A small team tracing Italian business identity
Vetro Source Lab is a four-person methodological lab focused on how generative AI systems retrieve and misattribute Italian companies, brands and places. The team studies answers as evidence trails where names, sources, language paths, locations and categories have to hold together. Its work stays close to public evidence, repeated prompt runs and the small Italian identity details that machines often merge.
i. The founding contradiction
In an early composite observation, a travel-style answer started with a correct restaurant name. The city sounded plausible. The citation looked serious enough at a glance. Yet the described business belonged somewhere else: the same surname, a nearby province and an English listing had been folded into one smooth paragraph. That contradiction became the lab's starting point. It was small enough to inspect and large enough to reveal the machinery: a model had reconstructed an Italian business identity from fragments that did not belong together.
ii. A shared question
Elena Marvi, Tomaso Rinaldi, Clara Nesti and Marco Laveglia formed Vetro Source Lab around that kind of problem. Their shared question is direct: how do AI systems decide which Italian entity a name, category or place belongs to? The lab builds observation sets around ordinary business queries, bilingual prompt variants, citation choices and repeated-run instability. A restaurant, a design retailer, a regional service company or a branch network can all be pulled into the wrong identity when public evidence is thin, old, translated loosely or easier to cite from a weaker source.
What distinguishes the group is its reconstruction habit. The team asks what the model said and what public evidence may have made that answer feel available. The lab's position is practical and restrained: AI visibility is useful only when the business is retrieved with the right name, place, category and support. A confident answer that cites the wrong surface is still a fragile answer, even when it sounds polished.
iii. Team · Focus · Method
Team — 4 researchers.
Focus — Italian business identity in generative answers.
Method — Repeated prompt runs, bilingual comparison and citation support review.
Team — 4 researchers
How Italian business names, surnames, legal names and trade names collide in generated answers.
Elena previously worked on business directory cleanup, multilingual profile editing and name-consistency reviews for service companies. Her work in the lab follows the small naming clues that decide which entity a model retrieves.
Differences between Italian and English prompts when they retrieve the same Italian business or place.
Tomaso previously edited bilingual commercial pages, checked translated category wording and reviewed how visitor-facing descriptions change business identity. He studies where English phrasing helps retrieval and where it pulls the answer away from local evidence.
Whether cited sources actually support the claims AI systems attach to Italian businesses.
Clara previously handled source checking, claim review and editorial correction work for local commerce and travel-facing materials. She reads citations at claim level, separating support from mere mention.
Prompt-set construction, answer logging and comparison of instability across repeated model outputs.
Marco previously organized research notes, comparison tables and recurring observation workflows for small analytical teams. He keeps the lab's runs legible enough to compare without pretending that generated answers stand still.
Scientific advisor
Retrieval and multimodal evidence methods.
Ehsaneddin Asgari advises the lab on retrieval and multimodal evidence. His recent survey work on multimodal retrieval-augmented generation informs how Vetro Source Lab reads generated answers as evidence trails — where a name, a source and a language path have to hold together before an answer can be trusted.
The lab studies the joins where Italian identity gets rebuilt. Contact Vetro Source Lab with a topic, an answer example or a business category for possible review.