How does Kindred make sure its sources are real?

Language models can invent plausible-looking citations. Kindred is built so that the references you see cannot be invented: the reference list is grounded by construction.

The short version

When Kindred analyzes a question, some phases search the web and gather real sources. The final reference list is then built only from those sources that were actually retrieved. Any reference that was not actually retrieved is removed before it reaches the page. So the Sources & References record is not the model's memory of what sounds like a citation; it is a record of documents the analysis really pulled in.

This is what “grounded by construction” means: the structure of the system, not a later check, is what keeps invented sources out. A reference has to come from the retrieved pool to exist at all.

Why this matters

A citation is only worth something if the source behind it is real. The most common failure mode for AI writing is a confident citation to a document that does not exist or does not say what is claimed. Building the reference list out of genuinely retrieved sources closes off the first half of that failure: the source is real, and it is one the analysis actually saw.

It also means a removal is not a quiet deletion. Every reference dropped for not being in the retrieved pool is escalated for operator review rather than silently discarded, because a dropped source is potential evidence that a model tried to invent one.

Going deeper · for technical readers

The web-searching phases pool the sources they actually retrieve into a single per-analysis pool. The final reference phase builds its list only from that pool, and a code-level filter drops any URL that is not in it. The filter runs in the one place where the model's full proposed source and the retrieved pool both exist, so a fabricated URL has no path to the page.

Each dropped source is captured at the moment it is dropped, with the tier it was proposed in, its model-assigned type, a bounded excerpt, and a conservative nearest-pool-match hint, then surfaced to an operator review queue. The capture never changes what is kept or dropped; it only makes the drop visible. The filtering decision and the escalation are independent of each other.

Grounding is also durable. Each phase's output, including the reference list, is persisted the moment it completes, so what you see is always read back from stored state rather than from a live stream that could be interrupted. The analysis runs as a detached job on the server; closing your laptop or losing the network does not affect it.

This account is consistent with our public security architecture, where the same guarantee is stated in engineering terms.

What this does not claim

Grounding by construction guarantees that a source is real and was actually retrieved. It does not, on its own, guarantee that the source says exactly what a claim attributes to it. That is the honest limit we state plainly everywhere: citations link claims to real retrieved sources, but the content of each claim is not yet mechanically verified against the source. That verification is on our roadmap.

So the right way to read a grounded reference is with confidence that the document is genuine and the awareness that the final judgment of whether it supports the claim is still yours to make.

Read the security architecture