Gravity that pushes, in one run in four hundred
A 2026 essay says a mass held in two places at once, with the results sorted afterwards, can push a probe away. We ran the protocol step by step. In the rare sorted subset the probe’s average momentum does shift away from the source, by close to the amount the paper predicts; the check written down first, that it matches the paper’s weak value at every sorting angle, missed narrowly. Over all runs together the pull is the ordinary one.
The verdict
Real effect, different reason
In a step-by-step quantum simulation, gravity from a mass in superposition pushes the probe away in the rare sorted subset, as the paper says, while the average over all runs stays the ordinary pull. The preregistered check that the push is the weak value missed narrowly: 2.13 % off at the sharpest sorting angle, against a 2 % bar. It is a conditional shift of an average, not a force anyone can use.
With the near branch’s kick at a thousandth of the probe’s momentum spread, the selected runs move away from the source by −3.54×10−3 of that spread. The weak value predicts −3.57×10−3.
Summed over every measurement outcome, the kick equals the ordinary attraction to within 5×10−13. Sorting cannot change the average, so nothing can be signalled or harvested.
A source that is simply in one place or the other (dephased) gives a kick between 2.1 and 4.3×10−4 in every subset: attractive, never negative.
Once the kick is no longer small next to the probe’s spread, the sorted subset stops being pushed. At a tenth of the spread it is already pulled.
The claim, as presented. The essay presents a scheme in which a mass held in superposition, with the results sorted afterwards, pushes a probe away in the sorted subset, while gravity stays attractive on average.
Popular Mechanics (Vlatko Vedral), “We Can Make an Antigravity Machine—And Unlock a 'Mind-Blowing' Future”, June 2026 · the episode
The registry’s note on sources. Read via the author's own repost and syndicated copies; the popularmechanics.com page itself was not opened. The paper behind it is Saldanha, Marletto and Vedral, arXiv:2602.12266.
The paper behind the essay is Saldanha, Marletto and Vedral, arXiv:2602.12266. The test is of the physics the essay describes; the “machine” and “future” framing is outside it.
Preregistered in HI-10d-postselection/PREREG.md (commit aa39064) before any code was written. Self-graded, with no outside review. The lab notebook is not public yet.
The setup
A source mass sits in a superposition of a near position and a far one. A probe particle, a wave packet, feels the softened 1/r pull of each branch while the simulation steps the Schrödinger equation forward. Afterwards the source is measured along a direction close to orthogonal to how it was prepared, and only the runs with one outcome are kept.
In those kept runs the probe’s average kick is a weak value: a weighted mix of the two branch kicks whose weights can fall outside 0 to 1. When they do, the average can point away from both positions of the source.
Below, ε is the near branch’s kick divided by the probe’s momentum spread, and δα is how far the sorting direction sits from exactly orthogonal. Kicks are in units of the momentum spread; negative means pushed away.
The checks
Every check below was written down before the code, except one replacement for a planted fault, noted under the table. Each planted fault had to turn its check red once.
| Check | Result | Numbers |
|---|---|---|
| The claim: sorted subset at ε = 10−3, δα = −0.05 | pass | −3.54×10−3 vs weak value −3.57×10−3; probability 0.25 % |
| All outcomes give the ordinary pull | pass | within 4.7×10−13 relative, every case |
| … with a planted signalling kick | red, as it should be | off by 0.39 |
| Probability is conserved through the hold | pass | 3.5×10−13 |
| … with a planted leak | red, as it should be | 2.0×10−3 |
| Dephased source never pushes | pass | every subset between +2.1 and +4.3×10−4 |
| 4 million sampled runs, ε = 0.02 | pass | −6.19×10−3 ± 8.4×10−4 vs exact −5.85×10−3 |
| … with a sampler that forgets the sorting | red, as it should be | reads +5.7×10−3 (attractive), z = 13.7 |
| Time step and grid halved | pass | changes below 1.3×10−9 |
| Matches the weak value at every δα within 2 % | missed | 2.13 % at δα = −0.02; 0.8 % or less away from the sharpest angles |
The misses. One check of ten missed: the match to the weak value is 2.13 % off at the sharpest sorting angle, where higher-order corrections are expected, against a 2 % limit. One planted fault (a sampler using the amplitude’s size) could not turn its check red by construction; it was replaced, after the fact and on the record, by the sampler that forgets the sorting, which did. The preregistered verdict rule did not cover a lone miss of the weak-value match. On review the verdict was first set to the claim holds (2026-10-12). On 2026-10-18, at Ben’s instruction, the lab’s verdict engine replaced that call: the push is there, and the check that it is the weak value is the one that missed, which the engine reads as real effect, different reason.
Turning gravity up
| ε | δα | Simulated kick | Weak value |
|---|---|---|---|
| 0.001 | −0.05 | −7.1×10−3 | −7.1×10−3 |
| 0.01 | −0.05 | −6.2×10−2 | −7.1×10−2 |
| 0.1 | −0.05 | +3.2×10−2 (pulled) | −0.72 |
| 0.1 | −0.3 | −3.4×10−2 | −6.2×10−2 |
| 0.5 | −0.05 | +0.38 (pulled) | −3.6 |
| 0.5 | −0.3 | +0.24 (pulled) | −0.31 |
The push is a weak-coupling effect, as predicted. The paper’s own Example 2, run through the same simulation, gives −1.96×10−3 at a probability of 8.0×10−4; the paper gives about −2×10−3 at 8×10−4.
The shift is hundreds of times smaller than the probe’s own momentum spread, and it lives in a fraction of a percent of the runs. No single probe can be seen to fly away, and there is no net force to build anything on.
- Decoherence from the lab walls and the environment, and a full 3-D geometry.
- The essay’s “machine” and “future” framing.
- The popularmechanics.com page itself; the essay was read through the author’s own repost.
Read: Saldanha, Marletto & Vedral, arXiv:2602.12266 · the essay, via the author’s repost. Prior art: Aharonov, Albert & Vaidman, Phys. Rev. Lett. 60, 1351 (1988), on weak values. No novelty claimed.