solid, the professional can move forward with confidence that would not exist otherwise.
The asymmetry is what decides the design. In a consumer context, fast wrong is recoverable. In a professional context, it is not.
Citations as a discipline, not a decoration
Making citations load-bearing meant removing every code path that could produce a factual claim without a source. That is harder than it sounds. Language models are optimized to complete text. If a generation step can invent a plausible number, it will, unless the architecture forbids it.
The mechanism is simple to state. Every tool call returns evidence along with its answer. The generation step is constrained to quote from or cite that evidence. If a question requires evidence the agent cannot retrieve, the agent says so.
That last part matters. "I could not find that section" is a valid answer. It is often the most valuable answer. A developer who hears "I could not find it" makes the phone call. A developer who hears a confident invention does not.
What we measure now
The metric we care about is grounding rate. What percentage of factual claims in a generated memorandum link to a retrievable source. For the core research tools, the rate is now above 97 percent. The remaining three percent are typically claims about the user's own context or about the agent's reasoning, which are explicitly not factual claims about the outside world.
We do not publish a confidence score to end users. The internal telemetry tracks grounding, retrieval hit rate, and citation click-through. The professional sees the answer and the sources. Everything else stays in the system.
The slower product is the better product
Nine seconds of research latency is a cost. It is also the most defensible nine seconds in a six-week entitlement timeline. We chose the slower product because the audience we serve cannot afford the faster one.