Here's what this paper means for actual humans: WHAT THIS IS ABOUT. When you build or use a quantum device, everything rests on knowing what its parts are: which degrees of freedom are "qubit 1," which are "qubit 2." The companion paper ("Forged Frames") showed that simple mathematical criteria cannot certify this; convincing fakes exist. This paper answers the question that result leaves standing: what is the MINIMUM experimental information that genuinely pins down a system's subsystem structure, and when can that information silently fail? WHAT WE FOUND, IN PLAIN TERMS. Four things. (1) EXACT LAWS INSTEAD OF MEASUREMENTS. Any wrong choice of parts must visibly distort the family of unentangled states, and we derive exactly how much, at every system size: closed-form constants (verified to seven decimal places), an explicit error bound, and a surprise. The two natural ways to measure "wrongness" (state distortion versus operator scrambling) turn out to be literally the same measurement, up to a fixed number, whenever the two parts have equal size. Wrongness cannot hide. (2) THE TWO-OPERATION THRESHOLD. You can pin down what your subsystems are by implementing just TWO generic operations and checking their statistics, while a THOUSAND mutually commuting operations tell you nothing at all. Diversity of control, not volume of control, is what fixes what your qubits are. The residual uncertainty is proportional to your measurement tolerance, with a constant we derive exactly and show how to compute in advance, so an experimenter can screen candidate operation pairs and pick a strong one (random pairs are typically several times worse, occasionally catastrophically so). (3) THE ALIASING LIMIT: CERTIFICATION CAN PASS WHILE TELLING YOU NOTHING. If you sample your certification data at special "resonant" times (set by the operations' energy scales), a maximally wrong assignment of subsystems passes the test perfectly. What your qubits are depends on when you looked. We characterize these blind spots exactly (for qubits they are universal and computable in advance; the danger windows shrink with the square root of your tolerance), and the fix costs almost nothing: sample at two different times, at least one away from the resonant set. With that, certification is restored; we prove it exactly and measure it. (4) REAL-LAB BUDGETS. We work out what certification costs in actual measurement shots (the error falls off unusually fast, inversely with the total number of shots rather than with its square root), how hardware noise floors the procedure, and how a simple reference measurement removes that floor to first order. A single seven-row table in the paper is effectively a lab protocol: it shows, condition by condition, which kinds of experimental access identify the subsystem structure (two generic operations: yes; one operation, commuting operations, resonant sampling: no). WHY IT MATTERS. For quantum hardware: this is a concrete recipe for certifying that the qubits you think you have are the qubits you physically have, with explicit tolerances, a sampling-time warning that is a real failure mode on calibrated hardware, and shot budgets. For foundations: the contested question "does physics determine its own subsystems?" gets a quantitative operational answer. Physics determines an operational subsystem CLASS (not one preferred coordinate representation of it), and that class is fixed by two operations sampled at two times, and by nothing less that we tested. WHAT IT OPENS UP. The seven-row identifiability table is a blueprint for a tabletop experiment on any small quantum platform; running it on hardware would move this from simulation to laboratory fact. On the theory side the paper leaves sharply posed open problems: the exact global constants when the two parts differ in size, the number-theoretic structure of long-time sampling (the failure times form a small-divisor hierarchy), and the extension from two parts to many. METHODS NOTE, BECAUSE IT MATTERS FOR TRUST. Every quantitative claim was pre-registered in the companion scripts before its run. The full ledger, including predictions that failed and one published-in-note constant that our own later theory overturned, is an appendix of the paper. All code, data, and working notes are in the deposit, seeded for exact reproduction. WHO AND HOW. A human-AI collaboration: research direction and standards of evidence by Ryan Cardwell (Crystalline Labs); design, code, derivations, and drafting by Claude (Anthropic), including autonomous sessions under the program's pre-registration rules, with adversarial reviews by other AI models incorporated and credited.
Here's what this paper means for actual humans: WHAT THIS IS ABOUT. When you build or use a quantum device, everything rests on knowing what its parts are: which degrees of freedom are "qubit 1," which are "qubit 2." The companion paper ("Forged Frames") showed that simple mathematical criteria cannot certify this; convincing fakes exist. This paper answers the question that result leaves standing: what is the MINIMUM experimental information that genuinely pins down a system's subsystem structure, and when can that information silently fail? WHAT WE FOUND, IN PLAIN TERMS. Four things. (1) EXACT LAWS INSTEAD OF MEASUREMENTS. Any wrong choice of parts must visibly distort the family of unentangled states, and we derive exactly how much, at every system size: closed-form constants (verified to seven decimal places), an explicit error bound, and a surprise. The two natural ways to measure "wrongness" (state distortion versus operator scrambling) turn out to be literally the same measurement, up to a fixed number, whenever the two parts have equal size. Wrongness cannot hide. (2) THE TWO-OPERATION THRESHOLD. You can pin down what your subsystems are by implementing just TWO generic operations and checking their statistics, while a THOUSAND mutually commuting operations tell you nothing at all. Diversity of control, not volume of control, is what fixes what your qubits are. The residual uncertainty is proportional to your measurement tolerance, with a constant we derive exactly and show how to compute in advance, so an experimenter can screen candidate operation pairs and pick a strong one (random pairs are typically several times worse, occasionally catastrophically so). (3) THE ALIASING LIMIT: CERTIFICATION CAN PASS WHILE TELLING YOU NOTHING. If you sample your certification data at special "resonant" times (set by the operations' energy scales), a maximally wrong assignment of subsystems passes the test perfectly. What your qubits are depends on when you looked. We characterize these blind spots exactly (for qubits they are universal and computable in advance; the danger windows shrink with the square root of your tolerance), and the fix costs almost nothing: sample at two different times, at least one away from the resonant set. With that, certification is restored; we prove it exactly and measure it. (4) REAL-LAB BUDGETS. We work out what certification costs in actual measurement shots (the error falls off unusually fast, inversely with the total number of shots rather than with its square root), how hardware noise floors the procedure, and how a simple reference measurement removes that floor to first order. A single seven-row table in the paper is effectively a lab protocol: it shows, condition by condition, which kinds of experimental access identify the subsystem structure (two generic operations: yes; one operation, commuting operations, resonant sampling: no). WHY IT MATTERS. For quantum hardware: this is a concrete recipe for certifying that the qubits you think you have are the qubits you physically have, with explicit tolerances, a sampling-time warning that is a real failure mode on calibrated hardware, and shot budgets. For foundations: the contested question "does physics determine its own subsystems?" gets a quantitative operational answer. Physics determines an operational subsystem CLASS (not one preferred coordinate representation of it), and that class is fixed by two operations sampled at two times, and by nothing less that we tested. WHAT IT OPENS UP. The seven-row identifiability table is a blueprint for a tabletop experiment on any small quantum platform; running it on hardware would move this from simulation to laboratory fact. On the theory side the paper leaves sharply posed open problems: the exact global constants when the two parts differ in size, the number-theoretic structure of long-time sampling (the failure times form a small-divisor hierarchy), and the extension from two parts to many. METHODS NOTE, BECAUSE IT MATTERS FOR TRUST. Every quantitative claim was pre-registered in the companion scripts before its run. The full ledger, including predictions that failed and one published-in-note constant that our own later theory overturned, is an appendix of the paper. All code, data, and working notes are in the deposit, seeded for exact reproduction. WHO AND HOW. A human-AI collaboration: research direction and standards of evidence by Ryan Cardwell (Crystalline Labs); design, code, derivations, and drafting by Claude (Anthropic), including autonomous sessions under the program's pre-registration rules, with adversarial reviews by other AI models incorporated and credited.
Here's what this paper means for actual humans: WHAT THIS IS ABOUT. The two companion papers established, for a system split into TWO parts, that you can certify what a quantum device's subsystems are from remarkably little: two generic operations, their statistics, two sampling times. Real devices have many parts. This paper carries the entire theory from two parts to many, and closes the last open corner of the two-part theory on the way. WHAT WE FOUND, IN PLAIN TERMS. Four things. (1) A LAW FOR EVERY CUT AT ONCE. Any wrong choice of parts distorts the family of unentangled states, and for many parts we derive exactly how much, for every way of cutting the device in two and every pattern of deformation across the parts: one closed formula whose weights land on exact fractions (verified to nine decimal places with a sampling-free method). The two-part law is literally the corner case. A structural surprise falls out: with many parts, rigidity is unavoidably direction-dependent; the meaningful object is a weight function over subsets of parts, not a single constant. (2) THE DIAGNOSTIC TOOLKIT GOES FULLY CLOSED-FORM. The standard "how wrong is this factorization" measures turn out to be exact linear functions of just two matrix norms of the frame (the same matrices behind textbook entanglement tests), with coefficients we give in closed form at every pair of dimensions, validated blind on dimension pairs the fit never saw. The two-part frontier constant decomposes into per-side constants with a clean derived formula, and the worst-case ratio of scrambling to visible distortion is computed exactly at three dimension pairs (28/11, 5/2, 265/138), each achieved at an exact-integer point by an explicit saved transformation. (3) CERTIFICATION IS PER-NODE. For many parts, which structure you pin down depends on which operations you certify, node by node on the lattice of ways to partition the device: a node is pinned exactly when your certified operations generate that node's algebra of local actions. Certify only one cut and the interior of a block stays COMPLETELY free (we exhibit an explicit transformation that passes every certificate while maximally scrambling the interior); add one more operation inside the block and it is still free; add two and everything locks. Counting operations is meaningless; per-node generation is everything. (4) A WEAK SECTOR BLINDS THE WHOLE DEVICE. The sampling-time blind spots found in the two-part theory (times at which certification passes while revealing nothing) survive with a sharper lesson: a forgery built for two resonant qubits, simply tensored with "do nothing" on the rest, passes every certificate of the full device while scrambling every cut that touches those qubits. On all-qubit hardware the dangerous sampling times are therefore a property of the whole machine, set by its calibration scale. The known fix (sample at two times, one generic) carries over unchanged. WHY IT MATTERS. For hardware: certifying a large device decomposes into certifying lattice nodes, each with a concrete requirement (generation-complete operations, non-resonant sampling), and the paper says exactly what an uncertified block costs you: everything inside it. For theory: the two-part identifiability program is now a special case of a lattice theory with closed-form constants throughout. HOW MUCH TO TRUST IT. Every quantitative claim was pre-registered in the companion scripts before its run. The appendix ledger lists every prediction: the hits, five failed predictions with their mechanisms, and one computation we invalidated ourselves before it could ship. Negative results rest on exact constructive witnesses, not on searches failing. All code, data, saved transformations, and working notes are in the deposit, seeded for exact reproduction, and every external reference was verified against a live source before citation. WHO AND HOW. A human-AI collaboration: research direction and standards of evidence by Ryan Cardwell (Crystalline Labs); design, code, derivations, and drafting by Claude (Anthropic), operating autonomously under a written research charter. The full arc, from empty directory to this draft, ran in one day of charter-governed iterations.
Version 4 (2026-08-25). Two additions from the correction campaign, carried as dated in-body correction records. The (2,5) entry, printed as a prediction in v3, is now census-exact -- and it was the wall law's first divergence test. At the constructed maximal-wall stratum point the stiff transverse count is exactly 10, so alpha_stratum(2,5) = 5; a fresh 60k ensemble's calibrated read is consistent with 5, with alpha = 10 (the superseded single-wall count's prediction) excluded at p = 0.000 and demonstrated power ~ 1; and a stratum-conditional paired ladder tracks the exact rank-10 Hessian form to <= 0.002 while a rank-20 comparator separates by >= 2.7 on identical draws. The law m_C(2,d) = 2(d-1) is census-exact at d = 3, 4, 5, now including the first point where it diverges (by 2x) from the superseded count. The (3,4) entry, printed open in v3, is now a PAIR of statements, and neither alone is alpha(3,4). The posted product-wall counting landed census-exact (alpha_stratum(3,4) = 5, so the asymptotic exponent satisfies alpha(3,4) <= 5), while the accessible window at n = 3e4 is steeper (calibrated 90 percent set [5.4, 6.9], with the previously read 4.0-4.1 excluded at power ~ 1): the codim-10 stratum's prefactor has not engaged at reachable depth, and window and asymptote have visibly decoupled. This is the scope lesson of the (d >= 3) x (d' > d) regime -- census plus conditional ladder measure the law, global windows measure the pre-asymptotic shoulder -- and it retro-explains the earlier fresh-seed ladders' instability as shoulder reads. Version 3 (2026-08-20). One SUPERSESSION: the qutrit-qutrit exponent closes. Versions 1 and 2 left alpha(3,3) open, v1 by a small-sample fit reading 3.25 and v2 by the bracket [2.3, 3.0] read off a descending window ladder. Both readings were artifacts of the estimator problem for which the pair-type table was already corrected in v2. The descent is attenuation tracking sample count rather than tail depth: an ensemble drawn from an EXACT alpha = 3.0 law descends by 0.62 across the ladder against the real data's 0.34, so a descending ladder is evidence FOR a constant exponent, not against one. Replacing the window fit with a calibrated Monte-Carlo test and inverting it over candidate exponents gives alpha(3,3) = 3.0 at p = 0.517 with a 90 percent confidence set [2.75, 3.10]; v1's 3.25 is excluded at p = 0.017. Power against alternatives is 0.990 / 0.590 / 0.120 / 0.575 / 0.990 / 1.000 at true alpha = 2.50 / 2.75 / 3.00 / 3.25 / 3.50 / 4.00. An independent re-audit at n = 10^4 reads flat near three on two seeds. Disclosed defect, carried in the paper: the test's own calibration row missed at the low edge (rejection rate 0.207 at a nominal 0.10 when the truth is alpha = 2.50, 4.4 standard errors high), so the lower end of the confidence set is softer than printed. This ensemble separates 3.0 from 3.25 with about 58 percent power and has none to separate 3.0 from 2.95. The (3,3) CARRIER family remains unattributed: the exponent is settled, the stratum that carries the accessible tail there is not. Version 2 (2026-08). Changes from v1: the pair-type exponent table is corrected (v1's alpha(2,4) = 2.08 was a 50-sample order-statistics fluctuation, placed at the 0.15th percentile of its own estimator's sampling distribution; three independent codebases now agree on the wall law alpha(2,d) = d, census-exact at d = 3 and 4); the transverse constant 1/12 is derived in closed form (diagonal cost 1/6, halved by the congruent partner's equal-cost correction); the (2,3) tail's carrier stratum is fully attributed (an isotropic-intertwiner family sharing the exponent 3); the qutrit-qutrit exponent is honestly bracketed open at [2.3, 3.0]; and the dispersion rule (tail-exponent fits print bootstrap dispersion or do not print) is adopted. The v2 ledger appendix carries the correction anatomy. Why is the multipartite certification constant heavy-tailed, where does the tail live, and what kills it? Localization: the credential pencil is exactly block-diagonal over party-subset supports; the tail is the worst pair; geometry statistics reconstruct from pair statistics (14/14 factor-band comparisons). The law: a frame-congruence theorem: whenever the two parties' operation frames subtend equal Bloch angles, the rotation intertwining them is an exact common kernel direction; the zero set is codimension one at generic angles, transverse growth is universally quadratic (mu = (Delta theta)^2/12), and the qubit-pair tail exponent is exactly 1/2 (the previously measured 0.39 is the finite-window slope of this law). The dichotomy: pairs containing a d >= 3 party measure exponents 2.08 to 3.69 because higher-dimensional operation pairs carry many local invariants that must all match: freedom helps the certifier. Engineering: heterogeneous dimension profiles tame the tail; qubit-pair presence splits all measured geometries into two clean classes; screening is a five-minute two-party calculation. Pre-registered; two entry gates that caught wrong conventions and all deaths ledgered.
Here's what this paper means for actual humans: WHAT THIS IS ABOUT. Quantum mechanics describes a system as a whole, but physics, and every quantum computer, depends on carving that whole into parts: this qubit versus that one, the system versus its environment. Mathematically, the carving is called a tensor product structure (TPS), and here is the uncomfortable fact this paper starts from: the carving is not written into the theory. It is extra structure someone has to supply. Several research programs have proposed that the dynamics alone (the Hamiltonian, sometimes plus a quantum state) singles out the "right" parts automatically, so that subsystems emerge rather than being chosen. WHAT WE DID. We tested that claim the direct way. We built quantum systems where the true division into parts is known by construction, because we engineered it in ourselves. Then we hid it, handed an optimizer only the ingredients the proposals say suffice (the Hamiltonian, or the Hamiltonian plus one state, or how entanglement grows in time), and let it search over every possible alternative carving to see whether the criteria recover the truth. WHAT WE FOUND. They do not. At every size we tested (4- up to 128-dimensional systems), the optimizer finds "fake" carvings that score better than the engineered truth on all of the simple criteria, often 40 to 90 times better, and the true carving is not even a local optimum: a search started exactly at the truth walks away from it. A single state's snapshot can be forged too. But the story has a second half. When we strengthened the test to require agreement on a whole family of unentangled states (the "product manifold") rather than a single snapshot, the surviving impostors turned out not to be alien worlds at all. They are the true carving rewritten in "dressed" coordinates, the same quasiparticle-style change of variables condensed-matter physics has used since the 1960s (Schrieffer-Wolff transformations), and in this final version we construct that transformation explicitly and show the optimizer's answer sits on it. Genuinely alien carvings that mimic the whole state family do not appear, and we measure quantitatively how impossible their job is (a "rigidity frontier"), anchored by an exact classical theorem. WHY IT MATTERS PRACTICALLY. (1) Any algorithm, simulation method, or interpretation-of-quantum-mechanics program that takes "the least-interacting factorization" or "the factorization where my state looks simple" as input inherits the failures documented here; we provide the counterexamples and the code to generate more. (2) The positive result sets the bar for what does work: agreement on a family of states identifies the physical parts up to a harmless, physically meaningful change of coordinates. (3) The whole test suite is released with fixed seeds, so every number reproduces, and the corrections we made along the way, including an inference we got wrong and later withdrew, are recorded in the paper rather than erased. WHAT IT OPENS UP. The paper ends with a sharpened question: how LITTLE experimental access suffices to pin down the parts? Not the full state family, but a handful of implemented operations known only through their measurement statistics? That question gets a quantitative answer (two operations, sampled at two times) in the companion paper archived alongside this one, which turns the boundary measured here into exact laws and laboratory protocols. WHO AND HOW. A human-AI collaboration: research direction and standards of evidence by Ryan Cardwell (Crystalline Labs); experimental design, code, numerics, and drafting by Claude (Anthropic), with adversarial reviews by other AI models incorporated and credited. All results reproduce from the archived scripts in the deposit.
Experimental supplement to arXiv:2611.08310, a companion manuscript on gate-optimal recovery designs (2026). This supplement specifies a concrete experiment for measuring the eleven-parameter small-angle coherent noise of a three-qubit register with three single-qubit Pauli gates, using a distance-three repetition code in both orientations. It gives the branch structure, the syndrome-to-Pauli look-up table for the three-gate design, the fixed sixteen-setting measurement table, shot and wall-clock budgets for three platform classes, a calibration and systematics checklist, an analysis pipeline, and a symptom-to-cause table. Two sections price the alternatives: a fixed recovery that does not resolve the syndrome saves a factor three in circuits and loses a factor five in precision floor, and a coherent unitary recovery keeps the circuit saving but replaces each single-qubit Pauli by a multi-controlled gate that every branch must compile. At three qubits no design in the scanned budget loses rank, so identifiability does not separate the gate counts; what depends on the working point is which three gates to use, not how many. The organising point is that the recovery gate count is a resource: three gates attain the lowest precision floor, the standard six-gate decoder is 22 times worse at a small-angle operating point and 2.3 times worse at the well-conditioned point of arXiv:2610.19779, and the ordering is stable over forty sampled working points while the size of the factor is not. The ordering survives when the look-up is compiled rather than measured, and the margin then widens from 5.6 to 7.7 in wall clock, because a compiled table charges every entry in every branch. A section added in version 1.5 asks whether that answer depends on the platform. A second device archetype -- a tunable coupler or iSWAP-style gate, whose always-on term is the exchange J(XX+YY)/2 -- carries one more parameter per bond, thirteen in all, and the two clusters are not nested: XX and YY commute, and ZZ is orthogonal to their plane. Enumerating all 64 design pairs at n=3 and all 3969 at n=5 gives the same gate budget for both archetypes, zero gates to reach full rank at n=3 and one at n=5; that count is a lower bound on the budget, not a recommended design, since the precision optimum is the three gates of the main text, where zero gates is 18 times worse at the median and 379 times worse at the certified point. At matched per-layer coherent infidelity the thirteen-parameter cluster is not worse (101.0 against 88.9 mrad on the local angles at 10^5 shots), and at the three-gate optimum the comparison widens (41.7 against 32.2 mrad). Choosing the wrong cluster is a bias, not a variance: the correct family leaves a residual of exactly zero, the wrong one leaves 0.4-0.6% in the fingerprint and moves the nine local angles by a quarter to a half, while its Cramer-Rao bound looks better. The decoded logical error cannot arbitrate at that bias level -- 200 random perturbations of the same total variation spread it over a factor 43 -- so the model family must be chosen by the residual. The same section closes with a composition law: a per-gate composition of Pauli probabilities is elementwise in the Walsh eigenvalues and therefore under-counts the coherent part of the depth-L noise by exactly L (measured slopes 1.98 for the generator against 0.99 for the per-gate rule). A closing section fixes the size of the gain and asks how far it extends. An exact prior removes 19.3% of the logical error at d=3, which is 4.3e-5 errors per code cycle; a three-gate fit captures 83, 96.5 and 99.5% of that share on the three platform classes. That share, not the accuracy of the estimate, is where every prior-based protocol stops, because 94% of the residual is weight-two aliasing. It is not a limit set by using Pauli probabilities: the same prior decodes both the coherent channel and its twirled version, so the drop from 19.3% to 11.6% under twirling prices the device rather than the estimate, and composing per-gate twirls changes at most two of the 256 decoder decisions. A generator rather than a channel is what additionally allows the coherent part to be inverted and the prior to be extrapolated to unmeasured working points. Carrying the same three-qubit fit to d=5 or d=7 obeys a closed form, G_d = G_3^((d+1)/4), so no extra precision is needed at larger distance: a 6.4 mrad rms on the eleven parameters moves the d=7 gain by 8.5%. What does not transport is the truncated part of the model---the 2(d-1)^2 three-body positions a d x d patch adds, none of which a chain fit covers---which costs 5% of the gain at d=5 and 10% at d=7; site-to-site spread costs another 3%. The law itself is a property of the decoder and applies to any prior, so this is a bound on what a three-qubit run can claim, not an advantage over Pauli-statistics protocols. Measuring at the target size instead of transporting avoids the truncation count but pays conditioning: a factor 68 from three to seven qubits. All numbers are model-level: no device data are reported, and the floors assume a calibration residual of 1e-3. Contents (v1.5): 1 What the experiment measures · 2 The three branches · 3 Syndrome decoding · 4 Settings and shot arithmetic · 5 Budget on three platforms · 6 Why the syndromes must be resolved · 7 What does and does not depend on the gate count · 8 Does the platform change the answer? · 9 How large the gain is, and how far it extends · 10 Calibration and systematics · 11 Analysis pipeline · 12 Diagnostics · 13 Checklist.