Randonautica's MMI statistics

D0ublezer0 · June 2, 2025, 9:11pm

Randonautica’s data processing era has begun!
And we can share some statistics about MMI usage in the app.

First I will remind you some terms:

Randonautica is an app, that uses PQ128MS quantum RNG to produce entropy, that is converted by binary word method into a uniform set of coordinates. Then it searches for density deviations among those dots and returns user a single location on the map, corresponding to maximum density deviation. Experiment tries to find out if user’s intention is able to influence the location generation in order to find the place of it’s manifestation. After reaching the location, users write a report to the app. Binary word is not the best method to convert entropy into dots, but the only suitable one for server with millions of users.

In randonautica there are several point types.
MMI related:
Attractor - the densiest area of random dots. It’s density shows the probability of Psi-influence and measured by Power and z-score values.
Void - the sparse area, that dots are avoiding. Can be analogous to Psi-miss, but also anomalous.
Strongest - Attractor or Void, depending on which z-score is bigger.
Non-MMI:
Pseudo - random point, created with pseudo-RNG
Quantum - random point, created with quantum-RNG

Before generating a point on the map, users are setting an intention, and after visiting it, they write a report. So we used ChatGPT to process their reports and find out, if their intention manifested during the trip or not. Thats how Manifestation Rate (MR) is measured.

So what do we know from the data processing:

We processed over 180k reports. Total manifestation rate is 17.4%.
If we compare MR for different point types, it looks similar among them, except for Attractor type, that has 1-2% bigger manifestation rate.

Type = “Attractor” (Total lines of this type: 19220)
Fulfilled=True: 2207 (19.13% within this type’)
Type = “Void” (Total lines of this type: 17295)
Fulfilled=True: 1621 (16.91% within this type)
Type = “Strongest” (Total lines of this type: 111152)
Fulfilled=True: 10050 (17.08% within this type)
Type = “Pseudo” (Total lines of this type: 15622)
Fulfilled=True: 1355 (16.68% within this type)
Type = “Quantum” (Total lines of this type: 17778)
Fulfilled=True: 1858 (18.09% within this type)

So in general, MMI affects Manifestation rate in a pretty small way. But what is interesting, that only “Attractor” point type has this 1-2% MR-boost.

Then we tried to compare average z-score and power values for total reports array and the sample, where intent was manifested. (z-score here is always bigger than 4 because algorithm is selecting only big values for output to increase the probability of MMI-influence)
Average Values:
Z-Score (All lines, abs.): 4.2285
Power (All lines): 3.6429
Z-Score (Fulfilled=True, abs.): 4.2343
Power (Fulfilled=True): 3.6571

So manifested intentions has 0.01% bigger density deviation, which is even smaller, so we tried to see different z-score/power ranges separately:

Good news is that Attractor type has bigger MR on the magority or ranges, so it is probably not just by chance. Also we can see, that MR is rising on bigger density deviations up to 30%, but the bad news is that really big deviations are rare, and that means, that bigger values has smaller sample size and less reliable. So the rise may be a statistical illusion here.

To be sure, we tried to look for users, who produce the biggest z-scores. If they have stronger Psi, it may prove MMI is working.
First we tried to look in top 100 users by their average z-score, taking into account that they should have at least 10 trips. Among those users MR is distributed pretty similar to random sample, so it doesn’t prove anything. Also their minimal z-score value is a minimal possible one (4.0). We decided to look for users, who have the biggest minimal z-score, assuming that their influence on MMI-interface is constant and shows up in 100% of trips. We found only 6 users with min. z-score = 4.1, that is not very big actually and their MR was below 20% and most of them are below 10%. So still not sure about it.

Semantic research of reports are still in the process.

WanderingIshiki · June 10, 2025, 9:39am

Awesome to see some data analytical results. I know you guys have been wanting to do that for years. Cheers for the share!

D0ublezer0 · June 16, 2025, 9:37am

A little summary of what we learned from this data and what it means for the whole Fatum project.

So, the key takeaways from the Randonautica data are that the MMI effect has low reproducibility (1%), but if we split the sample by attractor points into smaller samples by z-score, most of them will retain this 1% increase, which means that this signal is most likely not a simple error. That is, there is a signal, but it is weak in terms of reproducibility.

This explains why we were unable to compare different algorithms in the Fatum laboratory tests. Since more advanced algorithms have high entropy consumption, they could only be tested in one stream, and therefore the series of experiments were small, 10-20 trials per method (I could no longer focus on the target, looking at it for two hours is already tiring). With such sample sizes, the effect with a reproducibility of 1% will be below the error level. I tried to get around this problem by using plants as a source of Psi, but they did not show positive results.
In Randonautica, there is no such problem, since the experiment is carried out in parallel by thousands of people and the samples are large, so the effect can be seen. However, the problem is that only the configuration with the Binary Word method of generating points can be tested in this way, since all other methods consume too much entropy and the PQ128MS would not be able to support parallel generation for many users at the same time.

Thus, to answer the main questions of the project, a separate setup would be required with an application like a psi-trainer, but in a 2D version and the ability to collect user statistics, as well as a large array of RNG, about 100 times more productive than the current one.

What questions need to be answered:
The most important question is about the characteristics of the signal. Having answered it, you can understand in which direction it is worth improving the amplification algorithms to increase the effect. At the moment, two characteristics were distinguished in the Fatum project: Signal complexity and its persistence. Let’s consider the main options:

Low complexity - high persistence.
This is a classic option, which is assumed in most experiments with MMI. The signal is random changes of 1 bit with some probability throughout the session. RW amplification and, accordingly, Fatum systems based on generating points by wandering work well for this assumption. In this approach, an array of points is generated one after another, where each point makes a full cycle of wandering, after which the next one begins to be generated. Here, the method of finding an attractor point acts as an additional amplifier on top of the RW-amplified coordinates. However, the entropy consumption can exceed 200 MB per attractor point.
Low complexity - low persistence.
Here we assume that the signal is localized in a short time frame. That is, the user does not influence the RNG constantly, but in short bursts, when Psi consumes a certain “charge” and then stops until it is replenished. The signal in this case will look like islands of large series of flipped bits, but all data outside these islands will only lead to the convergence of the deviation back to zero. For this case, two approaches have been developed: First, the generation of points by the RW method does not occur one after another, but in parallel, that is, each new bit goes to a new point until everyone has accumulated the required number of wanderings. To prevent the effect from being smeared across the entire map, the sequence of points is determined based on their Hilbert coordinates, so new bits are transmitted to the nearest points. The disadvantage is that such a calculation takes more time. Also for this case, a multi-channel bi-entropy method was developed. This is an amplification method in which the bit stream is represented as several channels shifted by the corresponding number of bits. In each channel, a segment of bits is taken and the entropy level in it is measured using the bi-entropy method. Then the channel with the highest indicator is selected and a segment from it is amplified. In this way, the amplifier can identify islands of influence.
High complexity - low persistence.
Here we assume that the signal can be not only 1-bit, but also represent a complex bit pattern (Binary word). However, such a signal should still obey the principles of probability theory, and therefore the binary word will occur rarely and mainly at the moment of the beginning of the impact. In such cases, the classical method of generating points of the Randonautica project will work, and the radius of the attractor will be extremely small. Also, in order to avoid smearing such coordinates along the axes, it is recommended to use the Hilbert coordinate system.
Moderate-High Complexity - High Persistence.
The case when patterns occur regularly, however, following statistical laws, they will be somewhat distorted and contain many errors. In this case, it will be impossible to recognize the exact pattern, but some part of it will be reproduced. For this case, it is recommended to use gray codes in combination with the Hilbert coordinate system to reduce the impact of the error on the position of the points and neutralize the msb problem. The bi-entropy method can also be used to validate such a signal.

Thus, if you compare the results of all the above approaches in an experiment where users try to hit the target with an attractor point, in theory you can see which of them will have a higher signal than the others and make an assumption about the characteristics of the signal, which will allow you to correctly select the mathematical tools for further improvement.