{"id":"1029f29c-79d8-477c-a699-12484fa6e0cb","arxiv_id":"2412.14844","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"GAME Filter, a new software tool, recovers astrometric microlensing parameters from simulated Gaia DR4 astrometry for lens masses around 1 to 20 solar masses and distances up to 6 kpc.","lead":"The authors built and tested a software filter, GAME Filter, that detects gravitational microlensing events using only Gaia astrometry, without photometric data. They show with simulated Gaia observations that the filter can recover lens properties for strong events, including for faint sources and events whose peak lies outside Gaia's observing window.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Mock-data realism is the load-bearing risk: no blending and DR3 visit counts in the DR4 simulations mean the reported recovery rates and mass/distance reach are upper bounds until tested with realistic epoch astrometry.","rationale":"The reader's weakest assumption correctly identifies the mock-data realism as the most load-bearing risk. My reading of the full manuscript confirms that the simulated observations use DR3 visit counts, a Gaussian noise model, and fbl = 1, all acknowledged by the authors in footnotes and Section 4.3. The central claim—that the filter will work on real DR4 data and characterise lenses in the stated mass/distance range—depends directly on whether these idealisations preserve the signal. Blending is the most dangerous because it reduces the amplitude of the astrometric shift rather than merely adding noise, and its effect scales with the already-weak signals targeted in the G=19, θE ≈ 2 mas regime. The paper does not provide any quantitative blending test, so the reported numbers are best treated as upper bounds until such a test is done. I also note that the thresholds on MUWEmin and Lopt are calibrated on the same datasets used to report success rates, which can only inflate the fractions; however, the primary concern remains the realism of the mock data. The paper deserves credit for releasing the code, stating limitations explicitly, and avoiding overclaiming the event-rate predictions. Since the reader already issued a 'conditional' verdict based on the same concern, no change to the verdict is needed; the proposed concrete test would settle whether the concern actually degrades the headline numbers.","tokens_in":21715,"tokens_out":6233,"duration_ms":50952,"concrete_test":"Regenerate the lens_G19_N281_DR4 mock dataset with astromet/jaxtromet, adding a blended unresolved source with flux ratio fbl = 0.7 (i.e., 70% of the light from the microlensed source), keeping all other parameters and the GAME Filter thresholds fixed; recompute P20 and P10 for events with θE ≥ 2 mas and compare to Table 3. If P20 drops by more than the Poisson uncertainty from the current value, the claim that parameters are recovered at G=19 for θE ≳ 2 mas is not blending-robust, and the abstract's mass/distance reach needs to be revised downward.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that GAME Filter will identify and characterize astrometric microlensing events in Gaia DR4 rests on the assumption that the mock DR4 observations in Sections 2.2–3 are representative of real DR4 epoch astrometry. Three idealizations weaken this: (i) footnote 6 states that the visit counts Nv are taken from DR3 because the scanninglaw package does not provide DR4, so the sampling cadence is not the actual DR4 one; (ii) the noise is a Gaussian scatter along the scan (Section 2.2) with no correlated or systematic errors; and (iii) Section 4.3 explicitly assumes fbl = 1, i.e., no blending from the lens or unrelated sources. The last point is the sharpest: in crowded fields, a blended background source dilutes the centroid shift, so the same input event produces a smaller observed signal. The paper's own Equation 11 shows how a luminous lens changes the measured parallax, but no quantitative experiment is run to show how recovery rates (Prec, P20, P10) degrade as fbl decreases or when unresolved neighbours are added. Because the reported 'characterises lenses 1–20 M_sun up to 6 kpc' is derived from the same idealized simulations, it is an upper bound. The thresholds on MUWEmin and Lopt are also calibrated on the same datasets (Section 2.3), which can only bias the reported rates optimistically. These are external-validity issues, not internal inconsistencies; the code is public and the caveats are transparently stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Jankovič et al. present the GAME Filter, a software pipeline that fits a single-star-plus-microlensing model to Gaia epoch astrometry (along-scan positions) and selects candidate events using a reduced chi-square statistic (MUWE), an L-BFGS-B convergence criterion (Lopt), and boundary checks. They generate mock Gaia DR4 datasets with astromet for three source magnitudes, three sky positions with different visit counts, and an extended t0 range, plus a binary-star dataset to assess false positives. For each 50,000-event dataset they report the fraction of recovered events (Prec) and the fractions with parameters within 10% and 20% of the truth (P10, P20), study parameter degeneracies, and construct a lens mass-distance sensitivity map. The paper concludes that strong events are recovered with about 60% efficiency at G=14, fainter sources down to G=19 can still be characterized when the Einstein radius exceeds about 2 mas, visit counts above about 90 have little impact, events with their peak outside the DR4 window remain partially recoverable, and binary contamination is low.","tokens_in":22019,"tokens_out":6186,"duration_ms":56583,"significance":"If the reported performance transfers to real Gaia DR4 epoch astrometry, this would be a valuable tool for the first systematic astrometry-only microlensing search and would enable lens mass and distance estimates for events that have no photometric signal. The claim that astrometry breaks the photometric microlensing-parallax degeneracy is supported by the corner plots, which show a single cluster for (pi_EE, pi_EN) rather than the two clusters typical of photometric fits. The paper's strengths include the large mock-data campaigns (50,000 events per dataset), the public release of the code, the explicit false-positive binary test, and a transparent list of caveats. The significance is moderated by the idealized nature of the mock observations: the recovery rates and the 1-20 M_sun / 6 kpc reach should be viewed as upper bounds until the filter is tested on more realistic DR4 epoch astrometry with blending and the actual DR4 scanning law.","major_comments":[{"comment":"The mock DR4 datasets use visit counts Nv obtained from the DR3 scanning law because scanninglaw does not provide DR4; this is stated in footnote 6. Since the recovery statistics in Table 3 and the sensitivity map in Figure 11 depend directly on the time sampling of the epoch astrometry, the reported rates may not be representative of the actual DR4 cadence. Please rerun the analysis with the best available DR4 scanning-law prediction, or alternatively demonstrate robustness by testing several plausible DR4 cadences (e.g., different starting epochs or updated scheduling simulations) and show that Prec, P20, and P10 change only mildly.","section":"Section 2.2, footnote 6; Table 3"},{"comment":"The simulations assume fbl = 1, i.e., no blending from a luminous lens or unresolved neighbours, and Section 4.3 states that this assumption sets the lower boundary for the Einstein radius and hence for the derived lens mass and distance. With fbl < 1 the centroid shift is diluted and, as Eq. (11) shows, the measured Gaia parallax becomes a luminosity-weighted mix of source and lens parallax. Because the central claim that GAME Filter characterizes lenses at 1-20 M_sun and distances up to about 6 kpc is computed from these fbl = 1 simulations only, the reported reach is an upper bound. I request quantitative experiments in which fbl is varied over [0,1] or unresolved blended neighbours are added, showing how Prec, P20, and P10 degrade; if such simulations are not feasible, the abstract and conclusions should explicitly state that the quoted mass and distance ranges assume negligible blending.","section":"Section 4.3, Eq. (11)"},{"comment":"The acceptance thresholds (0.9 < MUWEmin < 1.1, Lopt < 0.015, and Lthresh = 0.01) are selected from histograms of the same mock datasets that are then used to report recovery rates. This in-sample calibration can only bias the reported Prec, P20, and P10 optimistically. Please validate the thresholds on an independent subset of the mock data (e.g., a cross-validation split or a separately generated calibration set), or fix the thresholds a priori and then report the recovery rates on unseen data. The relation between the Lopt < 0.015 acceptance criterion and the Lthresh = 0.01 restart criterion should also be clarified, since the reader cannot tell whether these are two different thresholds or the same threshold quoted with different precision.","section":"Section 2.3 and Appendix A.2"}],"minor_comments":[{"comment":"The captions refer to a dataset named 'lens_G14_N209_extended', but Table 3 lists the extended dataset as 'lens_G14_N281_extended'; Section 3.4 also uses 'lens_G14_N281_extend'. Please make the dataset names consistent throughout.","section":"Figure 7 and Figure 8 captions"},{"comment":"The sentence 'The strongest microlensing signal occurs at u0 = sqrt(2) theta_E and t0' is dimensionally inconsistent because u0 is dimensionless while theta_E is an angle. The maximum of Eq. (9) occurs at u = sqrt(2), i.e., at angular separation sqrt(2) theta_E; please rephrase (e.g., 'at u0 = sqrt(2) for a closest approach that sits at the maximum, or at separation u(t) = sqrt(2)').","section":"Section 4.1, item (i)"},{"comment":"The MUWE denominator is N - 11; when N <= 11 this reduced chi-square is undefined or negative. Since short events with sparse sampling are part of the claimed domain of applicability, the paper should state how such cases are handled in the code or exclude them explicitly from the analysis.","section":"Equation (10)"},{"comment":"The abstract states lens masses from approximately 1 to 20 M_sun, while the first bullet of the Conclusions states 0.1 to 20 M_sun. Figure 11 shows that sensitivity is below 10% for ML < 2 M_sun beyond 4 kpc, so the broader range in the Conclusions is not supported by the same sensitivity criterion. Please harmonize the quoted ranges and define the sensitivity threshold used for the headline claim.","section":"Abstract vs. Conclusions, Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is methodologically sound within its own mock universe, and the code availability and transparent caveats are commendable. The main risk is external validity: the headline recovery rates and the mass/distance reach are derived from fbl = 1 simulations with DR3 visit counts and in-sample threshold calibration. These are fixable with additional simulations or by softening the claims, so I recommend major revision rather than rejection. I would be comfortable with acceptance once the calibration question and the blending/cadence robustness are addressed or the claims are explicitly bounded."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, well-scoped methods paper with public code and a large mock-data validation. The headline claims are real for the idealized data they simulate, but the mock realism is the load-bearing risk: DR3 visit counts, no blending, and thresholds tuned on the same datasets mean the quoted recovery rates and the 1-20 Msun / 6 kpc reach are upper bounds until tested on realistic DR4 epoch astrometry or real data.\n\nWhat is actually new: GAME Filter generalizes the astrometry-only detection idea from Chen et al. (black holes) to general dark lenses, and it ships as an open-source tool. The mock study is big (50,000 events per dataset), the parameter-recovery and degeneracy analysis is careful, and the binary false-positive check is a useful addition. The authors are transparent: they flag in footnote 6 that visit counts come from DR3, in Section 2.2 that the noise is Gaussian, and in Section 4.3 that fbl = 1 is assumed. That honesty earns real credit.\n\nSoft spots: the stress-test note is mostly right. Thresholds on MUWEmin and Lopt are calibrated from histograms of the same mock datasets used to report recovery rates, so those rates are optimistic. The paper does not test how recovery degrades with blending, even though blending is arguably the largest real-world confound in crowded fields. Equation 11 shows how a luminous lens changes the measured parallax, but no simulation explores fbl < 1. The binary false-positive test uses uniform orbital distributions, a rough approximation the authors acknowledge.\n\nNone of this sinks the paper. The central argument—that the filter can find and characterize strong astrometric microlensing events in Gaia DR4—holds for the idealized case. The gap is external validity, not internal consistency.\n\nWho benefits: microlensing and stellar-remnant people preparing for DR4; anyone who wants a public detection tool. I would send it to a serious referee, with the ask that the authors either add a blending-degradation study or prominently caveat that the quoted reach is a best-case, isolated-lens scenario. Conditional acceptance is right.","headline":"Solid, well-scoped methods paper with public code; the headline recovery numbers are best-case because the mock data lack blending and use DR3 visit counts.","tokens_in":22609,"tokens_out":2818,"would_cite":true,"duration_ms":20249,"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":"GAME Filter recovers microlensing parameters from Gaia astrometry alone.","keywords":["astrometric microlensing","Gaia DR4","GAME Filter","microlensing parallax","lens mass and distance","mock Gaia observations","binary false positives","Einstein radius"],"falsifier":"A decisive test is to run GAME Filter on real Gaia DR4 astrometric time series for the two known microlensing events OB110462 and GDR3-001 and compare the recovered $\\theta_E$ and $\\boldsymbol{\\pi}_E$ with the published values; agreement within the mock-data error bars would support the filter, and disagreement beyond those errors would falsify the central claim. A complementary calculation is to regenerate the mock observations with $f_{\\rm bl} < 1$ and the actual DR4 scanning law: if the recovery rates and parameter accuracies fall far below the reported values, the quoted sensitivity range does not transfer to real data.","tokens_in":21539,"feed_emoji":"🔭","tokens_out":12175,"duration_ms":96930,"temperature":0.7,"pith_summary":"This paper develops and tests the GAME Filter, a software tool that hunts for astrometric microlensing events in Gaia's astrometric time series using only positional measurements, with no photometric light curve. The claim is that for strong events the filter recovers the lensing parameters—Einstein radius $\\theta_E$, event timescale $t_E$, impact parameter $u_0$, time of closest approach $t_0$, and the two microlensing-parallax components—and that the astrometric signal breaks the $\\pi_{E,E}$–$\\pi_{E,N}$ degeneracy that limits photometric microlensing. The paper validates the filter on mock Gaia DR4 datasets of 50,000 events each, varying source brightness, sky position, visit count, and whether the event peak falls inside or outside the mission window, and it reports that lenses of roughly 1–20 $M_\\odot$ out to about 6 kpc can be characterized. This matters because Gaia DR4, expected in 2026, will release full astrometric time series, and astrometry-only detection would let dark or faint lenses—including stellar remnants—be found without waiting for a photometric brightening.","feed_headline":"Astrometry-only filter recovers Gaia microlensing events","feed_subtitle":"Mock tests show lens mass and distance can be recovered before Gaia DR4 time series arrive.","key_machinery":"The carrying mechanism is the astrometric microlensing centroid shift $\\delta\\theta_C = \\frac{u(t)}{u(t)^2 + 2}\\,\\theta_E$, where $u(t)$ is the lens-source separation measured in units of the Einstein radius $\\theta_E$; a passing lens makes the unresolved light centroid trace an ellipse over the source's straight proper-motion track. GAME Filter evaluates the predicted along-scan deviation for trial parameters and minimizes the renormalized microlensing unit weight error (MUWE), a $\\chi^2$-like goodness-of-fit, using the L-BFGS-B algorithm with multiple initial guesses for $u_0$, $\\pi_{E,E}$, and $\\pi_{E,N}$. An event is declared recovered only when the minimized MUWE lies between 0.9 and 1.1 and the L2 optimality error is small, so the fitted ellipse is the object whose shape and orientation carry the lens mass and distance.","core_discovery":"The paper's central discovery is that a microlensing event can be identified and its lens characterized from the astrometric signal alone: the centroid of the two unresolved images shifts by $\\delta\\theta_C = \\frac{u}{u^2+2}\\,\\theta_E$ with respect to the unlensed track, tracing an ellipse whose size, shape, and orientation encode the six microlensing parameters. Fitting the along-scan residuals $\\Delta x_{\\rm obs}$ to this ellipse, and accepting only fits whose normalized unit-weight error is close to 1 and whose optimality error is small, the filter recovers true parameters for 61% of the simulated strong events at $G_0 = 14$, and recovers the microlensing parallax components as a single cluster, not the two degenerate clusters seen in photometry. The authors present this as the basis for characterizing lenses of roughly 0.1–20 $M_\\odot$ out to about 7 kpc when the signal is strong, with sensitivity concentrated at nearby high-mass lenses.","pith_inferences":["A natural next step is to fit the blending fraction $f_{\\rm bl}$ as an extra parameter; until then, the quoted mass and distance estimates should be treated as lower bounds in fields where the lens or a neighbour contributes light.","Because the mock visit counts use the DR3 scanning law while DR4's schedule is not yet public, re-running the filter on the true DR4 epoch geometry is the immediate validation test once the time series arrive.","The false-positive estimate was made with uniformly sampled binary orbits; converting the 5% confusion rate into a DR4 contamination prediction would require a realistic Galactic binary population, which the paper defers to future work.","If the recovery rates survive real DR4 data, the same residual-ellipse filtering strategy transfers to astrometric time series from other surveys, with epoch sampling being the main difference."],"forward_implications":["For bright, well-sampled events at $G_0 = 14$ with 281 visits, about 61% pass the filter; 55% of astrometry-only events ($|u_0|>1$) are recovered, rising to 82% when a photometric signal is also present.","Astrometry resolves the microlensing-parallax degeneracy: the recovered $\\pi_{E,E}$ and $\\pi_{E,N}$ form one cluster, because the two components set the ellipse orientation rather than just the light-curve shape.","Fainter sources weaken the signal, but at $G_0 = 19$ the parameters for events with $\\theta_E \\gtrsim 2$ mas are still recovered with accuracy around 80%.","Events whose peak falls outside the 2014.5–2020 DR4 window are still recovered, with the recovery fraction dropping about 15 percentage points and parameter errors rising.","Binary systems contaminate at roughly 5% for $G_0 = 14$, and the misclassified binaries appear as weak, low-$\\theta_E$, high-$|u_0|$ microlensing-like signals."],"supporting_citations":[{"why":"Supplies the standard microlensing relations for the Einstein radius and event timescale used throughout the paper.","marker":"Paczynski 1986"},{"why":"Derives the astrometric centroid-shift formula that the filter fits to the Gaia along-scan residuals.","marker":"Dominik & Sahu 2000"},{"why":"Shows that the astrometric centroid traces an ellipse and predicts that Gaia will see a large number of astrometric events.","marker":"Belokurov & Evans 2002"},{"why":"Introduces the microlensing parallax vector formalism the paper uses to explain why astrometry breaks the photometric degeneracy.","marker":"Gould 2004"},{"why":"Provides the simulation code used to generate mock Gaia tracks for single sources, microlensing events, and binaries.","marker":"Penoyre et al. 2022"},{"why":"Supplies the Gaia DR3 photometric microlensing catalogue whose events motivate the DR4 astrometric search.","marker":"Wyrzykowski, Ł. et al. 2023"},{"why":"Identified the single confirmed Gaia candidate GDR3-001, which serves as a sensitivity benchmark in the paper.","marker":"Jabłońska, M. et al. 2022"},{"why":"Previous astrometry-only pipeline focused on black-hole lenses, which this paper generalizes to other dark lenses.","marker":"Chen et al. 2023"}],"fun_headline_variants":["Gaia astrometry alone spots microlensing events","GAME filter uncovers microlensing from Gaia astrometry","Astrometry-only microlensing detection with Gaia","Mock tests recover lens properties from Gaia shifts","Astrometric microlensing: filter recovers lens mass"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The simulated Gaia DR4 observations—Gaussian scatter, no light blending, and visit counts taken from DR3 instead of the not-yet-available DR4 schedule—faithfully represent the real astrometric time series the mission will deliver.","fun_headline_variants_meta":{"raw":{"variants":["Gaia astrometry alone spots microlensing events","GAME filter uncovers microlensing from Gaia astrometry","Astrometry-only microlensing detection with Gaia","Mock tests recover lens properties from Gaia shifts","Astrometric microlensing: filter recovers lens mass"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000214,"raw_usage":{"total_tokens":1489,"prompt_tokens":1070,"completion_tokens":419,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":686,"completion_tokens_details":{"reasoning_tokens":341}},"tokens_in":686,"tokens_out":419,"duration_ms":3292,"temperature":1.0,"reasoning_tokens":341,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:51:32.547997+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test is to run GAME Filter on real Gaia DR4 astrometric time series for the two known microlensing events OB110462 and GDR3-001 and compare the recovered $\\theta_E$ and $\\boldsymbol{\\pi}_E$ with the published values; agreement within the mock-data error bars would support the filter, and disagreement beyond those errors would falsify the central claim. A complementary calculation is to regenerate the mock observations with $f_{\\rm bl} < 1$ and the actual DR4 scanning law: if the recovery rates and parameter accuracies fall far below the reported values, the quoted sensitivity range does not transfer to real data.","supporting_citations":[{"cited_title":"1986, , 301, 503","cited_arxiv_id":null,"evidence_quote":"Supplies the standard microlensing relations for the Einstein radius and event timescale used throughout the paper."},{"cited_title":"& Sahu , K","cited_arxiv_id":null,"evidence_quote":"Derives the astrometric centroid-shift formula that the filter fits to the Gaia along-scan residuals."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that the astrometric centroid traces an ellipse and predicts that Gaia will see a large number of astrometric events."},{"cited_title":"2004, , 606, 319","cited_arxiv_id":null,"evidence_quote":"Introduces the microlensing parallax vector formalism the paper uses to explain why astrometry breaks the photometric degeneracy."},{"cited_title":", Kruszyńska, K","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaia DR3 photometric microlensing catalogue whose events motivate the DR4 astrometric search."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Previous astrometry-only pipeline focused on black-hole lenses, which this paper generalizes to other dark lenses."}],"review_version":1}