The proton's baryon number may not be carried by its quarks — the gluon-junction hypothesis regains ground

What distinguishes matter from antimatter has a name: baryon number. It is taught as a property of quarks — one third each, three quarks, one proton. The STAR collaboration, at Brookhaven's RHIC collider, is publishing a measurement that runs against this teaching: in nuclear collisions, baryon number travels far better than the quarks that are supposed to carry it. The explanation that fits best is an old hypothesis that has remained in the margins — a Y-shaped gluon topology, the "baryon junction".

Source: science.org

In plain terms

A proton is three quarks held together by glue: the gluons. It has always been assumed that the proton's "baryon identity" — what makes it count as one unit of matter — is inscribed in the three quarks, one third each. The experiment tells another story.

Picture two trucks colliding head-on. After the crash, you look at the point of impact for what belonged to the trucks. The quarks are the heavy passengers: launched at full speed, they pass through and continue on their way, very hard to stop in the middle. Yet in the middle one finds far more "matter" than those heavy passengers could ever deposit. As if the truck's identity were in fact carried by something light and easy to tear off — a knot in the glue, not the passengers.

That is what the measurement suggests. A caveat: this is a bundle of converging clues that makes the quark picture improbable, not a direct detection of the junction. The generator that simulates the full evolution of the isobaric system, UrQMD, implements no junction at all and stays well below the data — which says where the problem comes from, without thereby supplying the model that would account for it.

Technical explanation

1. What baryon number is, and why nobody knows where it is stored. Baryon number $B$B is a conserved charge: a proton is worth $B = 1$B=1, an antiproton $B = -1$B=−1. Its conservation is what, to our knowledge, keeps the proton from decaying and hence keeps ordinary matter from evaporating. The usual assignment — $B = 1/3$B=1/3 per valence quark — is an accounting convention that works, not a measurement. Nowhere does quantum chromodynamics require this global charge to be localised on the quarks. The alternative proposed in the 1970s moves the charge onto the "glue": within a baryon, the three colour flux tubes issuing from the three quarks must meet somewhere for the whole to be colour-white. That Y-shaped meeting point is a non-perturbative gluon configuration, electrically neutral ($Q = 0$Q=0) and carrying $B = 1$B=1 all on its own.

2. The experimental lever: what can be stopped in the middle. A collision of ultra-relativistic nuclei defines two rapidity extremes, those of the beams ($y = \pm Y_{\text{beam}}$y=±Ybeam​), and a middle ($y \approx 0$y≈0). The whole question is how much baryon number manages to be slowed down all the way to the middle. Here the two hypotheses do not say the same thing, and this is where the actual mechanism plays out: valence quarks carry a large fraction of the nucleon's momentum, which makes them, in the paper's terms, "difficult to transport from $Y_{\text{beam}}$Ybeam​ to $y \sim 0$y∼0"; the junction, for its part, is made of low-momentum gluons, hence easy to detach and to deposit at the centre. In other words: if baryon number arrives in the middle in abundance, then it was not attached to whatever moves fast.

3. The isobar trick — cancelling the geometry to keep only the charge. Ruthenium 96 and zirconium 96 have the same mass number ($A = 96$A=96) but not the same number of protons ($Z = 44$Z=44 versus $Z = 40$Z=40, hence $\Delta Z = 4$ΔZ=4). Two nuclei of the same size: collision geometry, volumes and detection efficiencies are practically identical and cancel in a difference. What remains is a known electric charge gap. The reasoning is then direct: if the quarks carry both the charge and the baryon number, the two travel together and one expects $\langle B\rangle/\Delta Q \approx A/\Delta Z$⟨B⟩/ΔQ≈A/ΔZ — which, once normalised by $\Delta Z/A$ΔZ/A, gives $\langle B\rangle/\Delta Q \times \Delta Z/A = 1$⟨B⟩/ΔQ×ΔZ/A=1. The authors take care to qualify this 1 as a naive expectation, valid "if baryon number is carried by the valence quarks and all other effects are ignored". If the charge is carried by the quarks but the baryon number by a neutral junction, the latter arrives in excess and the ratio exceeds 1. Measurement: $1{,}84 \pm 0{,}02\ (\text{stat.}) \pm 0{,}09\ (\text{syst.}) \pm 0{,}16\ (\text{feed-down})$1,84±0,02 (stat.)±0,09 (syst.)±0,16 (feed-down) in the most central collisions, with a monotonic decrease from central (0–10 %) to peripheral (70–80 %). This decrease is not attributed by the authors to baryon transport but to a nuclear-structure effect, which the TRENTO model reproduces: zirconium's thicker neutron skin compared with ruthenium's progressively reduces the fraction of participating protons in Zr+Zr collisions as one moves towards the peripheral. That is also why the peripheral points must not be confronted directly with the generators — the authors write it explicitly.

4. What the double ratios actually measure. The raw quantities are
   $Q = (N_{\pi^+} + N_{K^+} + N_p) - (N_{\pi^-} + N_{K^-} + N_{\bar p})$Q=(Nπ+​+NK+​+Np​)−(Nπ−​+NK−​+Npˉ​​)
   $B = (N_p + N_n) - (N_{\bar p} + N_{\bar n})$B=(Np​+Nn​)−(Npˉ​​+Nnˉ​)
   and the charge gap between the two systems is written as a sum of double ratios,
   $\Delta Q \approx N_\pi (R2_\pi - 1) + N_K (R2_K - 1) + N_p (R2_p - 1),$ΔQ≈Nπ​(R2π​−1)+NK​(R2K​−1)+Np​(R2p​−1),
   with, for example, $R2_\pi = (N_{\pi^+}/N_{\pi^-})_{\mathrm{Ru+Ru}} / (N_{\pi^+}/N_{\pi^-})_{\mathrm{Zr+Zr}}$R2π​=(Nπ+​/Nπ−​)Ru+Ru​/(Nπ+​/Nπ−​)Zr+Zr​. The methodological benefit lies in the way the data were taken, and the authors put it thus: the ruthenium and zirconium beams were recorded "under nearly identical accelerator and detector conditions and analysed with the same procedure". A ratio of ratios built on two datasets taken under such conditions makes most of what is common to both disappear. That is what allows a statistical uncertainty at the percent level on so tenuous a quantity.

5. The weak link, named by the authors: the neutrons. $B$B includes the neutrons, which STAR does not measure directly. Their yield is estimated, relying on the deuteron/antideuteron ratio, through
   $N_n^{\text{pri}} = N_{\bar p}^{\text{pri}}\sqrt{N_d/N_{\bar d}}.$Nnpri​=Npˉ​pri​Nd​/Ndˉ​​.
   To this is added hyperon decay, estimated from their yields and known branching ratios, which contributes about 30 % of the total neutron yield. That is exactly why the result carries a third error bar, separate and dominant: $\pm 0{,}16$±0,16 of "feed-down", i.e. more than the statistical and systematic uncertainties combined.

6. The photonuclear test, cleaner because the photon brings nothing. In $\gamma+\mathrm{Au}$γ+Au events, one of the two nuclei interacts only through its electromagnetic field: the photon carries no baryon number. Any excess of baryons therefore comes from one side only, which removes contamination from the partner. One then fits the net-proton rapidity density with an exponential in distance from the beam,
   $f(y) \propto \exp(-\alpha_B \Delta y), \qquad \Delta y = Y_{\text{beam}} - y,$f(y)∝exp(−αB​Δy),Δy=Ybeam​−y,
   with $Y_{\text{beam}} = 4{,}06$Ybeam​=4,06. The parameter $\alpha_B$αB​ says how fast baryon-number transport runs out as one moves away from the beam: the smaller it is, the further the charge travels. Regge theory applied to the junction predicts $0{,}42 < \alpha_B < 1$0,42<αB​<1; the measurement gives $\alpha_B = 1{,}04 \pm 0{,}22$αB​=1,04±0,22. This figure must be read for what it is: the central value sits slightly above the predicted upper bound, and consistency with the interval rests solely on the $\pm 0{,}22$±0,22 uncertainty.

7. The control that makes the measurement defensible. A slope can always be an artefact of event selection. The authors therefore measure the same slope for antiprotons: $\alpha_{\bar B} = 0{,}02 \pm 0{,}05$αBˉ​=0,02±0,05, that is, flat. Antiprotons are created in pairs in the collision and do not transport the beams' baryon number; if they show no rapidity dependence while net protons do, then the observed dependence is indeed a property of baryon transport and not of the apparatus. It is an experimental control, and not a theoretical argument, that makes the artefact improbable here. Finally, in Au+Au collisions the mean slope is $0{,}64 \pm 0{,}05$0,64±0,05 and does not depend on centrality — consistent with the photonuclear value to within 1,7 σ — whereas the quark picture would lead one to expect a variation with the number of multiple scatterings.

Why It Worked

The experiment works because it turns a question of location — where a charge is stored — into a question of transport, which can be measured. The quantitative contrast is sharp: where the valence-quark picture naively leads to $\langle B\rangle/\Delta Q \times \Delta Z/A = 1$⟨B⟩/ΔQ×ΔZ/A=1, the data show 1,84, that is, nearly double, with a total uncertainty of order 0,2 when the three contributions are added. The generator that simulates the full evolution of the isobaric system, UrQMD, implements no junction and predicts a value of 0,5 to 0,7 — "significantly below" the experimental data, the authors write.

One must, however, state clearly the gap between what the measurement establishes and what one might be tempted to conclude from it. First, the comparison with the generators calls for a reading precaution. The authors do place on the same figure as their measured points five model calculations — UrQMD, TRENTO, PYTHIA 8.3, PYTHIA 8.3 with colour reconnection in mode 2, and HERWIG 7.2 — the fourth of which is precisely the one that dynamically produces junctions. But the PYTHIA and HERWIG values — about 0,5 to 0,6 for the default settings, $0{,}99 \pm 0{,}03$0,99±0,03 for the junction mode — are computed for the limiting case of proton-proton collisions at 200 GeV, and not for the isobaric system itself. Yet the only isobaric points whose number of participating nucleons approaches this limit are the peripheral points, and it is precisely about them that the authors give warning: because of the neutron-skin effect, they "should not be directly compared with the HERWIG and PYTHIA predictions". There is therefore no clean quantitative confrontation between a junction generator and the 1,84 of central collisions: the comparison is displayed, but it is not usable as a subtraction.

Second, the junction does not escape criticism on the other observable either: both PYTHIA 8.3 settings, including the one that produces junctions, significantly overestimate the slope measured in $\gamma+\mathrm{Au}$γ+Au — "perhaps because of the absence of certain physical ingredients in the model", the authors note. The junction wins a two-way comparison against the quark picture; it does not thereby achieve a fit.

Third, the authors announce no statistical significance in σ for rejecting the quark picture: their conclusion is that these results, supported by the Au+Au data previously published by STAR at several energies, "disfavor the valence-quark picture". To disfavor is not to exclude, and a model preference is not a detection. Finally, the object itself escapes calculation: a junction is non-perturbative, hence impossible to treat with the usual tools of perturbative QCD, which leaves the theoretical prediction as a broad interval rather than a value. Added to this are acknowledged sample limitations: the selected photonuclear events are biased towards high multiplicities, and the authors indicate that $\alpha_B$αB​ could depend on collision kinematics without the current measurements allowing the matter to be settled.

Causal Chain

Colour confinement: a baryon must be white → the three flux tubes issuing from the quarks must meet at a point → this Y-shaped topology is predicted as an object in its own right by Artru (1975) then Rossi and Veneziano (1977) → the junction, made of low-momentum gluons, separates from the fast quarks during a collision → it is slowed down to mid-rapidity while the valence quarks continue on their way → baryon number arrives at the centre in excess relative to electric charge → measured on Ru/Zr isobars, this ratio is 1,84 where the naive valence-quark expectation gives 1 → measured photonuclearly, where the photon brings no baryon, the transport slope is consistent with the window foreseen for the junction, its central value slightly exceeding the upper bound → the valence-quark picture is disfavored by three independent observables → but the junction generators are computed only for the proton-proton limiting case, and they overestimate the photonuclear slope → the electron-ion collider under construction at Brookhaven will inherit the question.

Anecdote

The four billion ruthenium and zirconium collisions — two billion per species — were not taken for that physics. In 2018, RHIC alternated these two beams — sometimes several times on the same day — to hunt for an entirely different phenomenon: the chiral magnetic effect. The reasoning was that ruthenium, with four more protons, generates a more intense magnetic field than zirconium and should therefore produce a more marked charge separation. To guard against their own desire to find, the STAR physicists set up a three-stage blind analysis, five independent groups working without knowing which species the data came from. The verdict, delivered in 2021, was negative: no chiral-magnetic-effect signal matching expectations. It is this dataset, assembled for one question and having failed to give the hoped-for answer, that today serves to weigh a half-century-old hypothesis about the nature of the proton.

Legacy and Current Data

What is at stake goes beyond the proton's internal bookkeeping. If baryon number is a topological property of the gluon field rather than a label carried by particles, then the scenarios that seek to explain why the universe contains matter and almost no antimatter — baryogenesis — are working on an object whose nature is not the one that was assumed. At this stage, one must be clear about what is not established: this measurement says nothing about baryon-number violation, provides no rate, and has no quantified implication for proton decay. What has been secured is more modest and more solid: a preferred transport mechanism, measured on two independent collision systems, plus a discrepancy with junction-free generators that the authors add to the file.

The researcher's view — open questions

This section is interpretation and not the paper's results. Three experiments would settle matters. The energy scan first: if $\alpha_B$αB​ really is a property of the junction and not an accident of kinematics, it should remain stable when $\sqrt{s_{NN}}$sNN​​ is changed — the authors themselves point out that their data do not allow this to be checked. The portability test next: the same measurement on other isobaric pairs, with a different $\Delta Z$ΔZ, would say whether 1,84 is a constant of the mechanism or depends on the pair chosen. The theoretical constraint finally: the junction being non-perturbative, only lattice QCD can in principle supply a quantified prediction of its contribution to baryon number. As long as that calculation is missing — and as long as the junction generators are computed only for the proton-proton limiting case, outside the regime where the 1,84 is measured — one will not be able to exclude that an ordinary, unmodelled transport effect explains part of the discrepancy. A methodological remark to close, which is mine and not the authors': displaying the feed-down uncertainty as a third, separate error bar, rather than folding it into the systematic, is a choice of transparency — it makes visible that the dominant limitation of this measurement is the estimation of the neutrons.

Sources

References verified during the fact-checking audit of 14 August 2026 — full text of the version deposited on arXiv consulted (Methods, Results and figure captions), publication metadata verified separately. The generator predictions cited here are placed back in the collision system for which the authors computed them.

• STAR Collaboration, "Tracking the baryon number with nuclear collisions", Science 393, no. 6812, pp. 727–731, 13 August 2026. DOI: 10.1126/science.ads5962 — peer-reviewed article.
• Full version deposited: arXiv:2408.15441, submitted 27 August 2024. arxiv.org/abs/2408.15441 — it is from this version that the detailed numerical values, the equations and the limitations cited here are drawn.

Background references

These references support the contextual reminders and do not come from the study under discussion.

• X. Artru, "Classical string phenomenology" (1975) and G. C. Rossi, G. Veneziano (1977) — original proposals of the baryon junction, cited as such in the study's references.
• STAR Collaboration, "Methods for a blind analysis of isobar data collected by the STAR collaboration", arXiv:1911.00596 — blind-analysis protocol of the isobar programme, for the anecdote on the origin of the 2018 data.
• Brookhaven National Laboratory, press release "Results from search for 'Chiral Magnetic Effect' at RHIC" (2021) — negative result of the search for which the isobaric collisions had been scheduled.