The Cresques Problem

Abraham Cresques was a cartographer on Majorca in the late fourteenth century. He is remembered for the Catalan Atlas of 1375, a lavish portolan chart commissioned by the Crown of Aragon and eventually gifted to Charles V of France. The atlas is stunning — gold leaf on vellum, wind roses radiating across the Mediterranean, miniature kings seated on thrones in the African interior. But its most interesting feature is invisible: the gap between what Cresques could see and what he chose to draw.

Cresques had never been to Mali. He had never crossed the Sahara. Yet his depiction of Mansa Musa — seated, holding a gold nugget, with a caption explaining the wealth of his empire — is among the most detailed representations of West Africa produced in medieval Europe. The information came through merchant intermediaries, translators, and second-hand reports passed along trade routes that connected Majorcan Jewish communities to North African networks ¹.

This is what I call the Cresques Problem: the challenge of faithfully encoding knowledge that arrives through intermediaries, translation, and hearsay into a form that others can trust. It is not a problem about accuracy in the modern empirical sense. Cresques was not trying to produce a satellite photograph. He was trying to build a representation that would be useful — to navigators, to diplomats, to the king who commissioned it — while being honest about the sources and methods that produced it.

Every knowledge system faces this problem. Wikipedia faces it when an editor summarizes a source they found through another source. A machine learning model faces it when training data encodes assumptions from its collectors. A courtroom faces it when expert testimony translates technical findings into plain language for a jury ².

The usual response to the Cresques Problem is to demand provenance: show your sources, trace the chain, verify each link. This is necessary but insufficient. Provenance tells you where knowledge came from. It does not tell you what happened to it along the way — what was compressed, what was interpolated, what was silently assumed.

What Cresques actually did was more subtle. He used visual conventions to signal confidence levels. Coastlines he could verify through firsthand pilot reports were drawn with precise portolan geometry. Interior regions relied on a different visual grammar — decorative, narrative, explicitly interpretive. The map itself encoded its own epistemology ³.

This is the principle I keep returning to: the best knowledge systems do not merely store and retrieve information. They make their own limitations legible. They build the epistemics into the interface. When you look at a Cresques map, you can see where his knowledge was strong and where it was speculative — not because he told you in a footnote, but because the visual language shifts.

Modern systems rarely achieve this. A search engine returns ten blue links with no indication of which ones rest on primary sources and which are third-hand summaries. A news aggregator displays headlines in identical typography regardless of whether the underlying reporting involved a six-month investigation or a rewritten press release. We have more information than Cresques ever imagined, and worse tools for signaling what we actually know versus what we are merely repeating.