On Mon, Feb 13, 2023 at 7:47 AM Kerim Aydin <[email protected]> wrote:
>
> The below CFJ is 4007. I assign it to G..
>
> status: https://faculty.washington.edu/kerim/nomic/cases/#4007
>
> =============================== CFJ 4007 ===============================
>
> The horses have been motivated this week.
>
> ==========================================================================
>
> Caller: ais523
>
> Judge: G.
>
> ==========================================================================
>
> History:
>
> Called by ais523: 13 Feb 2023 04:27:06
> Assigned to G.: [now]
>
> ==========================================================================
>
> Caller's Evidence:
>
> I just asked random.org to randomly pick one of Agora's horses (via
> shuffling the list and picking the first entry of the shuffled list)
> and a number in the range 1-3 inclusive, with the intention of using
> them for horse motivation if and only if they came out favourable to me
> (and not motivating the horses at all this week if they came out
> unfavourable - rerolling would mess with the distribution).
>
> The number was 1.
> The horse was Fargo.
>
> I like that result. As such, I motivate the horses, specifying Fargo
> (the galloper) and 1. (If this works, this causes Fargo's, Alexia's,
> and Destructor's Race Position to each increase by 1. See the CFJ below
> for a discussion on whether it worked or not.)
>
>
> Caller's Arguments:
>
> Does the process I used here qualify as "random", by Agora's
> definition? The probability distribution was initially random, and I
> couldn't predict the final choice in advance.
>
> The point of contention is therefore whether the possibility that I
> would reject a roll I didn't like would count as changing the
> probability distribution, causing the condition of rule 2505 to not be
> satisfied. However, because it's always possible to wait to see whether
> someone else motivates the horses before you decide to use your random
> numbers or not, this sort of bias always exists to some extent (getting
> smaller and smaller for the Horsened as it gets later and later on
> Sunday, but always there, and possibly confused somewhat by the
> possibility that the Horsened could succumb). As such, maybe the horses
> can't be motivated at all?
>
> Another possibility is that all that matters is that the *selection* is
> by a random process, and the *action itself* merely has to make use of
> the selected random numbers (if it occurs), but not necessarily in an
> unbiased/equiprobable way. (The Horsened has been taking other actions
> in between the random rolls and the motivation action, in a way which
> causes the motivation action to become more favourable for em, so e is
> also biasing the result of the motivation action in this sense.)
>
> ==========================================================================
Judge's Evidence:
Rule 2505/1 (Power=1)
Random Choices
When a Rule specifies that a random choice be made, then the
choice shall be made using whatever probability distribution among
the possible outcomes the Rule specifies, defaulting to a uniform
probability distribution.
The choice can be made using any physical or computational process
whose probability distribution among the possible outcomes is
reasonably close to that required by the Rules, and for which the
final choice is not trivially predictable by the selecting person
in advance. The selecting person SHOULD make the selection method
public, and SHOULD use a method for which the final probability
distribution can be readily confirmed.
Judge's Arguments:
Rule 2505 is essentially unchanged since it was enacted by proposal
7883 in September 2017. Proposal 7883 included a rather extensive
explanatory note as to its interpretation. While legislative intent
through explanatory notes is not rule text or precedent, it certainly
can provide clarity on definitions or term usage where the rules are
silent (and when it doesn't depart from common sense, precedent,
etc.). And the explanation in the proposal covers the exact situation
of this CFJ in the first sentence:
> [Note: "confirming the distribution" includes "confirming that the
> selector didn't just keep re-rolling until e got the desired result",
> (because that changes the final distribution for the whole process to
> be 100% whatever the selector was aiming for). The "trivially" in
> "trivially predictable" means you can use a PRNG, which is technically
> deterministic, but only "trivially" predictable if you run it with
> a set seed in advance. "Reasonably close" is to not get hung up on
> machine rounding, but also means we don't undo things if a calculation
> is *slightly* off (with *slightly* to be determined by judges etc.).
> Keeping it SHOULD instead of SHALL because I prefer to let the Officer
> have a little latitude. These are hard-earned lessons from the days
> when we dealt cards from a deck, and kept having to reset the whole
> deck because a single card had been missing from the draw].
https://mailman.agoranomic.org/cgi-bin/mailman/private/agora-official/2017-September/011834.html
So, when the rule text uses the term "final choice" it is including
steps in the generating process that may include decision points (an
example of an "acceptable" if deprecated decision point: "I need to
choose between 5 options and I've only got a d6 handy, so I'll re-roll
any 6s"). Further, I find that the "final" probability distribution
in question is only "final" when it is actually used (e.g. in this
case, when horses are motivated). This is because we *can't* know
the final probability distribution until the random number is actually
used. Even if a selector uses a third-party email service to send a
die roll to a public list, and the email confirms the die used, we
don't know that the selector won't follow it up with an email "To make
the final selection, I add +1 to the die roll and motivate the
horses". Until that is known, we don't know the "final" probability
distribution that was used for the selection. (Note: the "can't be
trivially predictable in advance" in R2505 refers to "in advance of
starting the random process" - if information such as initial die
rolls become known while the process is going on, that's fine).
So to answer the caller's question directly, taking eir rejection
procedure at face-value results in a final probability distribution
that is *not* reasonably close to the one required in the rules, so
the horses were not motivated this week.
The caller further contends that uncertainty in process may make
motivation impossible. It is true that we can never know the *exact*
probability distribution used. Again, if a player has a die server
email a number to the public list, intending to follow it up with a
motivation action, there was still perhaps a tiny probability that e
would be struck by lightning or otherwise be unable to complete the
roll (or in the "reroll 6s" example above, the vanishing probability
that the process doesn't terminate in reasonable time), so even in the
retrospect of a successful roll and motivation announcement, there was
a chance of "something else happening" as an outcome.
This is handled by the "reasonably close" clause in R2505, and it's a
sliding scale that's judged based on preponderance of evidence. On
one hand, if a dice server email is followed within minutes by a
motivation email that includes the die roll as expected, the chance
that something else could have happened is small enough for practical
purposes that the overall process was "reasonably close" to the
required distribution. Under such circumstances, it is absolutely
possible to motivate the horses. At the other extreme, let's say the
selector announces, after the dice server email, "the number was 4,
anyone else can use the number 4 to complete the random selection and
motivate the horses this week" (insert obligatory XKCD reference
here). In that case, we have absolutely no idea on the probability
that someone else decided to use 4 to do the motivation. So that
would fail.
In between those extremes, there's a wide range of edge cases. If an
hour passes between the die roll and the motivation email from the
same person, maybe there were shenanigans. If a whole day passes -
well the chance that the selecter follows through at some point
exceeds "reasonable" and perhaps it fails. Unfortunately, in using
the "reasonable" standard there's no hard and fast line, and the
integrity of the process must be based on preponderance of the
evidence. But in general, for at least the fairly-immediate case of a
few minutes delay and no evidence of e.g. roll withholding in the
process, the rule works as expected.
I find FALSE.
