Researchers have built a chart-generation method for AI "deep research" tools that refuses to draw a graph it can't actually back up.
The method, called Frame-Evidence Co-Adaptation (FECA), targets a quiet problem in AI tools that scour the web and compile reports with charts. Most of these systems draft a chart plan first, picking the entities, time ranges, and comparisons, then go looking for data to fill it in. When the web only partially supports that plan, systems either force the mismatch through or run a check afterward that flags the problem without fixing the chart itself. FECA instead treats the chart plan as a flexible frame that evolves alongside the evidence: the frame directs the search, and what comes back determines whether the frame gets accepted, revised, or dropped before anything renders. Tested on 100 real research topics, it improved numerical accuracy without degrading report quality or chart usefulness.
This matters because AI research tools already ship charts straight into reports people treat as reference material. A chart that quietly overstates what its data supports is worse than sloppy prose, since a graph carries an implicit claim of precision that text doesn't. The useful part of this work isn't a benchmark number - it's the diagnosis: charts go wrong because the plan gets locked in before anyone checks what's available to plot.
Post-hoc fact-checking has become the default patch for AI generation problems everywhere, from citations to summaries. FECA is a reminder that catching an error after the fact isn't the same as building a system that doesn't make it.