{"id":"9f36db8f-524d-43ea-a884-daf90d590bd1","arxiv_id":"2509.02203","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An ARIMA(3,1,5) model whitens DECAL sensor noise into residuals whose sign bits pass NIST tests, demonstrating feasibility of a low-rate true random number generator from this sensor.","lead":"This paper tests whether the electronic noise of a prototype particle-physics calorimeter sensor (DECAL) can be turned into random numbers. It fits a standard time-series model to the noise, converts the leftover unpredictable part into bits, and reports those bits pass the NIST randomness tests, though at a low rate.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Residual whiteness—the core assumption of the TRNG—is left unvalidated by the paper's own diagnostics; the 'truly random' claim is therefore unsupported.","rationale":"The reader's weakest_assumption correctly identifies residual whiteness as the linchpin. The paper's own evidence makes this suspect: the Ljung-Box df discrepancy (Sec. 3.3) and the neglected Box-Pierce correction (Sec. 2.5) are not resolved. My stress test reinforces the reader's view: the deterministic shuffle (Sec. 2.4) cannot manufacture entropy, so if residual whiteness fails, the NIST results on shuffled bits do not support 'truly random'. I find no additional independent objection beyond this. The paper is a feasibility study with honestly reported limitations (low bit rate, data not available); the concern is addressable by proper diagnostics and data release. Hence the CONDITIONAL verdict is appropriate; no change needed.","tokens_in":11117,"tokens_out":5230,"duration_ms":62881,"concrete_test":"Obtain the raw threshold-scan data from the authors and independently reproduce the differenced, mean-subtracted series. Fit ARIMA(3,1,5) using the same statsmodels routine, then compute the Ljung-Box Q at K=40 using the correct degrees of freedom K−p−q=32 and the Box-Pierce adjusted variance for residuals from a fitted ARIMA model. If the null of white noise is rejected at α=0.05, the residual assumption fails; if it passes, the residual-whiteness concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.4 generates bits from the sign of ARIMA residuals (Eq. 4), so the entire TRNG hinges on those residuals being i.i.d. Gaussian white noise. The paper's own Sec. 3.3 shows the Ljung-Box Q statistic follows χ²(K−q) instead of the theoretical χ²(K−q−p) (with p=3, q=5, K=40, i.e., df=35 vs. df=32), a standard symptom of an over-specified AR order or remaining structure in the residuals; Sec. 2.5 additionally states the Box-Pierce correction for parameter-estimation effects is neglected, and Sec. 3.3 explicitly calls the discrepancy 'outside the scope of this work.' Because the shuffle in Sec. 2.4 is a deterministic, data-dependent permutation, it cannot increase the entropy of the residual stream; if the residuals are not truly white, a deterministic post-processor may hide residual dependence from NIST tests without creating physical randomness. The abstract and Sec. 4 conclude 'truly random numbers' from NIST passes alone, which Sec. 2.5 itself notes is not a proof of randomness. Without raw data or code, the residual-whiteness claim is not independently verifiable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a TRNG based on DECAL sensor noise. A single pixel's threshold-scan Gaussian means are modeled as ARIMA(3,1,5); the sign of the estimated residuals gives bits, which are then shuffled by a deterministic algorithm. The authors report that the NIST SP 800-22 suite and additional diffusion tests are passed and estimate a bit rate of 5.4 bps. They conclude that the sensor noise can generate truly random numbers.","tokens_in":11448,"tokens_out":6400,"duration_ms":73212,"significance":"The application of time-series analysis to a particle-physics sensor is original and, if the residual-whiteness premise were established, would be a useful feasibility demonstration. The paper is transparent about the slow bit rate and mentions several limitations. However, the conclusion is stronger than the evidence: the residual diagnostics contain an acknowledged degrees-of-freedom discrepancy, the Box-Pierce correction is ignored, the shuffle is a deterministic permutation that cannot add entropy, and no entropy estimate is provided. These are load-bearing gaps.","major_comments":[{"comment":"The Ljung-Box result is the main evidence for residual whiteness. The observation that Q follows χ²(K−q) rather than χ²(K−q−p) contradicts the null distribution used for estimated ARIMA residuals. The paper calls this \"likely suboptimal choice\" and \"outside the scope\", but this is exactly the assumption on which Eq. (4) and bit generation rest. Please report the Q values, p-values, and the result with correct df and with the Box-Pierce correction (Sec. 2.5). Without this, the residuals cannot be claimed white.","section":"Sec. 3.3, Eq. (5)"},{"comment":"Even if the shuffle is deterministic and efficient, it is a permutation of the input bits and cannot increase entropy. If the residual bits contain bias or dependence, the shuffle may hide them from the NIST tests. The paper should test the unshuffled sign-of-residual stream, or at least show that the shuffle preserves the entropy of an i.i.d. source and quantify the min-entropy/entropy per bit (e.g., NIST SP 800-90B). The phrase \"truly random numbers\" is not justified by passing statistical tests.","section":"Sec. 2.4, shuffle algorithm"},{"comment":"The residuals are in-sample residuals from a model fitted to the same time series. Such residuals are not an independent sample of the physical innovation process, so NIST tests on them partly validate the fitting algorithm. An out-of-sample procedure (fit on a training interval, generate bits on a held-out interval, re-fit periodically and report parameter stability) would provide much stronger evidence. The current conclusion in Sec. 4 (\"truly random numbers\") overstates what the design can establish.","section":"Secs. 2.3-2.4 and Sec. 4"},{"comment":"The paper does not provide the raw data, processed time series, or code. Since the results depend on a specific pixel, tuning voltage, warm-up cutoff, ARIMA orders, and shuffle parameters, reproducibility is limited. Please include these materials or a detailed reproducibility description, including results for the other pixels claimed to behave similarly.","section":"Data availability"}],"minor_comments":[{"comment":"In Eq. (3), γjt should be γj; the subscript t is used both as index and in the expectation limit, which is confusing.","section":"Eq. (2)-(3)"},{"comment":"The algorithm says \"Skip the next Nshuffle segments\" and then \"when the Nshuffle-th segment is reached\", which is inconsistent. Clarify the indexing convention.","section":"Sec. 2.4"},{"comment":"The symbol q is used both for the MA order in Eq. (1) and for (n−2)/2 in the cumulative periodogram bounds. Rename one to avoid ambiguity.","section":"Sec. 2.5"},{"comment":"The residual Q-Q plot and histogram are visual; a formal normality test (e.g., Shapiro-Wilk) would be useful, although sign bits only require zero median.","section":"Sec. 3.3"},{"comment":"The caption states that \"p-value indicates the probability that the underlying p-values follow the uniform distribution\"; this is a misstatement. Revise to describe the P-value of the uniformity test used by NIST.","section":"Table 1"},{"comment":"Please clarify how \"repeat the scan 2000 times\" yields the 100000 scans mentioned in Sec. 3.4 when computing the bit rate.","section":"Sec. 2.2 / Sec. 3.4"}],"recommendation":"major_revision","confidential_remarks":"I would not reject on scope, but the \"truly random\" claim needs to be softened unless entropy is quantified. The manuscript does not provide data or code, which is a growing expectation. If the residual diagnostics cannot be fixed, the paper could still be published as a limited feasibility study with revised claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read this as an honest, clearly-written feasibility study, and I largely agree with the reader's CONDITIONAL verdict. What's new here is not the method—ARIMA whitening and NIST testing are standard—but the application to DECAL sensor noise. That is a legitimate extension, and the paper does it carefully: it reports the 5.4 bps rate without spin, it flags the Ljung-Box discrepancy, and it acknowledges that NIST passes are not proof of randomness. The residual diagnostics, diffusion tests, and NIST results are all presented clearly. Credit where due: this is the first ARIMA-based randomness characterization of DECAL noise, and the bits do pass all applicable tests.\n\nThe soft spot is real and load-bearing. The paper's own Fig. 10 shows the Ljung-Box Q statistic follows χ²(K−q) rather than χ²(K−q−p) with p=3, q=5. That is exactly the symptom you'd expect if the ARIMA model is over-specified or the residuals retain structure. The paper says this is 'outside the scope of this work' and then concludes the residuals are white noise 'to the degree of accuracy required.' That is a circular move: the bits are the sign of those residuals, so if the residuals are not truly white, the deterministic shuffle may be hiding dependence rather than removing it. The shuffle cannot add entropy. The paper also states it neglects the Box-Pierce correction, which is minor, but the degrees-of-freedom issue is not minor.\n\nOn the bigger claim: passing NIST on post-processed bits does not establish that the source has enough entropy to be called 'truly random.' The paper itself says this in Sec. 2.5, so the abstract and conclusion overstate. That said, this is an addressable issue, not a fundamental flaw. The measurement looks reproducible in principle, though raw data/code would help.\n\nWho is this for? People working on TRNGs from detector noise, and the DECAL collaboration itself. It deserves a serious referee, but only with the expectation of revision: either fix the Ljung-Box degrees-of-freedom analysis or weaken the conclusion to 'passes statistical tests' and avoid the phrase 'truly random.' I'd send it to peer review, and I'd tell the authors to address the residual-whiteness question head-on rather than deferring it.","headline":"A clearly-written feasibility study whose 'truly random' conclusion outruns the diagnostics; the Ljung-Box degrees-of-freedom discrepancy deserves a fix before publication.","tokens_in":11913,"tokens_out":1554,"would_cite":false,"duration_ms":20126,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A particle-detector pixel's noise, whitened by an ARIMA model, yields bits that pass the NIST randomness suite.","keywords":["true random number generation","DECAL sensor","sensor noise","ARIMA time series","white noise","NIST SP 800-22","random bit extraction","particle detector calorimetry"],"falsifier":"Re-run the pipeline with the shuffle removed and with Ljung-Box evaluated at the correct K-q-p degrees of freedom; if the un-shuffled bits or the residuals fail the randomness tests, the claimed sensor entropy is not the sole source of randomness.","tokens_in":11012,"feed_emoji":"🎲","tokens_out":6797,"duration_ms":68215,"temperature":0.7,"pith_summary":"Using the DECAL calorimeter sensor as a test case, the paper claims that a physical pixel's intrinsic noise can be converted into truly random numbers. The authors take threshold-scan data from one pixel, fit a Gaussian to each scan, and model the sequence of Gaussian means as an ARIMA(3,1,5) time series. They then take the sign of each model residual as a bit and apply a deterministic shuffle. The resulting bit stream passes the NIST SP 800-22 suite, the Ljung-Box test, the cumulative periodogram, and random-walk mean-squared-displacement checks. The rate is about 5.4 bits per second in the current setup, so the result is a feasibility proof rather than a high-speed TRNG.","feed_headline":"Sensor noise passes NIST randomness tests at 5.4 bits/s","feed_subtitle":"ARIMA whitening turns one calorimeter pixel's noise into bits that pass all applicable randomness tests.","key_machinery":"The load-bearing object is the ARIMA(3,1,5) residual sequence: the fitted model is supposed to strip the sensor's systematic temporal structure, leaving only white noise. The sign of each residual supplies one bit. A segment-based shuffle, whose parameters are derived from the bit stream itself, is then applied to break any small residual correlations. The Box-Jenkins workflow—ACF/PACF inspection, differencing, AIC/BIC order selection, residual diagnostics—is what justifies calling the residuals noise.","core_discovery":"The central discovery is that the stochastic component of DECAL sensor noise, after ARIMA whitening, behaves as Gaussian white noise whose sign bits pass standard randomness tests. Concretely: discard the first 10,000 scans for warm-up; fit a Gaussian to each threshold-scan count distribution via binned maximum likelihood; take the mean values as a time series; fit an ARIMA(3,1,5) model selected by AIC/BIC; compute the residuals; take the sign of each residual as a bit; apply a segment-based shuffle that swaps bits using patterns in the bit stream itself. The resulting stream passes all applicable NIST tests, and the residuals pass Ljung-Box, periodogram, and random-walk/MSD checks. The auth","pith_inferences":["The paper notes that the Ljung-Box statistic follows K-q rather than K-q-p degrees of freedom and that the Box-Pierce correction is neglected; this leaves open the possibility that some apparent whiteness comes from the fitted model and the shuffle, not the sensor itself. A direct comparison of the pipeline with and without the shuffle would separate those contributions.","Only one pixel is analyzed; if the noise streams of the 64 strips are independent, parallel readout could multiply the 5.4 bits per second rate by roughly the number of independent pixels, turning the feasibility result into a usable entropy source.","Passing statistical tests establishes local randomness, not physical nondeterminism; for cryptographic use, one would need an entropy model of the sensor and a conservative bias-free extraction rule rather than relying on test-passing alone."],"forward_implications":["If the claim holds, DECAL and similar pixel sensors can serve as physical entropy sources for random number generation.","The ARIMA residual stream is a statistical characterization of the sensor noise itself, which can inform calibration of future digital calorimetry experiments.","The bit-generation step is simple and online, so it can run continuously during normal sensor operation.","At 5.4 bits per second the current configuration is not a production TRNG; the paper's own conclusion is that optimizing the data acquisition chain is the path to practical rates."],"supporting_citations":[{"why":"Supplies the ARIMA model and Box-Jenkins workflow used to whiten the sensor noise.","marker":"[17]"},{"why":"Defines ACF/PACF tools used to identify AR and MA orders.","marker":"[18]"},{"why":"Earlier DECAL study showing threshold scans yield Gaussian-like count distributions.","marker":"[23]"},{"why":"Gives the binned maximum likelihood estimation used to extract each scan's mean.","marker":"[25]"},{"why":"Provides the innovations maximum likelihood algorithm for estimating ARIMA parameters.","marker":"[28]"},{"why":"The NIST SP 800-22 suite whose successful results ground the randomness claim.","marker":"[30]"},{"why":"Box-Pierce residual-autocorrelation correction that the paper explicitly neglects.","marker":"[31]"},{"why":"Defines the Ljung-Box test statistic used to check residuals for whiteness.","marker":"[32]"},{"why":"Documents the sensor warm-up and stabilization that justifies discarding the first 10,000 scans.","marker":"[21]"}],"fun_headline_variants":["ARIMA-whitened sensor noise passes NIST randomness tests","Sensor noise yields NIST-valid random bits after ARIMA whitening","Low-speed sensor noise still passes NIST randomness tests","Calorimeter noise becomes random bits via ARIMA whitening","Slow but sound: sensor noise passes NIST randomness tests"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The ARIMA residuals are genuine white noise—zero mean, constant variance, and no autocorrelation—so their signs are unbiased independent coin flips.","fun_headline_variants_meta":{"raw":{"variants":["ARIMA-whitened sensor noise passes NIST randomness tests","Sensor noise yields NIST-valid random bits after ARIMA whitening","Low-speed sensor noise still passes NIST randomness tests","Calorimeter noise becomes random bits via ARIMA whitening","Slow but sound: sensor noise passes NIST randomness tests"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001079,"raw_usage":{"total_tokens":4358,"prompt_tokens":755,"completion_tokens":3603,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":3517}},"tokens_in":499,"tokens_out":3603,"duration_ms":27823,"temperature":1.0,"reasoning_tokens":3517,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:45:28.553483+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the pipeline with the shuffle removed and with Ljung-Box evaluated at the correct K-q-p degrees of freedom; if the un-shuffled bits or the residuals fail the randomness tests, the claimed sensor entropy is not the sole source of randomness.","supporting_citations":[],"review_version":1}