Lab

The Lab is where we publish our measurements: how we test the AI that prepares your notes, what it gets right, what it gets wrong and what it costs.

In short: with a corpus of official texts cited to the article or paragraph and checkable at the source, a low-cost model drafts memos close to those of a frontier model working on the same corpus with our method, at 8 to 16 times lower cost. The figures, the method and our caveats are in the articles below.

• PROFESSIONALSCan AI draft a tax memo? What we measuredWe had Claude draft the same six memos, with and without access to the texts. On six frozen cases, what makes the quality is Luca Paper’s corpus and our method: with them, Claude scores 7.83/10, and our pipeline, on an inexpensive model, 7.62/10 on average; without them, Claude stays around 6/10. These six cases were used to tune the pipeline.13 October 2026
• METHOD AND COSTSClose to Claude Sonnet 5 on the same corpus, at a fraction of the cost: what we measuredIn our pipeline, an inexpensive model drafts memos close to those of Claude Sonnet 5 working on the same corpus: 7.62/10 on average over four runs, against 7.83/10 on one run (n = 6 per run, one grader). The model API cost is 8.3 to 16.2 times lower, and every citation in the corpus of official texts can be checked at the source. Caveats: the pipeline’s working cases, 7.83 obtained with the corpus and our method, Claude’s cost a lower bound.13 October 2026