Aviator Predictors: The Manufactured Proof, Measured

What the videos show: a second phone that "predicts" 356.85x

Our social-media sweep of 26–29 October 2026 captured the same staging across several Facebook groups and Instagram reels devoted to the game Aviator. A first phone runs the game. A second phone, held beside it, displays the result of the round in progress before it ends: a crash at 356.85x in one video, at 42.87x in another, with a cash-out shown at 42.37x. The posts attribute the performance to an artificial-intelligence prediction engine connecting to the servers of Spribe, the game's developer, and they funnel systematically to Telegram and to download sites.⁠^[1]

The sellers' public pages, read on 30 October 2026, show the whole funnel. One of them, Seller C's page, promises: "you have full control over airplane movement via our app" — and states the price without detour: "APP COST 150 USD", with a download button that leads to Telegram.⁠^[2] Another, Seller B's page, runs the same scheme: a landing page, a "client satisfaction" counter, and the same button to Telegram.⁠^[3]

What the video never shows: the round before, or the round after. Nor how many rounds were filmed before the one that was published. The entire demonstration rests on the order in which the images appear — first the "prediction", then the crash. It is precisely that order which the rest of this study tests.

A round's result is sealed before the round begins

Aviator is a "provably fair" game, and its developer publishes the mechanism. Before each round, the server generates a secret seed and shows its hash to all players: the fingerprint is public, the value is not. The round's result is then computed from that seed and the seeds of the first three players in the round, combined through a hash function (SHA-512, per the developer's documentation).⁠^[4] Once the round ends, the server seed is revealed: anyone can recompute the result and check it against the fingerprint displayed before the round. We replayed that verification step by step in an earlier study, on a real round taken from an operator's public records: the full verification, replayed figure by figure.

The direct consequence for the videos above: while a round is being played, its result exists as a hash — a string of characters that cannot be reversed. Knowing it would require knowing the server seed before its reveal, which is exactly the value the mechanism is built to disclose only afterwards. The "connection to Spribe's servers" the sellers claim would therefore have to read information those servers present to no one, in usable form, before the round ends — not even to the operator itself. And every round is independent: the next round's hash contains nothing of this one.

This mechanism does not make the game advantageous — Aviator's published return to player is 97%, and the house edge is 3%.⁠^[4] It does something else: it fixes what can be known, and when. Before the round: a fingerprint. After the round: everything. In between, the exact place where the videos claim to stand does not exist.

Predicting 356.85x to two decimals: the probability, in figures

The crash-game distribution law is documented, and we have deposited it with a one-million-round corpus⁠^[5]: the probability that a round reaches at least multiplier x is r ÷ x, where r is the published return to player (0.97 for Aviator).⁠^[4] Applied to the two values staged in the videos, it gives this.

The third column deserves a slow reading. Announcing before the round that it will stop at 356.85x — not "above 300", not "in the 350s": 356.85, to two decimals — is an event that occurs on average once every 13.1 million rounds. The videos present this event as the routine demonstration of a 150-dollar product, repeatable on demand. If the announcement were real, a single pre-announced, timestamped, verifiable hit would multiply the evidence for the "this seller knows" thesis by roughly 13 million (section 6). None of the captured posts announces anything in advance: all of them show.

Filming is not predicting: the arithmetic of staging

The same distribution explains why the video is easy to produce in the other direction. Film a half-hour session — about 180 rounds: the probability that at least one round reaches 42.87x during the session is 98.4%.⁠^[5] For 356.85x it is 38.7% per half-hour; over a few evenings of ordinary play, a round beyond 356x eventually passes in front of the camera. What remains is to present the recording with the "predictor" phone in frame, display the result as if it had been announced, and publish. The production cost of the proof is one evening of play — which the seller plays anyway, with demonstration stakes of their choosing.

The cash-out shown at 42.37x against a crash at 42.87x belongs to the same arithmetic: for a real cash-out to land in the half-unit just before the displayed crash, it is enough to cash out often, at many levels — or to select, among the filmed rounds, the one where the gap looks telepathic. Roughly one round in 3,700 naturally presents a crash in the window immediately following a 42.37x cash-out.⁠^[5] Selecting is not predicting: it is waiting for chance to produce the picture, then captioning it.

That is the fear these videos sell, and it should be named: the feeling that other players hold information you do not, and that every round played without it is a round lost in advance. That information does not exist — section 2 shows it cannot exist while the round is running. What exists is a market that monetises the exit from that feeling: 150 dollars for the app, then the subscription to the group where the "next signals" arrive.

The predictor graded as a bettor: the Kelly criterion

One question remains, which the videos never ask: what happens if you follow? We grade predictors the way our research stack has graded forecast models since 2026: as bettors. The method, drawn from Beuoy's work on Bayesian evaluation through Kelly betting and extended by Ramdas and co-authors⁠^[6], fits in one sentence: treat the predictor as a bettor staking its own bankroll on its own calls, and read its performance in the growth of that bankroll. A model can game a scoreboard it chooses; it cannot game a bankroll.

The calculation is elementary, because the signals carry no information about the round (sections 2 and 3): following a predictor amounts to cashing out at fixed targets without knowledge — exactly the situation of our study of the real cost of low cash-outs. For each target, the no-tool success rate is 0.97 ÷ target, the expectation is −3% of the stake at every target, and the Kelly-optimal stake — the one that maximises bankroll growth — is zero.⁠^[6] At any positive stake, the expected growth of the bankroll is negative. A few landmarks, at 5% of bankroll staked per round:

The complete table, all targets and all stake levels, is attached to this article as a download. It reads directly: at the apparently "safest" target, 1.20x — the one where you "win" four rounds out of five — a follower staking 5% per round sees their median bankroll fall to 16% of its starting level within 1,000 rounds, and the probability of having lost at least half exceeds 90% at almost every target. Winning often is not winning: the size of the losses, not their frequency, sets the slope. Our register of predictor claims, kept since October 2026⁠^[7], shows the same thing from the side of the marketed percentages: one online predictor's advertised "80% accuracy", set against its own 1.20x target, sits below the 80.8% any player obtains with no tool at all, simply by always cashing out at the same level (the full register: our crash-game predictor study).

And what if the videos were taken at their word? A predictor genuinely able to announce the crash to two decimals would have unbounded Kelly growth: it would stake its entire bankroll on every round and cash out just beneath the known point. Such an asset is not sold for 150 dollars on a download page — it is not sold. The gap between what the claim would imply and the price at which it is monetised is, in itself, a measurement.

What real proof would require — and no seller provides

This is not moving the goalposts. The proof that would convince is simple to describe, and our research stack formalises it as a sequential test: predictions announced before the rounds, timestamped by a third party, over a sample declared in advance, whose success rate beats the no-tool level (0.97 ÷ target) by a margin that survives the sample size.⁠^[6] A single exact two-decimal prediction, announced before a 356.85x round, would by itself carry an evidence factor of roughly 13 million in the seller's favour — the standard is not unreachable in principle. It is simply never approached: across all posts captured between 26 and 29 October, the number of predictions announced before their round is zero.⁠^[1] Every "proof" is a recording presented afterwards.

Two honest limits on this study. We do not buy or install the tools being sold — our rule is to purchase nothing from the sellers we measure — so the grading covers their public claims and demonstrations, not their code. And since the Telegram funnels are closed, the claims quoted come from the sellers' public pages and our dated platform captures. These limits do not touch the central result: it depends on no code. It depends on a hash, a division and a logarithm.

Our analysis: the same film, two production costs with nothing in common

Everything above rests on the cited sources and on our calculations in sections 3 and 5. What follows is our analysis, and it mixes three things we distinguish at every step: the findings of the cited sources, credited to them and cited; our calculations on their published data, flagged as ours, with their margins; and our interpretations, labelled as such.

Predicting costs tens of thousands of times more rounds than manufacturing. Our calculation, on the cited distribution law⁠^[5] : reaching at least 356.85x takes 368 rounds on average; landing exactly in the two-decimal bucket [356.85, 356.86) takes about 13.1 million. The ratio of the two expectations is 35,686 for 356.85x, and 4,288 for 42.87x (44 rounds against about 190,000). The ratio is no accident: for any displayed value v it is exactly 100 × v + 1 — the larger the displayed value, the wider the gap between predicting it and manufacturing it.

One evening of filming succeeds almost for certain — and that is what empties the video of evidential value. Our calculation, at 10 seconds a round as in section 3: over a three-hour filming block, or 1,080 rounds, the probability that at least one round reaches 356.85x is 94.7%.⁠^[5] Our reading: a video produced after the fact over such an evening is almost as expected if the seller manufactures it as if they truly predicted it — the ratio between the two likelihoods is about 1.06 (1 ÷ 0.947), practically 1, so the video separates the two hypotheses almost not at all. What would separate them fits in one word: anteriority, timestamped by a third party (section 6).

A follower's median bankroll halves within a few hundred rounds. Our re-cut of the complete grading table attached to this article⁠^[6] : at 5% of bankroll per round, the number of rounds after which the median bankroll is halved — computed from the exact median bankrolls in the attached table — is about 383 at the 1.20x target, 252 at 2.00x and 119 at 5.00x. Our reading: the target that "wins" in 80.8% of rounds⁠^[5] therefore halves the median bankroll in under 400 rounds — the frequency of wins does not set the slope; the expectation of −3% per round does.

Before you pay for anything: the checklist

1. Demand the announcement before the round. A prediction that exists only in a video published after the round is not a prediction. Require timestamped, public announcements, made before the rounds they claim to call.
2. Divide every percentage by its target. For any advertised "accuracy" at a cash-out target x, the no-tool level is 0.97 ÷ x (at 1.20x: 80.8%). At or below that, the tool adds nothing a button does not already do.
3. Require the declared sample. How many predictions, over what period, with what complete success rate — failures included? A seller who cannot state their sample does not have a result; they have a selection.
4. Compare the price with the claim. If the tool really announced crashes to two decimals, its value would bear no relation to 150 dollars. When the claim is infinite and the price is finite, the price is the one telling the truth.
5. Look at where the funnel leads. An anonymous page, payment or download via Telegram, no identifiable publisher, no recourse: if the seller disappears tomorrow, who do you write to?
6. Do not install apps from outside the official stores. Files distributed through these funnels often request broad permissions on the phone. The risk runs beyond the game: credentials, wallets, personal data.
7. Verify a completed round yourself. The game's fairness panel shows, after each round, the seeds that let you recompute it. It is free, it takes two minutes — and it is the only "prediction" in the system that can be checked: the one about the past.

And if you have already paid: stop following the signals before you stop anything else — every followed round costs 3% of the stake in expectation, app or no app. Keep screenshots of the promises and the payments; they document what you were sold.

Method and data: redo every calculation

Artefacts. Posts captured by our platform sweep between 26 and 29 October 2026 (public Facebook groups and Instagram reels promoting Seller A's and Seller B's products; staged values recorded: 356.85x and 42.87x, cash-out shown at 42.37x). Seller pages read on 30 October 2026: Seller C's (displayed price 150 USD, quotation reproduced in section 1)⁠^[2] and Seller B's⁠^[3] — public pages, cited as documents, with no purchase and no sign-up; the sellers' identities and addresses are recorded in our register — full list on request.

Mechanism. The fairness scheme published by Aviator's developer (server seed committed by hash before the round, player seeds, reveal and verification after the round); our complete replay of a verification is published in Provably Fair, Verified Step by Step. Aviator's published return to player: 97%.⁠^[4]

Probabilities and grading. The crash distribution law P(≥ x) = r ÷ x, that of our deposited one-million-round corpus (Study 01, DOI 10.5281/zenodo.23597144, documented model at 98% return; this article's calculations use Aviator's published 97%). Kelly grading under the method adopted in our research register (Beuoy; Ramdas et al.)⁠^[6]: growth g(f) = p·ln(1 + f(x−1)) + (1−p)·ln(1−f), median bankrolls from the exact binomial law. The complete table — six targets, four stake levels, median bankrolls at 100, 500 and 1,000 rounds and half-loss probabilities — is attached as a download: kelly-grading-aviator-predictors-2026-10-30.csv. No purchases, no sign-ups, no personal data processed.

Our publication rule. Sellers' claims are quoted word for word and measured; sellers are designated by letters (Seller A, Seller B, Seller C), their identities and pages recorded in our register — full list on request — and we attach no further label to any product or seller. The figures above are the whole of our conclusion — anyone can redo them from the cited sources.

Frequently asked questions

References

1. Chicken Road Productions platform sweep, 26–29 October 2026 — public posts promoting Seller A's and Seller B's products (Facebook groups, Instagram reels); staged values recorded: 356.85x, 42.87x, cash-out 42.37x. ↑
2. Seller C — public page read on 30 October 2026 (quotation and displayed price: 150 USD). ↑
3. Seller B — public page read on 30 October 2026. ↑
4. Spribe — published documentation of Aviator's provably fair mechanism: server seed committed by hash, player seeds, reveal and verification after the round; published return to player: 97%. spribe.co ↑
5. Dagor, A. — Crash-game bust multipliers: a one-million-round simulation, Zenodo, DOI 10.5281/zenodo.23597144 (distribution law and deposited corpus). ↑
6. Beuoy, M. — Kelly Betting as Bayesian Model Evaluation; Ramdas, A. et al. — work on optimal betting growth rates and power-one sequential tests (grading method adopted into the research register, 30 October 2026). ↑
7. Chicken Road Productions — predictor-claim register (kept since 27 October 2026): claims quoted word for word, each set against the no-tool level for its target. ↑