The CFO clicks the number. What opens?

A CHRO is presenting attrition figures to the executive committee. The CFO points at one number and asks where it comes from. In many companies the honest answer is a person: whoever built the spreadsheet three weeks ago, and who is now on holiday.

We want that question answered in one click, by anyone in the room who is allowed to see that record. That goal shapes more of our platform than any other single decision.

A link is not a source

The research on AI citations is uncomfortable. A 2024 study of long-document question answering found that proprietary models reached a citation F1 of only around 0.5 on two benchmarks, meaning that less than half of the statements in their answers were fully supported by the citations attached to them. A Stanford study the same year tested commercial legal research tools built on retrieval and found hallucination rates between 17% and 33%, and counted a particular failure as its own category: the citation is real, but it does not say what the sentence claims.

A May 2026 study of fourteen research agents made the distinction precise. Their links were valid more than 94% of the time and relevant more than 80% of the time. The factual accuracy of the claims against the sources cited ranged from 39% to 77%, and fell further as the agents made more tool calls. The link works. The claim in front of it may not.

For HR that gap is the whole problem. Nobody will check a footnote. Everybody will act on the sentence.

What makes a citation real

A label that says “source: HR system” is not a citation. We hold every citation to four properties, and a figure whose citation fails one of them is flagged rather than shown as fact.

1. Addressable. It points to one record at one moment, with two dates: the base-pay field of one employee, effective 1 March, as read at 14:03 today.
2. Readable in place. One click opens it in the page the person is already using, under their own permissions.
3. Reproducible. The same query tomorrow gives the same figure unless the underlying record changed, and when it did, the change is visible.
4. Owned. Every source names its system of record and the team that maintains it.

Four kinds of origin, shown on the figure

Numbers come from different places, and a reader should see the difference at a glance. Every answer we design distinguishes read from a record, computed by code from records that were read, quoted from a document down to the paragraph, and the Coworker's own reasoning, labelled so it can never pass for data.

The check runs after the answer, in code

The model writes its answer first. A separate step, written as ordinary code with no AI in it, then takes every figure in that answer and looks for it in the data actually read. A figure it cannot find is flagged. Because this step works from the data rather than from the model's account of the data, the model cannot argue its way past it.

Today the check catches a figure that appears nowhere in what was read. We are extending it to the harder case the 2024 legal study named: a figure that exists in the data, but not in the record the sentence points to.

Code can verify a figure. It cannot verify a judgement, a name in a sentence or the tone of a recommendation. That is why reasoning carries its own label, and why a person reads the brief before it goes anywhere.

Where it is still hard

1. A restricted source. If the person reading a brief may not open the record behind a figure, the figure should not be in their brief either. A shared brief has to respect the reader's permissions, not only the author's.
2. Effective dating. An HR record has the date a change took effect and the date it was entered. A citation that keeps only one of them cannot explain why last month's headcount changed after the fact.
3. Figures built from two systems. A cost per hire that joins recruiting and payroll has two owners, and the citation has to name both.

Why it is also a legal question

The EU AI Act requires high-risk systems, which include those used for promotion, task allocation and the evaluation of workers, to record events over their lifetime, and requires employers to keep those logs for at least six months. A brief whose every figure opens to its record is what that record-keeping looks like in practice.

If you are building this yourself

1. Start with one report people already argue about, and make every figure on it open to its record.
2. Separate read, computed, quoted and reasoning visually from day one. Adding labels later is much harder.
3. Store both dates on every citation: effective for, and read at.
4. Measure claim support, not link validity. A resolving link proves almost nothing.