Replace peer review with “peer replication” (2021)

99% of science is a waste of time, not just the papers. We just don't know which 1% will turn out not to be. The point is that this is making progress. As such, these 99% definitely are adding to the collective knowledge. Maybe they add very little and maybe it's not worth the effort but it's not nothing. I think one of the effects of AI progress will be allowing to extract much more of the little value such publications have (the 99% of papers might not be worth reading but are good enough for feeding the AI).

Is PL theory actually science? Although we call it computer science, I don't personally think CS is actually a science in the sense of studying nature to understand it. Computers are artificial constructs. CS is a lot closer to engineering than science. Indeed it's kind of nonsensical to talk about replicating an experiment in programming language theory.

For the "hard" sciences, replication often isn't so difficult it seems. LK-99 being an interesting study in this, where people are apparently successfully replicating an experiment described in a rushed paper that is widely agreed to lack sufficient details. It's cutting edge science but replication still isn't a problem. Most science isn't the LHC.

The real problems with replication are found in the softer fields. There it's not just an issue of randomness or difficulty of doing the experiments. If that's all there was to it, no problem. In these fields it's common to find papers or entire fields where none of the work is replicable even in principle. As in, the people doing it don't think other people being able to replicate their work is even important at all, and they may go out of their way to stop people being able to replicate their work (most frequently by gathering data in non-replicable ways and then withholding it deliberately, but sometimes it's just due to the design of the study). The most obvious inference when you see this is that maybe they don't want replication attempts because they know their claims probably aren't true.

So even if peer reviewers or journals were just checking really basic things like, is this claim even replicable in principle, that would be a good start. You would still be left with a lot of papers that replicate fine but their conclusions are still wrong because their methodology is illogical, or papers that replicate because their findings are obvious. But there's so much low hanging fruit.

I think the current system is just measuring entirely the wrong thing. Yes, fewer papers would be published. But today's goal is "publish papers" not "learn and disseminate truly useful and novel things", and while this doesn't solve it entirely, it pushes incentives further away from "publish whatever pure crap you can get away with." You get what you measure -> sometimes you need to change what/how you measure.

> If someone cares enough about the work to build on it, they will replicate it anyway.

That's duplicative at the "oh maybe this will be useful to me" stage, with N different people trying to replicate. And with replication not a first-class part of the system, the effort of replication (e_R) is high. For appealing things, N is probably > 2. So N X e_R total effort.

If you move the burden at the "replicate to publish" stage, you can fix the number of replicas needed so N=2 (or whatever) and you incentive the orginal researchers to make e_R lower (which will improve the quality of their research even before the submit-for-publication stage).

I've been in the system, I spent a year or two chasing the tail of rewrites, submissions, etc, for something that was detectable as low-effect-size in the first place but I was told would still be publishable. I found out as part of that that it would only sometimes yield a good p-value! And everything in the system incentivized me to hide that for as long as possible, instead of incentivizing me to look for something else or make it easy for others to replicate and judge for themselves.

Hell, do something like "give undergrads the opportunity to earn Master's on top of their BSes, say, by replicating (or blowing holes in) other people's submissions." I would've eaten up an opportunity like that to go really really deep* in some specialized area in exchange for a masters degree in a less-structured way than "just take a bunch more courses."

> Note I didn't say your machine, I said one machine of yours.

This thread discusses peer replication, this is not even an analogy.

I think you may be seeing the purpose of these papers differently to me, which may be the cause of this confusion.

The way you're describing a scientific publication is as if it were the end result of the scientific act. To use the software analogy, you're describing publication like a software release: all tests have been performed, all CI workflows have passed, QA have checked everything, and the result is about to be shipped to customers.

But talking to researchers, they see publishing more like making a new branch in a repository. There is no expectation that the code in that branch already be perfect (hence why it might only run on one machine, or not even run at all, because sometimes even something that doesn't work is still worth committing and exploring later).

And just like in software, where you might eventually merge those branches and create a release out of it, in the scientific world you have metastudies or other forms of analysis and literature reviews that attempt to glean a consensus out of what has been published so far. And typically in the scientific world, this is what happens. However, in journalism, this isn't usually what happens, and one person's experimental, "I've only tested this on my machine" research is often treated as equivalent to another person's "release branch" paper evaluating the state of a field and identifying which findings are likely to represent real, universal truths.

Which isn't to say that journalists are the only ones at fault here - universities that evaluate researchers primarily on getting papers into journals, and prestige systems that make it hard to go against conventional wisdom in the field both cause similar problems by conflating different levels of research or adding competing incentives to researchers' work. But I don't think that invalidates the basic idea of published research: to present a found result (or non-really), provide as much information as possible about how to replicate the result again, and then let other people use that information to inform their work. It just requires us to be mindful of how we let that research inform us.