The trace view, one year on

When Recall 2.5 shipped in June 2025, the trace view was the feature we were proudest of. For every answer Recall gave, a user could open a panel and see how it got there: which documents were retrieved, in what order, which passages were weighed, where the reasoning branched and why one precedent was preferred over another. We described it at the time as showing your working, and we expected it to be used the way a careful associate checks a junior’s research memo: every time, line by line.

Sixteen months later, with the trace view in daily use across several hundred legal teams, we know how it is actually used. It is not the way we expected. This is an account of the gap between the two, and of what we changed because of it.

What we expected

We designed the trace view for distrust. The argument was straightforward and we still believe it: a search system that gives legal answers without showing its sources is unauditable, and an unauditable system has no place in a matter file. The trace was meant to be read, and read often, so that over time the firm would build a calibrated sense of where Recall was reliable and where it was not.

We expected opening rates above 50 per cent in the first months, declining as trust built. We expected the detailed view, with every retrieval hop and every discarded candidate, to be the default. We expected the trace to be used mainly by the person who ran the query.

What happened

The trace view is opened on about 21 per cent of queries across active accounts. That figure has been stable, within a few points, since late 2025. It did not start high and decline; it started at around 30 per cent in the first month and settled quickly.

When it is opened, the reason is nearly always the same. Users open the trace when the answer surprises them. We can see this because trace opens cluster on queries where the top result is from an unexpected practice area, an unexpected jurisdiction or an unexpected time period, and on queries where the user’s follow-up query, within a minute, contradicts the first. The trace is not a routine check. It is what people reach for when something looks wrong.

The second finding is who opens it. Partners open the trace view on a much higher share of their queries than associates do. Our best reading, from conversations with firms, is that associates use Recall to find a starting point and move on, while partners use it to check a position they already hold and want to see the basis for any disagreement. The trace is, in that sense, an instrument for arguing with the system, and partners argue with it more.

The third finding was the one we had not anticipated at all. The trace view is used to train juniors. Several firms told us that supervising lawyers open the trace on an associate’s query, sitting beside them, to show how a precedent search should branch: that a limitation-of-liability question in a software licence should pull from the firm’s MSA precedents as well, that a governing-law point should check for recent matters in that jurisdiction first. The trace shows a research path, and a good research path is teachable. We built an audit tool; some firms are using it as a textbook.

What we changed

The detailed view was too detailed. Showing every discarded candidate made the trace long and slowed the thing people actually wanted, which was to find the step where the answer went somewhere unexpected. In Recall 2.6, in September 2025, the default became a compressed view: the retrieval path, the decision points, and the precedents that made the final answer, with discarded candidates behind a toggle. Opening rates did not change. Time spent in the trace fell by about half, which we read as people finding what they came for faster.

Citations had to survive translation. The 2.6 release also made the trace coherent across languages: a trace on an English query that drew on German precedents now shows the German passage, the English rendering Recall worked from and the alignment between them. Before that, cross-language traces were technically complete and practically unreadable, which may be part of why opening rates settled where they did.

Speed mattered more than we thought. Recall 2.7, in January 2026, made the trace view available immediately rather than assembled on demand. The old on-demand assembly took two to four seconds, and we had assumed that was acceptable for a feature used deliberately. The number of traces opened and closed within three seconds, which we take to be people checking one thing, rose noticeably after 2.7. A trace that costs a few seconds is consulted; a trace that costs nothing is glanced at, and glances are most of what auditing looks like in practice.

What we got wrong

We thought trust would be built by inspection. It appears to be built by the absence of surprise. The firms with the highest trust in Recall, measured crudely by query volume per lawyer and by how often the first result is used, are not the firms that open the trace most. They are the firms whose queries rarely return anything unexpected, because their playbooks and matter structure are clean and Recall’s results reflect that.

The trace view is doing its job. The job is different from the one we described. It is not a working that is checked every time; it is a record that is there when it is needed, and the knowledge that it is there seems to be most of its value.

We would build it again. We would describe it differently.