Wigner's Friend Experiments

Speculative If an extended Wigner’s friend experiment is ever run with a friend whose standing as an observer is not itself the disputed premise, then violating a local-friendliness inequality forces one of three assumptions about the world to be false: that every observed event exists absolutely rather than relative to someone; that a free choice here cannot change a probability there; or that settings can be chosen freely at all. The condition is the whole subject, and every experiment to date fails it. Section 5 is about the gap.

Established The magnitude is best stated as one metaphysical assumption eliminated that Bell’s theorem leaves standing. Bell’s theorem assumes locality, free choice and outcome independence, and its violation forces a choice among the three. Frontier Bong and seven co-authors argue that local friendliness drops the third and still yields inequalities quantum theory violates, so that the theorem “places strictly stronger constraints on physical reality than Bell’s theorem”. That is the prize: a sixty-year-old choice between locality and determinism replaced by a choice in which determinism is no longer one of the options. Speculative Whether the assumption set is genuinely weaker is contested in print, and section 15 states the objection.

Established The distance to the prize has a number, because the field invented a unit for it. Zeng, Labib and Russo define a branch factor — the gap between how hard it is to interfere two branches of a superposition and how easy it is to tell them apart — as an operational measure of how observer-like a friend is. On that scale the photonic friends used in every published test score 0.0; their own quantum-computer run reaches 16.0, with a seventeen-qubit friend on a trapped-ion machine; and they estimate that a human-level artificial agent would need of order 10¹⁹ logical qubits and a logical depth of 10¹⁴ operations. Frontier The upside is eighteen orders of magnitude away on the only scale anyone has proposed for measuring the distance.

Handwave Nobody gets a device out of this, and that should be said in the first section rather than the twelfth. There is no product and no customer; the beneficiary is the community choosing among interpretations of quantum theory, and the benefit is the elimination of options rather than the confirmation of one. Speculative The condition on all of it: the experiment needs a friend that can be coherently reversed after it has observed something, and the one concrete proposal in the literature — a reversible artificial intelligence on a fault-tolerant quantum computer — is costed by one of its own authors at more than 500 years of machine time for one second of thought.

Established Wigner’s 1961 puzzle is a disagreement about a description, not about a number. A friend inside a sealed laboratory measures a system and gets an outcome; Wigner, outside, describes the whole laboratory as evolving unitarily and assigns it an entangled state in which no outcome has occurred. Both descriptions are licensed by quantum theory as ordinarily used, and for the original scenario they make identical predictions for anything either party can check. That is why the original thought experiment settles nothing, and the IBM Quantum tutorial that compiles it into a five-qubit circuit says so plainly: the scenario “leads to the same measurement outcomes regardless of when or if an irreversible collapse happens” (vendor tutorial).

Established The modern subject begins when the scenario is extended so that two sealed laboratories are entangled and the disagreement becomes a measurable correlation. Brukner set out the two-friend version in 2018 and derived a no-go theorem for observer-independent facts; Frauchiger and Renner published a different extension the same year; Bong and co-authors turned Brukner’s scenario into a family of theory-independent inequalities in 2020. Frontier All three are no-go theorems about assumptions. None of them refutes an interpretation directly, and reading them as if they did is the commonest error made about this field.

Established Frauchiger and Renner kill a conjunction of three named assumptions and nothing else. Their Nature Communications paper proves that the following cannot all hold: (Q), that an agent using quantum theory may apply the Born rule to any system around her and trust a prediction made with certainty; (C), consistency, “in the sense that the different agents’ predictions are not contradictory”; and (S), that “from the viewpoint of an agent who carries out a particular measurement, this measurement has one single outcome”. The contradiction arises in a fraction 1/12 of runs. Frontier The paper’s own table assigns the cost: collapse models and Bohmian mechanics give up (Q); Copenhagen variants, consistent histories and QBism give up (C); many-worlds gives up (S). Speculative Every one of those positions was already committed to the sacrifice the table assigns it, which is why the theorem eliminated no working interpretation and started an argument instead.

Established Local friendliness kills a different conjunction, and this one is theory-independent. Bong and colleagues assume: Absoluteness of Observed Events — “an observed event is a real single event, and not relative to anything or anyone”; No-Superdeterminism — any set of events on a space-like hypersurface is uncorrelated with freely chosen actions after it; and Locality — “the probability of an observable event e is unchanged by conditioning on a space-like-separated free choice z”. From those three alone, with no assumption that quantum theory is correct, they derive inequalities on observable correlations. Frontier Quantum theory predicts violations, so if the friend is an observer, one of the three is false.

Established The premise that carries the weight is stated in the abstract and is conditional: “if quantum evolution is controllable on the scale of an observer”. Frontier The theorem is an implication, not a refutation, and its antecedent has never been satisfied. Speculative Bong and colleagues say so themselves, and leave the escape route open in print: “an open possibility is that the LF assumptions are valid, but that nature forever forbids the observation of violation of LF inequalities with observers”. Handwave A programme whose central theorem names, in its own text, a way the world could be in which the theorem never applies is a programme about an experiment, not about a result.

Frontier What survives all three theorems, taken together, is more than the coverage suggests. Relational and perspectival readings survive by denying absoluteness; Bohmian mechanics and dynamical-collapse theories survive by denying that an observer-scale system can be coherently reversed; many-worlds survives by denying single outcomes; superdeterministic and retrocausal accounts survive by denying free choice. Speculative The theorems narrow the price list. They do not pick a purchase, and no experiment performed so far has raised any of the prices.

Established The measured record is two photonic experiments, both from 2019–2020, and a short run of quantum-processor demonstrations since 2024. Every one of them uses a friend that its own authors decline to call an observer.

Established 2019, the six-photon test. Proietti, Pickston, Graffitti, Barrow, Kundys, Branciard, Ringbauer and Fedrizzi realised Brukner’s four-observer scenario at Heriot-Watt and reported it in Science Advances 5, eaaw9832. Their abstract is exact about what was done and what it shows: “In a state-of-the-art six-photon experiment, we realize this extended Wigner’s friend scenario, experimentally violating the associated Bell-type inequality by five standard deviations.” Frontier The inequality violated is a Bell-type inequality, not a genuine local-friendliness inequality; the distinction did not exist yet, because the LF polytope was published the following year.

Established 2020, the anchor: the first violation of a genuine local-friendliness inequality. Bong, Utreras-Alarcón, Ghafari, Liang, Tischler, Cavalcanti, Pryde and Wiseman published in Nature Physics 16, 1199–1205 a proof of the LF theorem together with a photonic test. For three binary-outcome settings the LF polytope has 932 facets in nine inequivalent classes; the genuine LF facet they target has an LF bound of 0, a maximum quantum violation of 1.345, and tolerates 18.3% white noise before the violation disappears. A second genuine facet reaches 0.880 with two-qutrit states at about 18.0% noise tolerance. Frontier Noise was never the limiting factor in this experiment, and that matters: the obstacle is not the apparatus.

Established What played the friend, exactly. The abstract says it in seven words: the theorem “is demonstrated in a proof-of-principle experiment where a photon’s path is deemed an observer”. In the body: “The friends are photon paths within the set-up” and “the measurements by the superobservers are photon-detection measurements”, on polarisation-encoded photons from type-I spontaneous parametric down-conversion in two orthogonally oriented BiBO crystals. Frontier The commission for this brief describes the anchor as an experiment in which the friend is a “measured path degree of freedom”. That is the one thing the path is not. Established The friend’s outcome is never measured and never recorded; it is encoded coherently so that the superobserver can undo it, and the only detections in the experiment are the superobservers’. Reversibility is the design, and a measured path would destroy it.

Established The size of the violation is the second correction, and it runs the other way from the field’s reputation. The paper reports that “all the experimental data points, except for the case of µ = 0.81, are at least two standard deviations away from 0, thus attesting the violation or non-violation of the inequalities with statistical significance”, across a state parameter µ from 0.744 to 0.992 at source fidelities of 0.9789 to 0.9887. Frontier The five-standard-deviation figure that attaches to this literature in secondary coverage belongs to Proietti and colleagues in 2019, on a Bell-type inequality, not to the genuine LF violation of 2020. Speculative By the standards of the loophole-free Bell tests this work is modelled on, a two-sigma effect is a preliminary result, and the field has not returned to improve it in six years.

Established The most instructive data points in the 2020 paper are the ones that do not violate anything. At µ = 0.80 and µ = 0.81 the experiment violates a Bell inequality while violating no LF inequality, which the authors single out as demonstrating “that it is possible to violate Bell inequalities without violating any LF inequalities”. Frontier That gap is the entire empirical content of the claim that LF is a stronger constraint than Bell, and it is a gap of two data points.

Established 2024–2025, the subject moves to quantum processors and acquires a scale. Zeng, Labib and Russo, in Quantum 9, 1851, compile extended Wigner’s friend scenarios into circuits and introduce the branch factor B = C_I − C_D, the interference complexity of two branches minus their distinguishability complexity. For an n-qubit GHZ friend, C_I = n and C_D = 1, so B = n − 1. They report violations of a semi-Brukner inequality whose theoretical maximum is 0.82843, sustained to B = 16 on Quantinuum’s H1-1 trapped-ion machine with a seventeen-qubit friend — 200 shots per trial over four trials, seven trials at size 17, error bars at three standard deviations — and to B = 4 on IBM’s Torino with 10,000 shots over ten trials. Frontier Their own framing is the most honest sentence in the experimental literature: prior photonic work sits at branch factor 0.0, and the ladder runs to an estimated 10¹⁹ for a human-level agent.

Frontier 2026, the circuit benchmarked rather than interpreted. Altman implemented on IBM’s ibm_fez backend a five-qubit instance of a Wigner’s-friend circuit family, measuring coherence witnesses on a four-qubit subset over 20,000 shots: visibility 0.8771 ± 0.0034, W_X = 0.8398 ± 0.0038, W_Y = −0.8107 ± 0.0041, and a phase-sensitive magnitude of 1.1673 ± 0.0040 against an ideal value near 1.414. Established The paper states that it “does not claim to test or discriminate among interpretations of quantum mechanics”. Speculative That is the correct register for a five-qubit run, and it is a useful contrast with how the 2019 and 2020 results were reported.

Frontier 2025, the first attempt at a friend with cognition in it, and it is a simulation. Zeng and Zhang trained an encoder-decoder convolutional network of 367,458 parameters to play the friend in an extended Wigner’s friend scenario, pre-training on 200,000 uncorrelated instances and training on 200,000 Born-rule instances with 8,000 held out, converging to a mean squared error of 0.025% by epoch 100, then generating 80,000 setting tuples at 100,000 measurements each. Established No quantum hardware was involved and no inequality was violated; the work is a framework and three proposed metrics. Speculative The authors concede the network “has not yet achieved the level of human cognition” and propose a quantum-computing version as the next step, which returns the problem to section 5.

Frontier The theoretical record is livelier than the experimental one, and in 2023–2026 most of the movement narrowed the claim rather than widening it. Utreras-Alarcón, Cavalcanti and Wiseman allowed the superobserver to query and reverse the friend repeatedly and proved that in such a sequential scenario “the violation of any Bell inequality within a sequential EWFS implies the violation of Local Friendliness” — so LF correlations collapse onto Bell correlations, and the gap between the two is a property of the scenario chosen rather than of the assumptions. Haddara and Cavalcanti then characterised the LF polytopes for arbitrary numbers of superobservers, friends, settings and outcomes, identifying exactly which scenarios collapse to Bell and which do not. Speculative The useful reading is that “LF is strictly stronger than Bell” is a statement about particular experimental designs, and the design used in 2020 is one of the ones where it holds.

Frontier 2025–2026, two results that move the target rather than the error bars. Del Santo, Manzano and Brukner argue in Physical Review Research 7, 033279 that the superobserver is not the privileged describer he is usually taken to be: each party holds information the other “fundamentally cannot have”, the two reside in different “bubbles”, and in a betting game the friend can out-predict Wigner. Mukherjee and Hance, in Physical Review Research 8, 023066 (22 April 2026), derive a timelike no-go theorem in which absoluteness of observed events, time symmetry, no retrocausality and a screening condition cannot all hold, and report that even a relaxed operational version of absoluteness is violated. Speculative Both results tighten the case against absoluteness without needing a friend at all, which is either the most promising development in the subject or an admission that the friend was never doing the work.

Frontier Against the whole programme stands a published objection that the theorems are weaker than advertised. Okon, in Entropy 27, 563 (2025), argues that Brukner’s theorem carries an unstated premise — that the superobservers’ results must coincide with the friends’ prior measurements — and that for local friendliness “it is not true that its assumptions are strictly weaker than those of Bell and it relies on the same mistaken application of standard quantum mechanics”. Established He observes that absoluteness of observed events contains two logically independent parts, and that its consistency component is not required for Bell’s theorem — so the comparison that makes LF “stronger” is not like for like. Speculative His conclusion, which the field has not answered in print, is that a true test “with observers of adequate complexity, is well beyond present technology”, and that this “greatly diminishes the force behind the LF result”.

Frontier The Frauchiger–Renner argument has its own standing rebuttal. Lazarovici and Hubert, in Scientific Reports 9, 470 (2019), argue that no inconsistency arises “if one considers a complete description of the physical situation”, and locate the error in an inference that applies the collapse rule for one agent and not for another while the superobserver is bringing decohered branches back into coherent superposition. Speculative Zukowski and Markiewicz reached a similar destination by a different route in Physical Review Letters in 2021. Two independent published rebuttals of the derivation, neither withdrawn, is an unusual state for a theorem eight years old.

Frontier Question one, and the question the whole subject reduces to: what makes something a friend? Three criteria are in print and they do not agree. Bong and colleagues left it deliberately open, offering instead a ladder: “experiments can test agents of increasing complexity; an experimental violation of LF inequalities with a given class of physical systems as ‘friends’ implies that either the LF assumptions are false, or that class of friends is not an ‘observer’”. Wiseman, Cavalcanti and Rieffel make thought sufficient, via an assumption they name Friendliness: if a system displays independent cognitive ability at least on par with my own, its communicable thoughts are as real as mine. Zeng, Labib and Russo replace the question with an operational proxy, the branch factor, which is computable from the circuit and says nothing about cognition. Speculative A photon’s path passes none of the three; a seventeen-qubit GHZ register passes the third at B = 16 and neither of the others. The field is arguing about the reference of a noun again, which is the same structural problem its sibling page reports.

Frontier Question two: is local friendliness really weaker than Bell? The claim is central — it is what makes an LF violation worth more than the Bell violations already in hand — and it is now attacked from two sides. Speculative Okon argues the comparison is malformed because absoluteness of observed events smuggles in a consistency condition Bell never needed. Frontier Utreras-Alarcón and colleagues show from inside the programme that in sequential scenarios the two coincide exactly. Speculative Neither shows the 2020 result is wrong. Both show that the headline is a claim about a scenario, and the scenario was chosen to make the claim true.

Frontier Question three: which assumption does a real violation actually kill? Speculative The theorem kills a conjunction, and an advocate of absoluteness can keep it by giving up locality or free choice, exactly as a local realist can keep locality in a Bell test by giving up measurement independence. Frontier What is different here is that the LF violation would leave locality standing in a sense Bell does not: Cavalcanti and Wiseman have argued that the violation has implications for quantum causality specifically because it does not require abandoning relativistic causal structure. Speculative The honest statement is that a violation with a genuine observer would make the relational and perspectival readings of quantum theory considerably more expensive to refuse, and would refute none of the alternatives outright.

Speculative Question four: is a friend that can be reversed a friend at all? Frontier Every version of the experiment requires the superobserver to undo the friend’s measurement, which means no irreversible record of the outcome may exist anywhere. Speculative A dynamical-collapse theory says that a system large enough to be an observer cannot be reversed, and therefore that the antecedent of the theorem is false as a matter of physics rather than engineering; the mass and complexity scales for that are the subject of Objective Collapse Models. Handwave An observer that leaves no trace of what it observed is not obviously the thing the word was coined for, and no paper in this literature defends the usage against that objection; they assume it and move on.

Speculative Question five: does any of this discriminate interpretations? Frontier The Frauchiger–Renner table assigns each interpretation a casualty, and in every case the casualty is something that interpretation had already given up. Speculative Wiseman, Cavalcanti and Rieffel claim one genuine kill: the stance that “quantum theory needs no interpretation” questions none of their assumptions and is therefore ruled out by their theorem. Handwave Ruling out a refusal to answer is a modest return on a no-go theorem, and it is the largest concrete interpretive claim this literature has made.

Speculative Question six: can the friend be mesoscopic rather than intelligent? The branch factor makes this a quantitative question for the first time: a friend is more observer-like to the extent that its branches are hard to recombine and easy to tell apart, and that is a property a memory register has without having thoughts. Frontier A run at B = 100 or B = 1,000 on a trapped-ion machine would not settle what an observer is, but it would make the denial that the friend is one carry a cost that grows with the number. Speculative Nobody has published where on that scale the denial is supposed to become unreasonable, and without such a number the ladder has no top.

Speculative Question seven, asked by nobody in the field and worth asking here: what would a null look like? Frontier If LF violations survive to B = 1,000 and then stop, the natural reading is a collapse mechanism, and the subject becomes an unusually clean test of one. Speculative If they survive indefinitely and the friend remains a register, the subject has produced a theorem whose antecedent nobody can satisfy. Handwave No paper in this literature states a pre-registered prediction for where the ladder should break, which means no outcome of the ladder can surprise anyone in it.

Established Bottleneck one, and the only one that matters: the friend. Eighteen orders of magnitude separate the best demonstrated friend from the friend the theorem is about. Branch factor 16 today; an estimated 10¹⁹ logical qubits and logical depth 10¹⁴ for a human-level agent. Speculative Nothing else in this brief is a bottleneck by comparison, and any section that treats loss, noise or statistics as the problem has mistaken a solved difficulty for the hard one.

Established Bottleneck two: the cost of reversibility, which has been costed and the number is bad. Rieffel’s estimate, reported in Scientific American in June 2024, is that for thoughts a human would take one second to think, the quantum-computed agent “would require more than 500 years”, with error correction alone needing roughly 1,000 physical qubits per logical qubit. Frontier Wiseman’s own assessment in the same piece: “In principle, I can’t see any reason that it can’t be done, but it is significantly harder than I had initially thought.” Speculative The 500-year figure is the single most useful number in this subject and it appears in a magazine rather than in the paper, which is a fact about how the field reports its own difficulties.

Frontier Bottleneck three: gate speed, not gate count. The proposal requires the agent to think and be reversed at something near human rates, and the authors flag this explicitly as “particularly challenging”, taking the cognitive task to be about a second. Speculative Quantum gate times are slow relative to classical logic and error-corrected logical gate times are slower still by the correction cycle; the deficit is multiplicative with depth, which is where the 500 years comes from. Handwave No published roadmap for a fault-tolerant machine targets a clock rate set by a metaphysical experiment, and none will.

Established Bottleneck four: the loopholes the photonic experiment left open, which its authors enumerate. A fully rigorous demonstration “would require, in addition to a more plausible ‘observer’, implementing shot-by-shot randomized measurement settings and closing separation, efficiency, and freedom-of-choice loopholes”. Frontier Every one of those is standard practice in loophole-free Bell testing and has been since 2015, and none has been done for a local-friendliness inequality in the six years since the theorem was published. Speculative The reason is not difficulty; it is that closing them on a photon-path friend improves the least interesting assumption in the set.

Speculative Bottleneck five: there is no agreed criterion of observerhood, and the three on offer are not commensurable. Thought, branch factor and complexity-ladder position measure different things; a system can score high on one and zero on another. Handwave A field that cannot say what would count as satisfying its theorem’s antecedent cannot say when its experiment has been performed, and this one has been in that position since 2020.

Frontier Bottleneck six, prior to all of them: the measurement problem itself. Okon’s charge is that standard quantum mechanics “is simply unable to make concrete predictions” for scenarios in which a measurement is later reversed, because the theory does not say when the state-update rule applies. Speculative If that is right, the experiments are testing a prediction the theory does not unambiguously make, and the no-go theorems are theorems about a formalism rather than about nature. Handwave The objection is answerable in principle by any interpretation that fixes the update rule, which is to say the field’s bottleneck is the problem it was built to adjudicate.

Established This subject depends almost entirely on a machine nobody is building for it. The decisive experiment needs a fault-tolerant quantum computer large enough to hold a thinking agent and fast enough to run it forwards and backwards, and every component of that requirement belongs to Fault-Tolerant Quantum Computing. Speculative The dependency runs one way and it is total: nothing about a Wigner’s friend test would change a fault-tolerance roadmap, and the roadmap sets the date.

Frontier The second dependency is a criterion, and it sits in a different discipline. Whether a running artificial agent is an observer is not a question quantum foundations can answer with a circuit, and the candidate suppliers are Machine Consciousness, Synthetic Consciousness and Consciousness Research. Speculative The last of those reports that the largest adversarial test the field has run damaged both theories it tested and that every objective marker in use was calibrated against verbal report. Handwave A criterion for observerhood is therefore not arriving from that direction on any schedule this brief can state, and Cavalcanti’s own move is to sidestep it: intelligence is measurable, whereas there is no possible test of whether anyone else is conscious.

Speculative The third dependency is a rival prediction, and it is the one that would make the programme interesting fastest. Objective Collapse Models predicts that reversal fails above some scale; local friendliness assumes it does not. Frontier The branch-factor ladder is, without anyone having designed it as one, a search for the scale at which collapse would bite — which is a stronger reason to climb it than the reason the field gives.

Frontier A fourth, cheap and missing: an experimental community. Two photonic groups produced the two experiments, six years ago, and neither has returned. Speculative The hardware results since have come from quantum-software groups and a solo author running circuits on cloud machines, which is a different kind of activity with a different standard of evidence.

Frontier The tests below are ranked by what a result would settle, and the ranking inverts the field’s own, which is ordered by what can be run this year.

Speculative Tier one, and the decisive experiment for this brief: run a local-friendliness test with a friend that is plausibly an observer — a coherently controlled artificial agent, or failing that a mesoscopic memory large enough that denying it observer status costs something. Frontier Every other component already exists: the inequality is derived and theory-independent, the violation is predicted, the noise tolerance is 18.3%, and the photonic version has been run. Speculative The missing component is the friend, and the proposal actually in the literature is Wiseman, Cavalcanti and Rieffel’s QUALL-E: a human-level artificial intelligence running on a universal quantum computer, whose thoughts make it an observer under their Friendliness assumption and whose substrate makes it reversible under Bennett’s logical reversibility. The cost is stated: roughly 10¹⁹ logical qubits at a logical depth of 10¹⁴, about 1,000 physical qubits per logical qubit, and more than 500 years of machine time for one second of thought. Nothing else on this list would settle the question, and nothing about this is scheduled.

Frontier Tier two, and the one already running: climb the branch-factor ladder until something breaks. Established The ladder has three rungs so far — 0.0 for photonic friends, 4 on IBM Torino, 16 on Quantinuum H1-1 with a seventeen-qubit friend — and each rung is a hardware problem with a known recipe. Speculative A run sustaining a semi-Brukner violation at B = 100 would be the largest coherent “observer” ever placed in superposition and reversed, and a run that fails at a reproducible B, with the failure tracking system size rather than gate error, would be the first positive evidence for a collapse mechanism from any experiment anywhere. Frontier This is the cheapest high-information experiment in the subject and it needs no new physics, only machine time.

Frontier Tier three: close the loopholes on the photonic test. Established The 2020 authors list what is missing — shot-by-shot randomised settings, spacelike separation, detection efficiency, freedom of choice — and all four have standard solutions in the loophole-free Bell literature. Speculative It would raise a two-sigma result to a proper one and would not touch the objection everyone actually has, which is why it has not been done and probably will not be.

Frontier Tier four: run the sequential scenario as a control experiment. Utreras-Alarcón, Cavalcanti and Wiseman showed that when the superobserver may query and reverse the friend repeatedly, LF violation and Bell violation coincide. Speculative Building both scenarios on the same apparatus and showing the predicted collapse of the gap would demonstrate that the experimenters control the distinction they are claiming to measure. Handwave No field enjoys running the experiment that shows its headline is design-dependent, and this one has not.

Speculative Tier five: a timelike test, which needs no second laboratory. Frontier Mukherjee and Hance’s 2026 theorem puts absoluteness of observed events in conflict with time symmetry, no retrocausality and a screening condition in a single-laboratory temporal scenario, and they report that a relaxed operational version of absoluteness is violated too. Speculative If that inequality can be measured on existing hardware it removes the spacelike-separation requirement from the programme entirely, which would be the largest simplification the subject has had; nobody has announced the measurement.

Speculative Tier six, not an experiment and the cheapest useful act available: publish the number at which the ladder is supposed to be convincing. Handwave A pre-registered statement of the branch factor, or the cognitive capability, at which the authors would accept that a friend is an observer would convert a sequence of demonstrations into a test. Speculative Without it, each new rung is reported as progress and each denial that the friend is an observer remains free.

Speculative A null on tier one is the outcome this brief expects more of, and it is informative. Frontier If a reversible agent can never be built, or if reversal fails at a scale well below cognition, then the antecedent of every theorem in this brief is false as physics and the whole class of results is about a possibility nature does not offer. Speculative That would be a substantial discovery, and it is the one this literature is least organised to claim.

Frontier The engineering requirement for the tier-two ladder is machine time on a trapped-ion processor and nothing else. The published rungs used 200 shots per trial across four trials, seven at the largest size, with three-standard-deviation error bars on Quantinuum H1-1, and 10,000 shots over ten trials on IBM Torino. Established The trapped-ion machine reached branch factor 16 where the superconducting machine stopped at 4, and the difference is gate fidelity and all-to-all connectivity rather than qubit count. Speculative That makes the next rungs a purchasing decision rather than a research programme, which is the most actionable sentence in this brief.

Speculative The requirement for tier one is the requirement for fault tolerance, restated with a clock attached. Roughly 1,000 physical qubits per logical qubit at current code assumptions, of order 10¹⁹ logical qubits, a logical depth of 10¹⁴, and the whole computation run forwards and then exactly backwards without an irreversible record escaping. Handwave Every large programme in Fault-Tolerant Quantum Computing is aimed at machines many orders of magnitude smaller than that, and none of them is aimed at running the computation in reverse.

Frontier One engineering constraint is peculiar to this experiment and is not a scaling problem: the agent must not leak. A single photon of thermal emission carrying which-outcome information out of the friend’s register destroys the reversal, which makes the isolation requirement stricter than for computation, where a leak costs fidelity rather than validity. Speculative This is the same constraint that limits macroscopic superposition experiments generally, and the mass and coherence-time record is tracked by Objective Collapse Models rather than here.

Speculative For the photonic route the engineering is conventional and solved elsewhere. Closing detection efficiency, spacelike separation and setting randomisation on an entangled-pair source is the standard loophole-free Bell recipe, built with superconducting nanowire detectors and short low-loss paths, and the components come from Precision Quantum Sensing and Quantum Networks. Handwave The hard part of this experiment has never been the optics.

Frontier The sibling page is Indefinite Causal Order, and the two subjects share an evidential shape precisely enough that the comparison is diagnostic rather than decorative. That page owns the process-matrix formalism, the quantum switch and the causal-inequality hierarchy, and this one does not restate any of it. Speculative What the two share is the move: derive an inequality from a small set of assumptions about agents in sealed laboratories, violate it photonically, and then discover that the dispute is about whether the physical thing playing the role in the theorem — a control degree of freedom there, a friend here — is the kind of thing the theorem is about. Frontier The instructive difference is in the numbers: that page reports two independent device-independent violations at 24 and 18 standard deviations, and this one reports a genuine-inequality violation at about two, six years old and unrepeated.

Frontier Objective Collapse Models is the page that would falsify this one, and it is the most useful adjacency here. Local friendliness assumes an observer-scale system can be coherently reversed; collapse models predict it cannot, above some scale. Speculative Read together, the branch-factor ladder is a collapse experiment with a foundations justification attached, and it is a better collapse experiment than it is a friendliness experiment.

Frontier The nearest existing page in the corpus is Consciousness Research, and the relationship is a demand this page makes that that page cannot meet. That brief owns the empirical science — the 256-subject adversarial test that damaged both theories it examined, the markers calibrated against verbal report — and none of it supplies a criterion for whether a running program is an observer. Speculative This page is distinct from it in exactly the way the commission says: it is about observer-relative facts, not about what observers are made of, and it needs the other question answered without being able to answer it.

Speculative The supply adjacency is Fault-Tolerant Quantum Computing, and the demand adjacency is Artificial General Intelligence. The decisive experiment needs both to arrive, and it needs them to arrive in the same box: a human-level agent running natively on a fault-tolerant quantum machine rather than on silicon. Handwave Neither programme has any reason to produce that combination, and a classical agent simulated on a quantum computer is what the proposal actually requires, at the cost section 5 gives.

Speculative Machine Consciousness and Synthetic Consciousness own the question the Friendliness assumption answers by fiat. Handwave Wiseman and colleagues do not claim to solve it; they assume cognitive parity is sufficient and say so. That is an honest assumption and it is the load-bearing one.

Frontier There is no programme, no budget line and no consortium; there are about five groups and one funded thesis lineage. The theory comes from Griffith, Vienna, ETH Zurich and a handful of philosophy departments; the two experiments came from Griffith and Heriot-Watt; the hardware work since 2024 has come from a quantum-software company and individual authors using commercial cloud machines. Speculative The subject has no institutional home and has not asked for one.

Speculative The institutional requirement that would do the most costs one meeting: a published convention on what a Wigner’s friend experiment claims. Three certificates are in circulation — a Bell-type inequality in an extended scenario, a genuine local-friendliness facet, and a coherence witness on a compiled circuit — and secondary coverage moves between them without marking the transition. A one-page statement distinguishing them, plus the branch factor of the friend used, would remove most of the confusion in the popular literature at no research cost.

Speculative The second is a pre-registered observerhood threshold, agreed before the next rung is climbed rather than after. Handwave The asymmetry is currently free in both directions: a violation is reported as evidence against absoluteness, and a sceptic denies the friend was an observer, and neither side pays. A number fixed in advance is what converts that into a test, and the field has had six years to publish one.

Frontier The third has money in it and belongs to someone else’s budget. The tier-two ladder needs commercial trapped-ion machine time, which is bought by the hour; the tier-one experiment needs a fault-tolerant machine, which is bought by the decade. Speculative The only realistic funding route for the decisive experiment is as a demonstration on hardware built for something else, which makes the timeline in section 13 a forecast about other people’s programmes.

Frontier The ethical question here is unusual: the decisive experiment requires creating a thinking agent, letting it observe something, and then deleting the observation by running it backwards. That is the protocol, not a side effect — reversal is what the theorem needs. Speculative The authors are aware of it and cite the literature on the moral consideration of artificial entities in the same paper that proposes the experiment, which is more than most proposals do.

Speculative The tension is sharp because the experiment’s premise is the ethical problem. The Friendliness assumption says that if a system has cognitive ability on a par with mine, its thoughts are as real as mine. Handwave An experiment whose validity depends on the friend’s thoughts being real, and whose method depends on erasing them, is not obviously coherent as an ethical matter, and no paper in this literature works the objection through. Speculative The nearest analysis on this map is Machine Consciousness, and it does not treat reversal.

Frontier The communication exposure is the same one every foundations result has, and here it is measurable in headlines. The 2019 six-photon result was reported in the popular press as showing that objective reality does not exist; the paper claims a five-standard-deviation violation of a Bell-type inequality in a scenario where the observers are photons. Speculative The gap between those two statements is the whole of section 15, and the cost of it is borne by the credibility of quantum foundations generally rather than by the groups involved.

Speculative There is no dual-use question specific to this subject and no data-protection question at all. Handwave Nothing here processes anyone’s information, nothing is deployable, and the only near-term societal contact is through the phrase “objective reality” appearing in headlines. Anyone who tells you otherwise is selling something.

Speculative The civilizational question is what it would mean if observed events were not absolute — not what a device would do, because there is no device and there is no plausible path to one. Every institution that adjudicates facts assumes that what happened happened for everyone: courts, instruments, scientific replication, accounting. Handwave A demonstration that absoluteness fails would not change any of them, because the failure is confined to a regime where one party can coherently reverse another party’s laboratory, and no institution operates there.

Speculative The sober version is a change in what physicists are entitled to assume, and it is not small. Relational and perspectival readings of quantum theory move from being one option among several to being the cheapest option; Frontier the alternatives survive but pay in locality, in free choice, or in denying that the reversible agent was an observer, and each of those payments has consequences elsewhere in physics. Speculative That is the realistic ceiling on this subject’s importance, and it is an importance internal to physics.

Speculative The asymmetry of outcomes is worth stating because the null is the more consequential result. A violation with a genuine observer confirms what the formalism already predicts; a reproducible failure of reversal at a stated scale would be the first positive evidence for a modification of quantum theory in a century of looking. Handwave The programme is organised to celebrate the less informative outcome, and that is a structural feature of no-go-theorem-driven research rather than a criticism of anyone in it.

Handwave Any claim beyond this — that observer-relative facts license a technology, a communication channel, or a revision of how societies establish truth — is an assertion, and this page does not make one.

These horizons track the branch-factor ladder and the fault-tolerance roadmap it depends on, because those are the only two schedules this subject actually has:

• 10 yr: Frontier Expect the branch-factor ladder to be climbed by one or two orders of magnitude on commercial trapped-ion and neutral-atom hardware, from 16 towards a few hundred, reported as semi-Brukner or local-friendliness violations with larger GHZ friends. Speculative Expect no loophole-free photonic LF test, because closing loopholes on a photon-path friend improves the least contested assumption. Expect at least one paper claiming a friend with cognition in it, running a small neural network as the friend on a quantum processor, and expect the dispute about whether it counts to be the same dispute as today’s. Frontier Expect the timelike no-go line to be measured, since it needs no second laboratory.
• 25 yr: Speculative This is the window in which the ladder either breaks or does not, and the break is the interesting outcome: a reproducible failure of reversal tracking system size rather than gate error would move the subject from foundations into experimental physics overnight. If it does not break, expect the field to have settled on branch factor as the operative definition of a friend and the philosophical objection to have moved on to why that definition is the wrong one. Handwave The modal outcome on this horizon is no break, no agreed criterion, and a larger number.
• 50 yr: Handwave The tier-one experiment requires a fault-tolerant machine of order 10¹⁹ logical qubits running a thinking agent forwards and backwards, against an estimate of more than 500 years of machine time per second of thought at present assumptions. Speculative Nothing forbids the assumptions improving by many orders of magnitude, and nothing in any published roadmap suggests they will; the honest statement is that the date is set by a curve nobody has drawn.
• 100 / 250+ yr: Handwave Beyond useful forecasting. The defensible structural claim is that a theorem whose antecedent requires a technology and whose consequent requires a definition is settled by the technology arriving or by the definition being agreed, and this one has had neither for six years while its inequality sat there, derived and ready.

• Depends on Four pages on this map, and the dependency is asymmetric in every case. Fault-Tolerant Quantum Computing owns the machine the decisive experiment runs on and sets its date. Artificial General Intelligence owns the agent that would have to run on it. Machine Consciousness and Consciousness Research own the criterion for whether that agent is an observer, and the latter reports that the field’s own markers were calibrated against verbal report, so the criterion is not arriving soon. Objective Collapse Models supplies the rival prediction that reversal fails above some scale, which is what the branch-factor ladder is really testing.
• Requires (not on this map) Five conditions, in the order of the tokens above. The criterion. A statement, agreed before the next result, of what makes a system an observer for the purposes of the theorem; three incompatible answers are in print and a photon’s path satisfies none of them. The machine. A fault-tolerant quantum computer able to run a thinking agent and reverse it exactly, estimated at 10¹⁹ logical qubits, logical depth 10¹⁴, about 1,000 physical qubits per logical qubit, and more than 500 years per second of thought. The loopholes. A photonic local-friendliness test with detection efficiency, spacelike separation, shot-by-shot randomised settings and freedom of choice all closed, which the 2020 authors list as missing and which nobody has supplied. The reversal. A demonstration that a system which has actually recorded an outcome can be coherently reversed, which is the antecedent of every theorem here and has never been done above the scale of a few qubits. The convention. A published statement distinguishing a Bell-type violation in an extended scenario from a genuine local-friendliness violation from a coherence witness on a compiled circuit, and reporting the friend’s branch factor alongside the result.
• Enables No typed enabling edge is claimed, and that is the finding rather than an omission. A confirmed violation with a genuine observer would change which interpretations of quantum theory are cheap and which are expensive; it would not enable a device, a channel or a measurement anyone would buy. The one downstream consequence worth naming is negative: a reproducible failure of reversal at a stated scale would be direct evidence for Objective Collapse Models, which is the only page here that would gain a measurement from this one.
• Adjacent Indefinite Causal Order runs the same argumentative structure on operation order rather than on observers, with an experimental record an order of magnitude stronger in sigma and the same unresolved dispute about what the physical role-player is; the two pages should be read together. Synthetic Consciousness owns the question the Friendliness assumption answers by fiat. Precision Quantum Sensing and Quantum Networks build the detectors and sources a loophole-free photonic version would need. Quantum Gravity is adjacent only rhetorically: nothing in this subject requires it.

Established “Experiments have shown that objective reality does not exist.” They have shown that certain correlations violate an inequality derived from three assumptions, on the premise that a photon’s path is an observer. The 2020 abstract states the premise as a conditional — “if quantum evolution is controllable on the scale of an observer” — and the paper explicitly allows that “the LF assumptions are valid, but that nature forever forbids the observation of violation of LF inequalities with observers”. Frontier Nothing in the experimental record bears on whether facts about your kitchen are observer-relative, and no author in this field claims it does.

Established “The six-photon experiment violated a local-friendliness inequality.” It did not; genuine LF inequalities did not exist in 2019. Proietti and colleagues report violating “the associated Bell-type inequality by five standard deviations” in Brukner’s four-observer scenario. The first violation of a genuine LF facet is Bong and colleagues in 2020, and the paper’s own most striking data points show a Bell violation with no LF violation at all, at state parameter 0.80 and 0.81. Frontier Merging the two is the commonest factual error in coverage of this field, and it converts a two-sigma result into a five-sigma one.

Frontier “In the anchor experiment the friend’s path was measured.” The opposite. Established The abstract says “a photon’s path is deemed an observer”, the body says “the friends are photon paths within the set-up” and that “the measurements by the superobservers are photon-detection measurements”. The friend’s outcome is never detected; it is encoded coherently so the superobserver can undo it, and if it were measured the experiment would not work. This brief’s commission describes the anchor as a “measured path degree of freedom”; the correction is recorded here rather than smoothed over, because the reversibility is the point.

Frontier “Frauchiger and Renner proved that single-world interpretations are inconsistent.” They proved that (Q), (C) and (S) cannot all hold, with the contradiction arising in 1/12 of runs. Established Their own table has Bohmian mechanics — a single-world theory — giving up (Q) and remaining consistent. Frontier Lazarovici and Hubert argue in Scientific Reports that “no such inconsistency occurs if one considers a complete description of the physical situation”, locating the error in an inconsistent application of the collapse rule; Zukowski and Markiewicz published a separate rebuttal in Physical Review Letters. Speculative Two unwithdrawn published rebuttals is not the profile of a settled theorem.

Frontier “Local friendliness assumes strictly less than Bell.” Its authors say so and it is the claim that makes the programme worth running. Okon denies it in Entropy 27, 563, arguing that absoluteness of observed events has two logically independent parts and that its consistency component is not among Bell’s assumptions, so the sets are not nested. Utreras-Alarcón, Cavalcanti and Wiseman show from inside the programme that in sequential scenarios LF violation and Bell violation are equivalent. Speculative The defensible position is that LF is stronger than Bell in the specific scenario the 2020 experiment used, which is a narrower claim than the one in circulation.

Frontier “These experiments test whether consciousness causes collapse.” No experiment in this literature tests that, and no paper in it claims to. Speculative The Friendliness assumption is about cognitive parity, not consciousness, and Cavalcanti’s stated reason for framing it that way is that intelligence is quantifiable whereas there is no possible test of whether anyone else is conscious. Handwave That is a sidestep and the authors present it as one.

Frontier “A quantum computer has now run the experiment with an artificial-intelligence friend.” It has not. Established The largest hardware friend on record is a seventeen-qubit GHZ register at branch factor 16 on Quantinuum H1-1; the 2026 IBM run used five qubits and its author states it “does not claim to test or discriminate among interpretations of quantum mechanics”; the neural-network friend of 2025 was trained and evaluated entirely in simulation. Speculative The gap between the largest friend built and the friend the theorem needs is about eighteen orders of magnitude on the field’s own scale.

Speculative “The experiment is impossible in principle.” Frontier Wiseman’s own position is that he cannot see a reason it cannot be done, only that it is significantly harder than he first thought. Speculative The impossibility claims on offer are physical rather than logical — a collapse mechanism, or a limit on coherent control of complex systems — and each is a testable proposition rather than a proof. Okon’s weaker claim, that the test “with observers of adequate complexity, is well beyond present technology” and that this “greatly diminishes the force behind the LF result”, is the one this brief endorses.

Speculative “Branch factor measures observerhood.” It measures how hard two branches are to recombine relative to how easy they are to distinguish, which is a circuit property. Frontier It is the only quantitative proxy anyone has proposed and it is genuinely useful, because it puts the photonic experiments at 0.0 and makes the ladder legible. Handwave Nothing connects a number on that scale to having thoughts, and the authors do not claim it does; the estimate that a human-level agent sits near 10¹⁹ is an extrapolation from resource counts, not a measurement of anything.

Speculative “This is philosophy, not physics.” Established The inequalities are theory-independent, derived from stated assumptions, with computed bounds, quantum maxima and noise tolerances — 932 facets, a bound of 0, a maximum of 1.345, 18.3% white-noise tolerance — and two of them have been measured. Speculative The part that is not physics is the antecedent: what counts as an observer. The useful criticism is not that the field is untestable but that its most-quoted claim and its most-testable claim are different claims.