Evidence beats novelty

In sectoral-AI calls, the panel has stopped believing "state of the art." Here is what it rewards instead.

The most common mistake in AI and smart-industry applications is leading with how advanced the model is. Evaluators in these calls have read hundreds of proposals claiming state-of-the-art performance with nothing behind the claim, and they have learned to discount the phrase on sight. What scores is the opposite posture: a measured result, a real benchmark, a named deployment environment, and an honest statement of technology readiness.

"We improved throughput by a measured margin on this dataset, validated in this operational setting" outscores "our novel architecture" every time. The reason is structural. These calls fund adoption and impact, not research for its own sake. The panel is looking for evidence that the thing works and will be used, not that it is clever. Cleverness without evidence reads as risk.

The discipline, then, is to resist the novelty reflex and let the data carry the case. Where there is no evidence yet, the right move is not to inflate. It is to say so and scope the work to produce it. An honest, lower TRL that the panel believes beats an impressive one it does not. Credibility is the scarce resource in a crowded field, and it is built with numbers you can stand behind.

From our practice · Eucade