I tend not to use the word "law"... Probably if I had to choose a meaning, I'd associate the term with the bare equations, not a causal model. After all, we say pV=nRT is the "ideal gas law", but I hope it's clear (from above) that this law can be used many different ways, in different models.

(And, by the way, thanks for engaging in conversation about this! I can already tell it’s going to make the eventual paper much stronger.)

And, sure, in the interventionist view, it’s far more natural to picture *literally* putting on a jacket (or pushing a particle into some position), as opposed to “setting the outdoor temperature to cold”. But really, they’re both (imagined) external interventions in different-scale models.

I don’t think you’re using different notions of causality; you’re being very consistent about the interventionist framework, as far as I can tell. What’s happening is that you’re switching from model to model, and the causality is different in each one. Different models, not different definitions.

Right -- you're zooming out, asking bigger questions, as if the cause might be "out there, somewhere". And it *is* out there, IF you zoom out. And it will instead be out further if you zoom out further.

And if you don't zoom out at all, the cause is still right there, right where you started.

I suspect what you’re thinking here is you could always change the scope (zooming out) so that you could re-assign the “input” causes to position settings of something. But that’s changing the model.

I could also zoom out even further and say the cause was your decision to set that position! ;-)

"Which things are causes" is already ascribing causes to reality. "Which model variables are causes?" is much safer, for a given model.

And sure, models are sometimes "mental", but that doesn't mean there's not an objective fact of the matter about which DAG is the best one to use in a given case.

Well, I should add that, for an experimentalist, there's much more of a fact-of-the-matter about what the intervention variables are.

But, as I hope is clear from examples 1-4, you can change the experiment in a way that completely changes the causal model, and still use the very same equation.

Oh, no -- it's entirely model-dependent, in almost every case. The causation is *in the model*, in a very real sense, not in the bare facts of the system.

But when we're doing physics, we're using models, we're imagining interventions. We're using causal reasoning. That's real *reasoning*, anyway!

Good! I forgot to specify that the box has a piston on it (and is kept at the same temperature, etc).

But do you really think the location of the piston relative to the box is causing the pressure to change in cases #1 and #3?

I see the act of moving the whole box to a new pressure as the cause.

Ah, be careful here -- you're being tempted to either zoom in or zoom out, changing the scope of the model. As soon as you do that, it's a different model, with different causal structure. Stick with the scope of the model given, and I think everyone will agree.

I suspect most people will agree on the cause-effect relationship in all of these, even though it changes... And notice it's not the computational order! The cause is the variable that you are setting, from outside the problem. The effect is the variable correlated with that intervention.

3) A box of gas is moved between two regions with different pressure. You see the volume change by ∆V. Solve for ∆p.

4) A box is of gas is physically compressed, such that the change in pressure is ∆p. Solve for ∆V.

Here's a fun example I think I'm going to include in the paper... identify the cause(s):

1) A box of gas is moved between two regions, changing the pressure by ∆p. Find the change in volume ∆V.

2) A box of gas is physically compressed, changing its volume by ∆V. Find the change in pressure ∆p.

But any variable can either be observed or intervened upon, depending on context. There's nothing special about pressure or velocity here. Our causal reasoning is based on the perspective of what you think you're able to externally change, in principle. And that perspective is context dependent.

Of course, one point of the paper is that there’s no “one right way” to use an equation; the intervention variables change with context, and sometimes are not even used. But then it’s not really a model, just a set of facts. I still like my claim that all *models* allow counterfactual interventions.

That’s fascinating that you’re thinking in more causal terms about the ideal gas law than about projectiles! I have some students going through physics textbooks (for a paper about causality in physics) and they’ve found typically just the opposite: causal language for F=ma, no causation for pV=nRT.

Aren’t *all* physics models about counterfactuals?

Again: There's a difference between a solution and a model. Think of a solution as a single run of a model. Sure, a single output is self-consistent, but that's not enough. The *models* of subsystems of the universe have to be self-consistent, and that's where you define signals. No retrosignaling!

Yes, that's what I've been trying to say here. In a consistent retrocausal model, all the retrocausality has to be at the level of hidden variables. Hidden = no signal, and signals are defined in terms of agents. For consistency, we can't signal into the past, even if we can cause the hidden past.

That's right -- they're measuring phase shifts, and *inferring* things about the past. No direct measurement of what they're talking about is even possible.

That’s all I need to know; I’m not going to use your model, because it is telling me something false. I can’t signal into the past. If that’s what this model does, it’s wrong; it empirically fails. If you want anyone to consider it, you’d better fix this flaw; show why you can't signal backwards.

Look, Jarek. You’re an agent. I’m an agent. You’re modeling particles, with some causal model (related here in insufficient detail, that’s not important). You’re asking me to use the same model of these particles. But your model tells me that I should easily be able to signal into the past.

Thanks for your interest! The talk is supposedly going to be uploaded soon, at that same link. I'll re-post it when it happens.

I think your argument is blurring together (A) and (B), guessing that (B) will reflect the same symmetries as (A) on a smaller scale. But it can’t, at least not at the level of the input-output structure. That (B) symmetry will be broken by the asymmetric agents, even if (A) is nicely symmetric.

The connection is that agents must (somehow) result from the original model (A), and those agents must inherit a time-asymmetry (probably for reasons similar to what you have in mind). That asymmetry in turn informs both the correct (B) models and the feasible signals. That gives a signaling-arrow.

The only way to discuss signals is to discuss (B), where some agent (not an electron!) outside some particular subsystem is modeling that subsystem. That’s the sort of model that has forward- but not-backward- signaling, from the agent’s perspective (the agent who is using the useful (B) model).

This is a tough discussion because we don’t know the details, only broad strokes. But regardless, it’s important to distinguish between (A) a model of the whole universe, and (B) a model of some subsystem in that universe. You seem to be mainly thinking about (A), but there is no signaling in (A).

Thanks for coming! Wasn't it the middle of the night for you?

I've tried to explain my view, above. Maybe you could tell me your view (in text, please, not in those ultra-busy slides; I'm not looking at them).

You can easily signal to the future but not into the past. Why is that, do you think?

Happening soon, in two hours!

It looks like you might need to "register".