Why every Luminol answer cites its source

Every answer Lumerian gives carries its evidence. Prism's code finds the quote behind every claim, verbatim or near-verbatim, on the cited page. Aperture links every finding to its crop, its module and the expert who audited it. Helm's assistant produces no numbers of its own. Three ways that document and image models fail explain why, and so does the way the people we build for make decisions.

01Right answers, wrong sources

Most document benchmarks score only the final answer. CiteVQA scores the answer and the region cited for it together. Across 711 PDFs averaging 40.6 pages, the best of 20 models scored 76.0 on that joint measure and the best open-source model 22.5; models often gave the right answer while citing the wrong place ^[1]. Without the citation, those answers would have looked correct.

02What models leave out

On long, multi-document legal cases, current models are more prone to leaving key information out than to inventing it, and they do worse as cases get longer ^[2]. A missing easement or a missing work order raises no error. The defence is an answer that shows what it read. Prism reads every page up to a 3,000-page budget and lists each unread page with its reason, so a reviewer can see what was left out.

03Real documents are harder

Benchmarks are mostly clean. Across 58 document-parsing models, scores fell by an average of 11.61 points from clean renders to real captures ^[3]. Title instruments are often decades-old scans; inspection imagery is captured in wind and glare. A cited answer lets a person look at the scan for themselves.

04How the people we build for decide

Engineers, lenders, title teams and crews sign their names to decisions, and recent public AI projects in this sector keep the final call with a person. When PJM reviewed site-control evidence for 811 interconnection applications with Tapestry's HyperQ, the system returned page-level citations to the PJM engineers who made the final determinations ^[4]. PNNL's benchmark for drafting environmental reviews grounds each draft in sources with pointers for expert review, and its researchers state that the systems still require human oversight ^[5].

05When there is nothing to cite

Sometimes the evidence is not there, or not clear. Then the right answer is to say so. Conformal risk control gives a formal way to decide when a system must abstain, and it comes with a hard floor: if a step's error rate is above the tolerance, some share of items has to go to a person whatever the method ^[6]. In Prism today, a row whose evidence is silent reads Unknown and goes out with an information request, and every row rated 4, the top of the scale, goes to a person. In Aperture, a conflicting module identity goes to the identity review queue.