{"id":"5ae94a27-f58f-4f12-ba98-735367922e44","arxiv_id":"2607.14925","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DD4AO, a data-driven AO controller that optimizes an IIR filter against the measured disturbance spectrum, held stable closed-loop AO on-sky and improved Strehl from 28.8% to 33.9% over the integrator on Arcturus.","lead":"This paper reports on-sky tests of DD4AO, an adaptive-optics controller that learns the atmospheric disturbance spectrum and builds a custom filter to cancel it. On the bright star Arcturus it beat the standard integrator by five Strehl points, and on a faint binary it matched performance while saving deformable-mirror stroke.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported 5% Strehl gain may reflect a detuned integrator baseline; the paper's own admission that the integrator gain 'may have been set too low' (§5.1) undermines the comparative claim.","rationale":"I agree with the reader's assessment. The paper is a well-written validation report with honest limitations, but the primary claim is comparative, and the baseline tuning is explicitly questioned by the authors. The concrete test would settle whether the advantage is real by eliminating the detuning confound. The verdict should remain CONDITIONAL, pending such an audit or re-analysis. The reader's weakest_assumption is exactly the load-bearing concern: fairness of the comparison baseline. The paper's own admission in §5.1 that the integrator gain may have been set too low, combined with the single-representative-observation basis for the Strehl number, makes the 5% improvement unestablished. No other concern (e.g., vibration suppression or hour-long stability) is as central to the paper's main claim. Therefore, I do not change the CONDITIONAL verdict, but I emphasize the need for a quantitative baseline-tuning audit.","tokens_in":6121,"tokens_out":3162,"duration_ms":27557,"concrete_test":"Using the recorded pseudo-open-loop telemetry from the Arcturus session, perform an offline gain sweep of the integrator on a per-observation basis (e.g., optimize the integrator gain to minimize the residual PSD low-frequency content or to maximize the Strehl via maoppy) for each 15-s snapshot. Then recompute the mean and standard error of the Strehl for the best-tuned integrator and compare with the DD4AO values. If the tuned-integrator Strehl is within the per-observation uncertainty of the DD4AO Strehl, the reported 5% improvement is not a robust DD4AO advantage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is the 5% Strehl improvement (28.8%→33.9%) over the integrator on Arcturus, reported from a single representative observation without error bars. The paper's own text undermines the comparison: §5.1 states 'the integrator gain may have been set too low' and 'was not always optimal,' and §4 notes a hand-tuned scalar optical gain was applied to all controllers. If the integrator baseline was run below optimal gain, the low-frequency residual excess attributed to DD4AO's shaping could instead be the integrator's detuned response. The residual RMS margins across the five Arcturus observations are described as 'not always statistically significant,' and the Strehl difference is not accompanied by a distribution or uncertainty. Therefore, the magnitude of DD4AO's advantage over a properly tuned standard controller is unestablished. This is load-bearing because the paper's conclusion that DD4AO is 'a robust and adaptive control solution' for RISTRETTO/SAXO+ rests on that comparative margin.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents DD4AO, a frequency-domain data-driven controller for adaptive optics, implemented on the PAPYRUS on-sky testbed. The controller is designed via convex optimization using the measured disturbance periodogram and stability/input constraints, and the real-time pipeline allows instantaneous switching between DD4AO and standard integrator/OMGI controllers. The authors report stable closed-loop operation over hour-long exposures on Arcturus and HD137909, and claim a 5% Strehl ratio improvement over the integrator on Arcturus (28.8%→33.9% at 1310 nm), plus reduced DM stroke and vibration-peak suppression on the fainter binary.","tokens_in":6359,"tokens_out":3479,"duration_ms":38963,"significance":"If the comparative claims hold, this is an important milestone: it shows that a data-driven IIR controller can be deployed in a real-time AO loop, adapt every 20 s, and be switched against standard controllers in service. The work combines a real-time implementation with measured latency (0.04 ms), a convex formulation with stability guarantees, and a two-star on-sky validation campaign. These strengths make the paper a useful contribution to the AO control literature. However, the central quantitative claim — a 5% Strehl gain over the integrator — depends on the fairness of the comparison, and the paper itself provides evidence that the baseline integrator was not optimally tuned. The lack of error bars on the Strehl numbers and the admitted lack of statistical significance in the residual RMS margins mean that the magnitude of the DD4AO advantage over a properly tuned standard controller is not established. The circularity of validating on the same periodogram used for optimization also limits the strength of the robustness claims.","major_comments":[{"comment":"The headline 5% Strehl improvement (28.8%→33.9%) is quoted from a single representative observation with no uncertainty or repetition. The residual-RMS comparison across the five Arcturus observations is described as 'not always statistically significant', and the paper states that 'the integrator gain may have been set too low' (and that a hand-tuned scalar optical gain was applied to all controllers). These admissions directly undermine the comparative claim: the low-frequency excess in the integrator residual is exactly what a detuned integrator gain would produce. Please provide the Strehl distribution over observations, a properly tuned integrator baseline (e.g., a gain sweep), or a sensitivity analysis showing that the 5% gain is not an artifact of the integrator being run suboptimally.","section":"§5.1, §4"},{"comment":"The claim that DD4AO 'consistently used less deformable mirror stroke' on HD137909 is not quantified: Fig. 13 shows command standard deviations but no values, effect sizes, or statistical tests. In this low-SNR regime the performance is equal within uncertainty, so the stroke advantage is a secondary result that could depend on the specific gain settings of OMGI and the integrator. Please report the stroke values, their scatter, and a test of significance, and state the gain settings used.","section":"§5.2, Fig. 13"},{"comment":"Part of the validation is circular by construction: the controller is re-optimized every 20 s against the measured disturbance periodogram, and the residual PSD plots in Figs. 9 and 15 show rejection of the same measured vibration peaks that were fitted in Eq. (1). This demonstrates that the optimizer matches its fitting target, but it does not independently validate robustness or predictive power. To support the 'robust and adaptive' conclusion, please show an out-of-sample assessment — for example, residual PSDs from a later time window compared with the periodogram used for the optimization, or the closed-loop performance after the disturbance characteristics change.","section":"§2, §3, §5"}],"minor_comments":[{"comment":"Figure numbering inconsistency: the text says 'Figure 7 shows the command standard deviation in volts', but the caption of Figure 7 reads 'Total residual PSD'. The subsequent reference to Figures 8 and 9 as residual RMS per mode and total PSD appears inconsistent with the captions. Please renumber or correct the in-text references.","section":"§5.1"},{"comment":"The mode-grouping strategy is described qualitatively. Please specify how many groups were formed, how similar the periodograms need to be for grouping, and whether the optimization is per group or per representative mode.","section":"§3"},{"comment":"Reference [3] (Boccaletti et al. 2022) lacks publication details (journal, volume, pages, or arXiv identifier). Please complete the citation.","section":"References"},{"comment":"The switching protocol is stated as 15-second snapshots in a 1-minute cycle. It would help to report the number of switching cycles per observation and the guard time excluded for transients after each switch.","section":"§4"}],"recommendation":"major_revision","confidential_remarks":"The paper is promising as a demonstration of a real-time data-driven controller, but the central comparative claim is not yet supported. The authors' own statements about the integrator gain being too low and the hand-tuned optical-gain scalar suggest the baseline may have been unfairly handicapped. The missing error bars on Strehl and the circularity of the PSD-based validation are also serious concerns. I would encourage the editor to request a major revision that addresses these points with additional analysis rather than rejecting outright, since the underlying system and experiment are valuable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuine engineering validation of an adaptive frequency-domain controller (DD4AO) integrated into a real-time AO pipeline, not a new control-theory result. The paper deserves a serious referee, but the central quantitative claim — the 5% Strehl improvement on Arcturus — is weaker than the abstract implies.\n\nWhat's actually new and good: DD4AO is an IIR controller synthesized from measured disturbance PSDs using Karimi and Kammer's convex method, and the authors have made it run live: re-optimizing every 20 seconds, keeping the loop closed and stable across hour-long exposures, switching instantaneously between controllers for fair-ish comparison, and keeping hard-real-time latency at 0.04 ms. The on-sky results on two stars are real data, and the low stroke usage on the faint target is interesting. The paper is refreshingly candid about its limitations.\n\nSoft spots: the headline 5% Strehl gain (28.8% → 33.9%) comes from one representative observation with no error bars or distribution. The residual-RMS margins across five Arcturus observations are \"not always statistically significant,\" and the paper itself says the integrator gain \"may have been set too low\" and \"was not always optimal.\" That matters because the claim is comparative: if the baseline integrator was detuned, the margin measures that detuning as much as DD4AO's advantage. The hand-tuned optical gain applied to all controllers also weakens the comparison, though at least it was the same for all. And because the controller is re-optimized against the measured periodogram, the residual PSD plots show the fitted target, not an out-of-sample prediction — so the vibration suppression is a demonstration of adaptation, not generalization. No code or data are released.\n\nNone of this kills the paper. It is a solid integration-and-validation report, and the authors have flagged most of these issues themselves. But the abstract's 5% Strehl claim should carry a confidence interval and a baseline-tuning audit; otherwise it will mislead readers.\n\nAudience: AO control specialists and instrument teams (RISTRETTO, SAXO+). I'd bring it to a reading group for the pipeline and the switching methodology. I would not cite the Strehl number as established, but I might cite the pipeline description.\n\nRecommendation: send to peer review. A good referee can push for the statistical detail and data release. This is the kind of work that should be published, after the numbers are reported honestly.","headline":"An honest on-sky demo of continuously adaptive AO control, but the headline Strehl gain is one snapshot and the integrator baseline may have been running detuned.","tokens_in":6925,"tokens_out":2348,"would_cite":true,"duration_ms":23721,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DD4AO, a data-driven adaptive-optics controller, improved on-sky Strehl by 5 percentage points over the standard integrator at 1310 nm while keeping the loop stable for hour-long exposures.","keywords":["adaptive optics","data-driven control","predictive control","IIR filter","on-sky validation","RISTRETTO","PAPYRUS"],"falsifier":"On a test bench with a repeatable injected disturbance (e.g., a rotating phase screen or a calibrated vibrating mirror), tune the integrator and OMGI gains to their best values by gain sweep and re-measure Strehl and residual PSD; if the 5% advantage at 1310 nm shrinks to the noise or reverses, the claimed gain is an artifact of baseline detuning. Also run DD4AO with its 20-second re-optimization disabled to test whether the benefit survives without adaptation.","tokens_in":5995,"feed_emoji":"🔭","tokens_out":5684,"duration_ms":51623,"temperature":0.7,"pith_summary":"The paper reports on-sky validation of DD4AO, a data-driven adaptive-optics controller that builds an IIR filter from measured disturbance power spectra via convex optimization. The authors aim to show that such a controller can run in real time, remain stable for hour-long exposures while adapting continuously, and match or beat standard controllers. On the bright star Arcturus, DD4AO improved the Strehl ratio from 28.8% to 33.9% at 1310 nm compared with the integrator, with gains concentrated in low-order modes and low temporal frequencies. On the faint binary HD137909, DD4AO produced equal or better correction while using less deformable-mirror stroke and damping vibration peaks that remained in the other controllers' residuals. If these results hold, data-driven IIR control is a deployable solution for next-generation extreme adaptive optics.","feed_headline":"Adaptive-optics controller beats integrator by 5% Strehl","feed_subtitle":"Hour-long on-sky runs show the data-driven controller stayed stable while lifting Strehl from 28.8% to 33.9% at 1310 nm.","key_machinery":"The central object is the DD4AO IIR filter itself, synthesized directly from non-parametric frequency-response magnitudes by a convex Second-Order Cone Program. The algorithm takes the periodogram of the pseudo open-loop reconstruction as an estimate of the disturbance spectrum, then minimizes a mixed H2/H-infinity sensitivity objective — residual rejection weighted against high-frequency command amplification (the complementary sensitivity function with a sigmoid stroke penalty) — subject to stability constraints. Filters are recomputed every 20 seconds, and modes with similar periodograms are grouped to keep the optimization scalable; the resulting hard real-time filtering runs at 0.04 ms","core_discovery":"DD4AO is a frequency-domain, data-driven controller: it estimates the disturbance frequency magnitude from pseudo open-loop telemetry, then solves a convex optimization to synthesize an Infinite Impulse Response (IIR) filter that minimizes the H2 norm of the residual sensitivity while bounding high-frequency stroke through the complementary sensitivity function and a sigmoid weight, with stability constraints included. The paper's central empirical claim is that this controller held the loop closed and stable for hour-long exposures, re-optimizing every 20 seconds, and outperformed the standard integrator on Arcturus by 5 percentage points of Strehl at 1310 nm. On the fainter star, it used n","pith_inferences":["The paper's comparison baseline was deliberately detuned — a hand-tuned scalar gain was applied uniformly and the authors concede the integrator gain 'may have been set too low' — so the 5% Strehl margin likely overstates DD4AO's advantage over a fully optimized standard controller; the qualitative 'equal or better with less stroke' claim is on firmer ground.","A controlled experiment with a repeatable injected turbulence spectrum and individually tuned baseline gains would isolate how much of the gain comes from DD4AO's 20-second adaptation versus its static IIR shaping.","One testable prediction: disabling the periodic re-optimization should degrade performance on slowly evolving vibration frequencies while leaving the broadband component of the gain intact, which would separate the two mechanisms.","The same controller could be applied to other closed-loop systems where disturbance spectra are time-varying, such as tip-tilt platforms or beam stabilization, wherever a pseudo open-loop measurement is available."],"forward_implications":["If the reported gains are real, data-driven IIR control can replace fixed-gain integrators in extreme adaptive optics without sacrificing loop stability, even over hour-long exposures.","The 5-point Strehl gain at 1310 nm (28.8% to 33.9%) on a bright star, if reproducible, is a concrete sensitivity improvement for high-contrast instruments like RISTRETTO.","Lower DM stroke on faint targets implies reduced actuator effort and potentially longer deformable-mirror lifetime, plus more headroom for aggressive correction.","Vibration-peak suppression without explicit vibration modeling suggests DD4AO's data-driven shaping can attenuate telescope vibrations that standard controllers leave in residuals.","The mode-grouping strategy and 0.04 ms real-time latency indicate the approach can scale beyond the 195-mode test toward higher-order systems."],"fun_headline_variants":["Data-driven AO controller gains 5% Strehl on sky","Frequency-domain controller beats integrator by 5% Strehl","DD4AO: stable hour-long loop, 5% Strehl boost","New controller lifts Strehl 28.8% to 33.9% on sky","Adaptive optics controller: 5% Strehl gain in test"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the comparison is fair even though the standard controllers were run sub-optimally: a single hand-tuned scalar gain was applied to all controllers and the paper notes the integrator gain may have been set too low, so the size of DD4AO's advantage over properly tuned baselines is untested.","fun_headline_variants_meta":{"raw":{"variants":["Data-driven AO controller gains 5% Strehl on sky","Frequency-domain controller beats integrator by 5% Strehl","DD4AO: stable hour-long loop, 5% Strehl boost","New controller lifts Strehl 28.8% to 33.9% on sky","Adaptive optics controller: 5% Strehl gain in test"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000291,"raw_usage":{"total_tokens":1586,"prompt_tokens":840,"completion_tokens":746,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":650}},"tokens_in":584,"tokens_out":746,"duration_ms":7233,"temperature":1.0,"reasoning_tokens":650,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T00:39:53.284627+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a test bench with a repeatable injected disturbance (e.g., a rotating phase screen or a calibrated vibrating mirror), tune the integrator and OMGI gains to their best values by gain sweep and re-measure Strehl and residual PSD; if the 5% advantage at 1310 nm shrinks to the noise or reverses, the claimed gain is an artifact of baseline detuning. Also run DD4AO with its 20-second re-optimization disabled to test whether the benefit survives without adaptation.","supporting_citations":[],"review_version":1}