Ideas are easy. Testable ideas are not.

The difficult part of research ideation is rarely producing another suggestion. It is finding a direction that is different enough to matter, credible enough to pursue and concrete enough to test.

Anyone who has run a lab knows the cost of getting this wrong. A student spends weeks surveying neighbouring fields. A group debates whether an idea is genuinely new or only sounds new. A promising direction dies because nobody could say what experiment would settle it.

General-purpose AI has made the first step cheap. Ask for ten research directions and you will get ten, fluently worded. But fluency is not mechanism. A smooth answer does not show why the predicted effect should happen, does not separate evidence from inference, and does not tell a team what to do on Monday morning.

That gap is why I built ScopeIdea, a web-based research ideation platform. It takes a technical brief and turns it into a ranked portfolio of research concepts, each designed to show how the idea would work, what it should be compared with and what experiment could begin to evaluate it.

Here is how it works. Several research engines examine the same brief in parallel, from different angles, including neighbouring research and analogies from other domains. Their strongest candidates are combined and ranked on novelty, potential impact, feasibility, supporting evidence, fit to the gap you stated and clarity of execution. You choose how bold the search should be, from Practical to Rupture.

What survives is developed into a card. Each card sets out a technical mechanism, an explicit comparator, a validation pathway, a readiness assessment and a decisive first experiment. I chose those fields because they force the right questions into view. What causes the predicted result? Against which baseline should it be judged? What observation would make me continue, revise or stop?

Here is what that discipline looks like, as an illustration and not as output from the platform. Suppose the brief is a sensor that detects a trace gas in humid air and lasts a year in the field. “Use a new porous material” is not an idea. A research-ready card has to say how the gas interacts with that material, and why the interaction should change a measurable signal. It has to name the comparator, an established sensor whose response to humidity is already known. And it has to propose the decisive first experiment: expose both sensors to the same gas concentration across the full humidity range, and ask whether the new response stays distinguishable from the humidity effect. If it does not, the idea stops there, cheaply. If it does, the next experiment is obvious. Either result is useful, and a vague idea can never produce one.

This is the same instinct behind my thinking about trust in sensors. A reading is only useful if you know how far to believe it, and an idea is only useful if you know how you would find out that it is wrong. A comparator and a first experiment are how an idea earns that.

I want to be careful about what a platform like this can claim. A score orders concepts according to stated criteria. It does not establish that the top-ranked idea is correct, or even new. ScopeIdea is decision support. It does not replace a literature review, domain expertise, experimental judgement or authorship. Evidence labels and comparators make scrutiny easier, but researchers still have to check the cited work, test the assumptions, assess safety and confirm novelty before acting on any recommendation.

I also want to be plain about the evidence. ScopeIdea is publicly available, but its effect on research outcomes has not been independently benchmarked. That is the evaluation that matters most, and I would rather it came from outside. A fair test would compare ScopeIdea with conventional brainstorming and with general-purpose AI, using blinded portfolios of ideas judged on novelty, technical validity, diversity and testability. It would also measure how often experts reject unsupported claims, and whether different reviewers reach consistent judgements.

Until that work is done, I would suggest a modest way to use it. Treat the output as a structured set of hypotheses to argue with. A doctoral student can move from a broad topic towards a defensible experiment. A principal investigator can compare several mechanisms before committing laboratory time. A development team can record why one direction advanced and another did not, which is a record most teams never keep.

Good research still depends on people who read the papers, run the experiments and are honest about what the results show. A tool can make the first draft of the thinking faster and more explicit. It cannot do the thinking for them.

If you work on a problem where the next step is unclear, try it on your own brief, and tell me where it fails. The failures will teach me more than the successes.