Is gravity the thermodynamics of quantum entanglement?

A computational research program. Every claim below carries its evidential status; every number was produced by a public script you can run in minutes; ideas we refuted ourselves are listed, not hidden.

Not peer-reviewed · human–AI collaboration · all code public · last updated June 2026
Paper (PDF) · Interactive demo · Full index of results · Prior-art audit · Earlier work (archive)

The idea

Quantum entanglement behaves like a finite resource: a system that is strongly entangled with one partner has less entanglement available for others (monogamy). The vacuum of quantum field theory is saturated with entanglement — the amount between a region and its surroundings grows with the region's boundary area, the same way black-hole entropy does.

The hypothesis under test: spacetime geometry is the bookkeeping of this resource. A region's area measures its entanglement budget; when matter systems entangle with each other, that entanglement is taken from the vacuum between them; and Einstein's equations express the requirement that the books balance at every point. Distance, force, dark energy and the arrow of time should then all be readable as properties of how vacuum entanglement is distributed, spent, and lost.

This is not an original premise — it builds on Jacobson (1995, 2015), Van Raamsdonk (2010), Verlinde (2016) and the holographic literature. What this program adds is (a) monogamy as the specific conservation principle, which among other things forces the attractive sign of gravity, and (b) an insistence on testing each step numerically, on lattice models where everything is exactly computable, before believing it.

Four postulates

Entanglement is the only fundamental quantity. No spacetime, no fields at the foundation — a quantum state and the entanglement between its parts. axiom
A region's entanglement budget is its boundary area. Measured on lattice vacua (area law). One amendment forced by measurement: the budget is held mostly in collective (many-party) form — pairwise entanglement accounts for only 11–40% of it. measured · amended
Matter–matter entanglement is debited from the vacuum. Passed its designed falsification test: importing entanglement into two sites of a lattice vacuum drains their vacuum entanglement monotonically to exactly zero, and the totals balance to four decimal places. tested
Decoherence is the cost of irreversible records. Originally: "every entanglement transaction pays a universal fee." Our own exact simulation falsified that — in flat space the adiabatic cost is exactly zero — and the postulate was amended: the cost is what gets radiated (through horizons, or by sudden changes). falsified amended

Three results, with their figures

1. The vacuum measures distance in resistance, not in travel time

On a lattice with spatially varying stiffness, there are two natural distances between points: the resistance of the path (Σ 1/c) and the wave travel time (Σ 1/√c). They disagree inside any "gravitational lens." We pre-registered the prediction that vacuum correlations follow travel time — and were wrong in all eight test configurations: entanglement follows the resistance metric, the same quantity that sets a topological particle's mass in the companion graph model. One geometric object under both sides of the field equation.

2. A single metric is the zero-noise configuration

That result raises a problem: such substrates generically carry two metrics (static/resistive vs. causal), while our universe demonstrably has one — gravitational waves and light from the same merger arrived within one part in 10¹⁵ of each other's speed (GW170817). We found the welding condition: the two metrics coincide exactly when mass density tracks stiffness, m·c = const, which is uniform wave impedance — the textbook condition for zero reflection. Measured: the same sharp obstacle reflects 19% of an incident wave in the generic substrate and 0.5% in the welded one.

Reflected fraction versus mass profile exponent; minimum at the uniform-impedance point s=1
Reflection off an identical obstacle, as the substrate's mass profile is varied. m = c−s; the minimum sits at s = 1, the uniform-impedance ("one metric") point — 39× quieter than the generic case. Since every reflection is an emitted record (decoherence), a bimetric vacuum continuously radiates and decoheres what crosses it. Conjecture, clearly labeled as such: our sky has one metric because the alternatives are dissipatively unstable. Details & script.

3. In a massless vacuum, the position of a mass is public information

Place a point mass in a lattice vacuum and ask: which distant regions could tell, from their local state alone, where it is? In the massless vacuum, every one of 28 disjoint fragments can — the position is redundantly recorded everywhere, which is the information-theoretic signature of an objective, classical fact (Zurek's "quantum Darwinism", here found in the static vacuum itself). Give the field a mass, and the record dies at the correlation length: geometry becomes private.

Per-fragment discrimination of the mass position vs distance, massless vs gapped vacuum
How well each vacuum fragment knows where the mass is. Massless field: power-law decay — every fragment on the lattice can discriminate the position (28/28 above threshold). Gapped field: machine zero beyond the correlation length (1/28). Our universe gravitates through a massless field; this may be why everyone agrees on where things are. Details & script.

4. A scaling law found by accident — and verified by prediction

Part of the vacuum's correlation between two regions can never be extracted as usable entanglement (it is "locked": the mutual information persists while the distillable part is exactly zero). Hunting for where the locked fraction equals the observed dark-energy fraction (0.68), we found not a point but a law: the 0.68 contour sits at a fixed absolute gap of 1.83 lattice units between the regions — independent of their size. The law, fitted at four sizes, predicted the fifth (κ*=2.229) before we computed it (2.228).

kappa* versus region radius with the constant-gap law and the out-of-sample point
The constant-gap law. κ* = separation/radius at which 68% of the inter-region correlation is locked. The red square is an out-of-sample test: predicted from the smaller sizes before being measured. Whether this has anything to do with dark energy is an open question, flagged as such; the lattice law stands on its own. Details & script.

Everything else, in one line each

What we got wrong, in public

Four of our own pre-registered predictions were refuted by our own runs. We keep them visible because they are the evidence that the methodology can say no to us:

Open problems

Reproduce everything

Every figure and number on this page comes from a short Python script (plain numpy; minutes on a laptop): n2_withdrawal.py, n5_reserve.py, n10_broadcast.py, n11_two_metrics.py, n12_weld.py, d1_fstar.py, and the data files alongside them. The standalone paper on the locked-fraction results: PDF. An out-of-sample prediction (result 4) is timestamped in the public git history before its measurement.

Method, honestly

This program is a human–AI collaboration: the AI generates and audits derivations and code at high speed; the discipline that makes this acceptable is mechanical, not rhetorical — falsification criteria written before each run, two adversarial prior-art audits (≈100 search agents each) whose corrections are posted on the affected pages, and refutations published with the same prominence as successes. None of this substitutes for independent human review, which is the program's explicit next step and the reason the paper above exists.

Closest prior work, credited where each result builds on it: Jacobson; Faulkner–Guica–Hartman–Myers–Van Raamsdonk; Verlinde; Casini–Huerta; Eisler–Peschel; Klco–Savage; Adesso–Illuminati; Zurek; Riedel–Zurek; Danielson–Satishchandran–Wald; Wilson-Gerow–Dugad–Chen; Wheeler; Sakharov; Kleinert; Friedman–Sorkin. Full audit: prior-art page.