# Provenance and accuracy

> Research outputs carry their sources, and every step of the trail can be inspected, across the whole product.

## The trail, by output

| Output | Its trail |
| --- | --- |
| Literature answers | Citations to the papers, through [Lacuna's typed links](/lacuna) back to the originals |
| Generated documents | Every figure and table links to its exact location in the source PDF |
| Code claims | The sandboxed run itself; files, line numbers, outputs |
| Network answers | Arrive in your conversation with their context attached, from a verified researcher |

_[diagram: Claim, page, original. Inspect any step.]_

## The paper audits its own claims

Appendix A of the Lacuna preprint is a per-claim audit table: each extracted claim, paired with the evidence that supports it. The statuses include "Supported," "Supported (limitation)," and, in one row, "False (useful failure case)."

## On being wrong

Systems built on language models can be wrong, this one included. The design response is inspectability everywhere, plus measurement in public. We benchmark our own stack in a preprint ([arXiv:2606.26246](https://arxiv.org/abs/2606.26246)), and the numbers include the unflattering ones: on survey-grade citation, the best F1 in the test, Lacuna's own, is 0.052 against GPT-Researcher's 0.039. Better than the alternative and nowhere near solved, and published anyway. That is the standard this product wants to be held to.

## Your part

Click the sources. Follow a figure to its page. Ask Althea to show its trail when one is not visible, and treat a missing trail as information.