{"id":"7f6547c6-56f7-4acd-a634-dfdd49493c78","arxiv_id":"2607.17886","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The charge-learnability transition in U(1)-monitored circuits is controlled by I_loc = q·I_read(η), so measurement probability and strength trade off along constant-information curves.","lead":"This paper studies quantum circuits with a conserved charge, where random weak measurements leave a record from which an observer tries to guess the total charge. It finds that the point where guessing succeeds is set by a single 'informational power' of the measurement, regardless of whether measurements are strong and rare or weak and frequent.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim assumes single-site I_loc determines the global record–label information I_true; this is untested and in tension with Sec. V's admission that spacetime correlations remain essential. Exact comparison along a constant-I_loc contour would settle it.","rationale":"The paper is cleanly written and makes an honest conjectural statement in Sec. V, and the exact record–label mutual information calculation (Eq. 24) plus the Fano-bound comparison in Fig. 7 are valuable decoder-independent checks. However, the abstract's claim to 'establish informational power of local measurement as a unifying principle' goes beyond what is shown. The single most load-bearing step is the identification of a local, single-readout quantity I_loc with the global phase boundary. This step is unsupported analytically, is challenged by the paper's own concession that space-time correlations remain essential, and is backed only by small-system Binder crossings without error bars or extrapolation. The proposed exact I_true comparison along a constant-I_loc contour is decisive because I_true is already computed exactly at the accessible sizes; if equal-I_loc protocols give different I_true, the collapse cannot be fundamental and the central claim is not established. Because the authors themselves label the criterion conjectural, the appropriate verdict remains CONDITIONAL rather than ACCEPT or REJECT; my analysis does not change the reader's verdict, hence UNCHANGED.","tokens_in":12076,"tokens_out":9231,"duration_ms":91953,"concrete_test":"Compute exact I_true(Y;M) via Eq. (24) for L=8, T=16 at four points on the I_loc=0.20 contour: (q=0.20, η=1), (0.40, 0.480), (0.60, 0.371), (1, 0.277), using at least 10^5 trajectories per point with bootstrap standard errors. If the four I_true values are not equal within ~2σ, then the full record–label information is not determined by I_loc and the phase boundary cannot be organized by local informational power alone. If they are equal, repeat at the mutual-information boundary I_loc≈0.16 to confirm; this directly tests whether the admitted space-time correlations are encoded in I_loc.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the learnability phase boundary is a contour of I_loc(q,η)=q I_read(η) (Eq. 21), unifying projective, weak, and probabilistic weak measurements. The load-bearing condition is that this single-site, single-readout quantity controls the full space-time record–label mutual information I_true(Y;M) (Eq. 23), because I_true is what fundamentally limits charge inference. This condition is not derived, and it is in tension with the paper's own Sec. V admission that 'space-time correlations within the measurement record remain essential.' Sparse projective measurements and dense weak measurements with the same I_loc produce different multi-point correlations: few sharp events versus many weak events, with different backaction and different accumulation of local information into a global constraint. Moreover, I_read is a maximum over input ensembles; the actual information a readout carries about the global label is suppressed by the local charge entropy within each sector, and this suppression can depend on (q,η) separately. The numerical support is finite-size Binder crossings at L=6–10 without error bars, extrapolation, or scaling collapse, and only the I_loc≃0.20 contour is tested in detail. Thus the organizing-variable claim rests on an unverified equivalence between local informational power and the full record–label statistics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies charge-learnability transitions in U(1)-symmetric monitored random circuits when the local measurements are probabilistic weak measurements, i.e., a measurement probability q and a measurement strength η are independently tunable. For a single readout, the authors correctly derive the binary asymmetric channel, its error probability ϵ(η), and the single-readout informational power I_read(η)=1−h2(ϵ(η)). They then define a local informational power I_loc(q,η)=q I_read(η) and conjecture that the finite-size learnability phase boundary is organized by contours of constant I_loc, thereby unifying probabilistic projective, deterministic weak, and probabilistic weak measurements. They introduce a label-sensitive cross-entropy diagnostic to distinguish unbiased, biased, and antibiased SEP decoders, and they compute an exact record–label mutual information I_true=I(Y;M) as a decoder-independent benchmark, comparing decoder accuracy to the Fano bound. The central numerical evidence consists of finite-size Binder crossings and variance peaks for L=6,8,10, with the exact mutual-information calculation performed only up to L=8.","tokens_in":12464,"tokens_out":3317,"duration_ms":37072,"significance":"If the constant-I_loc organizing principle holds, the paper would provide a genuinely useful single-quantity description of charge learnability across very different monitoring protocols, with potential experimental relevance for comparing projective and weak readouts. The exact record–label mutual information calculation is a valuable decoder-independent benchmark, and the cross-entropy variance diagnostic is a sensible methodological contribution for detecting mismatched decoders. The paper is also commendable for explicitly stating in Sec. V that the organizing criterion is conjectural and that space-time correlations remain essential. However, the central claim is supported only by finite-size numerics without scaling collapse, error bars, or an independent test of the I_loc hypothesis against the exact record–label information, so the significance currently rests on an unverified equivalence between a single-site readout quantity and the global statistics of the full measurement record.","major_comments":[{"comment":"The central claim that the transition boundary follows constant I_loc=q I_read(η) is posited rather than derived. The numerical test in Sec. III.A only follows the single contour I_loc≃0.20, using Binder crossings for L=6,8,10 with no error bars, no finite-size extrapolation, and no scaling collapse; the crossing locations are selected visually. This is insufficient to establish that I_loc, rather than some other combination of q and η, organizes the boundary, especially because Sec. V concedes that space-time correlations within the record remain essential. The authors should test the constant-I_loc prediction along at least one additional contour and, more decisively, compare exact I_true values at fixed I_loc for different (q,η) pairs. If I_true differs at fixed I_loc, the proposed one-parameter organization is not correct.","section":"Sec. II.B and Sec. III.A, Eq. (21)"},{"comment":"The exact record–label mutual information is computed only for L=6,8 with 10^4 trajectories, and the claimed collapse of SEP accuracy versus I_true is not quantified. No measure of residual η dependence at fixed I_true is given, so the statement that 'q and η primarily tune the information carried by the record' is not distinguished from a simple empirical correlation. The authors should quantify, for example, the spread of accuracy values at fixed I_true across different (q,η) points and system sizes, and provide error bars or confidence intervals for the Binder crossings of B_I used to place the I_loc≃0.16 boundary in Fig. 4.","section":"Sec. IV, Eq. (24) and Fig. 7"},{"comment":"The selection of the decoder-dependent transition scales I_loc≃0.17, 0.20, 0.25, and the decoder-independent scale I_loc≃0.16 appears to be made from the same finite-size data later used to draw the phase diagram in Fig. 4. This creates an in-sample fitting risk: the constant-I_loc curves are not predicted from independent data. A stronger test would be to fix I_loc, compute the predicted transition point on a new (q,η) cut that was not used to select the scale, and show that the Binder crossing or variance peak lands at the predicted location within quantified uncertainty. Without such an out-of-sample test, the agreement in Figs. 2 and 3 is anecdotal rather than a validation of the organizing principle.","section":"Sec. III.B, Figs. 5 and 6"}],"minor_comments":[{"comment":"The notation T=2L is stated as 'total circuit depth' but the following sentence equates it to 12, 16, and 20 alternating brickwork unitary layers for L=6,8,10. Clarify whether T denotes the number of unitary layers or the number of half-layers, and make the counting of 'layers' consistent with the figures.","section":"Eq. (22)"},{"comment":"The projective matching condition is written as q_s = q I_read(η). The subscript s on q_s is used only once; define it explicitly as the measurement probability in a projective protocol, or replace it with a clearer symbol such as q_proj.","section":"Sec. II.B, around Eq. (21)"},{"comment":"All finite-size diagnostics are plotted without error bars, and the vertical dashed lines marking transitions are described as 'selected' without a quantitative criterion. Adding error bars from trajectory subsampling and a defined crossing-extraction rule (e.g., linear interpolation between adjacent sizes) would make the comparisons reproducible.","section":"Figures 2–7"},{"comment":"The text uses 'mutual infor(Iloc = 0:16)' with apparent typographical corruption; also the figure legend uses unicode subscripts inconsistently with the body text. These should be cleaned to avoid confusion about which curve is which.","section":"Fig. 4 and text after Eq. (29)"},{"comment":"The concluding paragraph appropriately labels the criterion as conjectural. However, the abstract and Sec. I state the organizing principle without this caveat. Moderating the abstract language or explicitly adding 'finite-size evidence suggests' would better match the actual level of support.","section":"Sec. V"}],"recommendation":"major_revision","confidential_remarks":"The paper's channel-level calculations are correct and the cross-entropy diagnostic is a useful contribution. The main risk is that the abstract and summary overstate a conjecture that the authors themselves qualify in Sec. V. The manuscript can likely be brought to publishable form by (i) reframing the central claim as a supported conjecture, (ii) adding an out-of-sample constant-I_loc test and error bars or scaling collapse, and (iii) quantifying the I_true comparison in Fig. 7. I do not see a fatal internal inconsistency, but the evidence as presented is not yet strong enough for the advertised unifying principle."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I've read the paper through. It's a genuine extension of the learnability story to probabilistic weak measurements, and the authors are more honest in the conclusions than they are in the abstract. The clean part is the single-readout information theory: the channel in Eq. (14), the binary symmetric channel reduction, and the closed-form I_read(eta) are all correct. Defining I_loc = q I_read(eta) is a natural guess, and the constant-I_loc test is a legitimate thing to try. I also think the cross-entropy variance is a real addition: the antibiased decoder example, where the Binder ratio says q ~ 0.17 and the cross-entropy variance says q ~ 0.25, is a convincing demonstration that the label-blind diagnostic is misleading. And the exact record-label mutual information computation, though limited to L=6,8, is an honest decoder-independent benchmark; the Fano bound check and the rough collapse in Fig. 7 support the interpretation.\n\nThe soft spots are real, but not fatal. The central claim that I_loc organizes the phase boundary is a conjecture; there is no derivation connecting a single-site, single-readout quantity to the global record-label mutual information. The authors concede in Sec. V that space-time correlations remain essential, which sits in tension with the idea that the boundary is a simple function of I_loc. The numerics are also thin: L=6-10, no error bars, no scaling collapse or extrapolation, and crossings are read by eye. The representative contour I_loc ~ 0.20 is tested on several cuts, but the reference value itself is read from the projective-limit data, so the test is somewhat self-consistent rather than ab initio. And no code or data are released, which makes the finite-size claims hard to verify. None of these are fatal; all are addressable.\n\nThe abstract overstates the result: 'establish ... unifying principle' should be 'consistent with' or 'suggest.' The paper itself, in Sec. V, is appropriately cautious.\n\nThis paper is for people who work on charge sharpening, learnability transitions, and monitored-circuit diagnostics. It deserves a serious referee. I would send it to review, and ask for code/data, error bars, and a better finite-size analysis. The conjecture is worth taking seriously, and the cross-entropy diagnostic alone justifies a referee's time.","headline":"A well-scoped extension with an honest conjecture; the numerics are thinner than the abstract implies, but the cross-entropy diagnostic and exact-MI benchmark earn it a referee.","tokens_in":12899,"tokens_out":4324,"would_cite":true,"duration_ms":38594,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P15","81P45","68Q12"],"pacs":["03.65.Ta","03.67.Mn","89.70.Cf"],"model":"deepseek-v4-flash","headline":"Weak and projective measurements learn a conserved charge at the same local information rate","keywords":["charge learnability","monitored quantum circuits","weak measurements","informational power","U(1) symmetry","posterior entropy","mutual information","cross entropy"],"falsifier":"One concrete test: compute the finite-size Binder crossings for two protocols with identical I_loc but with one using high-q/low-η (e.g., q=0.8, η=0.314) and the other low-q/high-η (q=0.2, η=1); if their crossing points differ by more than the finite-size scatter, the single-parameter organization fails. A sharper falsifier for the I(Y;M) benchmark: evaluate Eq. (27) for L=6 and L=8 with exact propagation; if the Binder crossings of I(Y;M) do not converge to a constant I_loc value or if the variance peaks do not align with the crossings, the claimed decoder-independent scale of ~0.16 is an art","tokens_in":11968,"feed_emoji":"⚛️","tokens_out":1372,"duration_ms":16927,"temperature":0.7,"pith_summary":"This paper asks whether a machine decoding the measurement record of a U(1)-symmetric quantum circuit can learn a globally conserved charge when measurements are weak and probabilistic, rather than the idealized projective case. It claims that the transition from failure to success is organized by a single local quantity, the informational power I_loc = q·I_read(η), so different combinations of measurement probability q and measurement strength η lead to equivalent learning when they supply the same local information per measured site. It also introduces a label-sensitive diagnostic, the cross-entropy variance, to separate genuinely informative decoders from confidently wrong ones, and uses an exact record–label mutual information to establish a decoder-independent benchmark for this learnability limit.","feed_headline":"One information scale sets the charge-learning boundary","feed_subtitle":"Weak and projective measurements cross into learnability at the same local informational power, unifying their phase diagrams.","key_machinery":"The central object is the 'informational power of a local measurement,' defined as I_loc(q,η)=q·I_read(η), where I_read(η)=1−h2((1−η)²/[2(1+η²)]) is the mutual information between a single weak-measurement outcome and the local charge eigenvalue, maximized over input ensembles. This one number replaces the two-parameter (q,η) description of the measurement protocol and is claimed to be the organizing variable of the phase boundary. It is tested by comparing Binder-ratio crossings and cross-entropy variance peaks along constant-I_loc curves; the record–label mutual information I(Y;M) supplies the intrinsic ceiling.","core_discovery":"For a monitored circuit with conserved charge, the finite-size learnability boundary—the crossover from a record too weak to infer the label to one that is sufficient—is approximately organized by the local informational power I_loc(q,η) = q·I_read(η), where I_read(η) is the single-readout informational power of a weak measurement of strength η. Projective measurements (η=1) and weak measurements with the same I_loc are claimed to yield similar decoding behavior, so the transition is controlled by one combined scale rather than by measurement probability and strength separately. The paper further shows that cross-entropy variance, unlike posterior-entropy Binder ratios, correctly identifies","pith_inferences":["The proposed one-parameter organization likely extends to other conserved quantities (e.g., Z2 or free-fermion charges) where a similar single-site informational power can be defined; if the mechanism is generic, a family of 'learnability phase diagrams' indexed by I_loc would emerge.","Because I_loc is purely local, the framework implicitly assumes that cross-site, cross-time correlations within the record do not independently shift the boundary. One testable extension: vary the circuit's scrambling rate while holding I_loc fixed; any shift in the boundary would expose a correlation-dependent correction.","The exact mutual-information benchmark, being decoder-independent, suggests a natural operational meaning for 'learnability': a label is learnable from a record family exactly when I(Y;M) exceeds the Fano floor for target accuracy, independent of the classical decoder used.","The paper treats spatial structure coarsely through a one-dimensional resource; one could extend the decoder to higher dimensions or nonlocal measurement patterns to probe whether I_loc remains the controlling variable or whether geometry enters the effective scale."],"forward_implications":["If the claim holds, all local monitoring protocols—probabilistic projective, deterministic weak, probabilistic weak—can be placed on a single phase diagram parameterized by I_loc, simplifying the search for circuits that can or cannot reveal a conserved charge.","Decoder benchmarking gains a quantitative rule of thumb: a decoder that is confidently wrong (antibiased) will be mis-located by label-blind diagnostics, so cross-entropy variance should be the default finite-size transition probe whenever decoder fidelity is in question.","The exact record–label mutual information gives an intrinsic threshold that decoder performance cannot surpass, so discrepancies between a decoder's transition point and the I(Y;M) boundary directly measure extraction efficiency rather than record content.","The reported collapse of projective and weak data onto one curve of accuracy versus I(Y;M) suggests that the measurement record's total charge information—not its detailed readout history—is the controlling resource for learning.","Experiments that can tune mid-circuit readout strength can test the constant-I_loc predictions directly by comparing transition points at matched I_loc, without needing to vary both q and η independently."],"fun_headline_variants":["One informational scale sets the charge-learning boundary","Weak and projective measurements unify via local informational power","Charge learnability collapses onto a single local measure","Informational power of local measurement governs charge detectability","A single number predicts when records reveal conserved charge"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that a single-site, single-readout quantity (I_loc) controls the phase boundary, so that two protocols with the same I_loc produce statistically equivalent records for charge inference; the authors themselves note in Sec. V that space–time correlations within the measurement record remain essential, so this reduction is an assumption without an analytic derivation.","fun_headline_variants_meta":{"raw":{"variants":["One informational scale sets the charge-learning boundary","Weak and projective measurements unify via local informational power","Charge learnability collapses onto a single local measure","Informational power of local measurement governs charge detectability","A single number predicts when records reveal conserved charge"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000172,"raw_usage":{"total_tokens":1057,"prompt_tokens":633,"completion_tokens":424,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":377,"completion_tokens_details":{"reasoning_tokens":352}},"tokens_in":377,"tokens_out":424,"duration_ms":4865,"temperature":1.0,"reasoning_tokens":352,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T16:42:44.388340+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One concrete test: compute the finite-size Binder crossings for two protocols with identical I_loc but with one using high-q/low-η (e.g., q=0.8, η=0.314) and the other low-q/high-η (q=0.2, η=1); if their crossing points differ by more than the finite-size scatter, the single-parameter organization fails. A sharper falsifier for the I(Y;M) benchmark: evaluate Eq. (27) for L=6 and L=8 with exact propagation; if the Binder crossings of I(Y;M) do not converge to a constant I_loc value or if the variance peaks do not align with the crossings, the claimed decoder-independent scale of ~0.16 is an art","supporting_citations":[],"review_version":1}