Experiment Non-Repudiation

This is not meant to be long or very technical. It is just me trying to put words on something I have been thinking about for a while: I think computational science is missing a whole field, and I want to give it a name.

NB: when I say science here, I mean any field that relies on computation to produce its results. Computer science first, but also physics, biology, economics, climate, etc…

The question

Every number in a published paper is, in a way, an unsigned statement. We sign contracts, emails, bank transfers, even software releases. But the results that science is built on? We don't sign them. A paper says a model reached 94.7% accuracy, a simulation converged, a solver ran in 3.2 seconds, and we just take the author's word for it.

So here is the question I keep coming back to:

‘Given a computational result, can its author later deny that it came from this code, on this machine, at this time? And can anyone else falsely claim that it did?’

Today, the honest answer to both is YES. Results can be changed after the fact, failed runs can disappear, code can be swapped, and nothing leaves a trace. And it works the other way too: an honest author has NO way to prove that their number is really the one their code produced. Nothing binds the result to the computation. I call this the repudiation problem.

What Turing did

In 1936, Turing didn't ask “what can computers do?” in general [2]. He asked one very precise question: can a machine decide whether another machine will halt? To even ask it, he had to define exactly what a machine is. Then he showed where the line falls between what can be decided and what cannot. That one question, and that one line, gave birth to a whole field.

I think we need the same kind of move for experiments. Not ‘is this result true?’, because honestly, no one can answer that from the outside. But a narrower question, one that can actually be answered: can this result be denied?

Evidence is testimony (for now)

Let's be honest about what “evidence” means in computational science today. The author reports a result, and someone may be able to rerun it. Mostly, no one does. Code changes, environments drift, the hardware is gone, the data is private.

Let's consider this example: a conference C receives 30,000 submissions. A reviewer R gets a paper whose experiments ran on 200 GPUs for six months. Even with the code, R cannot check anything in a two-month review window. R decides based on how convincing the paper is, and on trust. So in practice, evidence is testimony. We believe the number because we believe the person.

I think this definition has to change. A computational result should count as evidence only when it is bound to the computation that produced it, so that its author cannot deny it and no one else can forge it. Evidence should move from what the author SAYS to what the computation itself commits to. To me, this is not an engineering detail. It is an epistemic change in what it means for a computational claim to be known.

A simple definition

Let's keep it simple. An experiment is code C, run on a machine M, at a time t, producing a result r. A passive observer watches the run. It never reruns the code and never reads the data. Then a third party, holding a key K that the author does not control, signs what the observer saw:

σ = Sign_K( h(r ∥ C ∥ M ∥ t) )

Verify_K(σ, r, C, M, t) = 1   ⇒   r was produced by C on M at t

A result is non-repudiable when two things hold: the author cannot produce a valid σ for a run that never happened, and once σ exists, the author cannot disown it.

Where the line falls

Like Turing's question, this one has a limit, and I think it is important to say it clearly. A passive observer can make the origin, integrity, time and completeness of a computation undeniable. It CANNOT make its correctness undeniable.

Why? Because to know a result is correct, you have to redo the work or trust the inputs. An observer that does neither can tell you exactly what ran, on what, and when. It cannot tell you the method was right. And that is exactly why this is its own field. Reproducibility asks: can the result be obtained again? Verifiable computing asks: was the calculation done right? Experiment non-repudiation asks something else: can the result be denied?

How big is this?

Honestly, I don't know. I don't think anyone knows yet how big this field could become. It touches every discipline that computes. It touches how papers are reviewed, how benchmarks are trusted, how AI agents are held accountable, and probably a lot more that we haven't thought of.

What I do know is what happens if we ignore it. If computation cannot produce evidence that stands on its own, results become claims. Claims that cannot be checked invite shortcuts. Shortcuts that cost nothing spread. And slowly, one unverifiable number at a time, trust wears away until the whole thing can't hold its own weight. Put simply: either computation learns to prove itself, or it slowly stops counting as science.

End Notes

In our NeurIPS 2026 position paper, we argue that computer science conferences should require non-repudiable experimental results [1]. K-Veritas is a first working instance of this idea: a passive observer, a third-party signature, every run counted, and a public record anyone can check. It is a starting point, not THE answer.

The field is open. I hope others will join and shape it, criticize it, and take it further than we can alone. Because at the end of the day, verification should not depend on who you are, how famous you are, or which lab you come from. Everyone's work should be verifiable.

@misc{keita2026computerscienceconferencesrequire,
  title={Computer Science Conferences Should Require Nonrepudiable Experimental Results},
  author={Mamadou K. Keita and Christopher Homan},
  year={2026},
  eprint={2605.08586},
  archivePrefix={arXiv},
  primaryClass={cs.CR},
  url={https://arxiv.org/abs/2605.08586},
}