Spectroscopic Binary Detection as Agent-Callable Tools: Detecting 40,000+ Main-Sequence Binary Candidates from SDSS DR19 APOGEE Spectra

arXiv:2608.10866 · astro-ph.SR, astro-ph.IM · Submitted 2026-08-11 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Astrophysics Radio. The week's best astrophysics papers, unpacked for curious ears.

Vera: Next we'll be talking about the paper "Spectroscopic Binary Detection as Agent-Callable Tools: Detecting 40,000+ Main-Sequence Binary Candidates from SDSS DR19 APOGEE Spectra".

Jocelyn: The paper was written by Serat M. Saad and Yuan-Sen Ting from The Ohio State University and Max Planck Institute for Astronomy.

Vera: Stay tuned as we take you through the paper and discuss its implications.

Paper discussion segment 1: Vera: We're back, and the paper on the table is "Spectroscopic Binary Detection as Agent-Callable Tools: Detecting forty thousand plus Main-Sequence Binary Candidates from SDSS DR19 APOGEE Spectra", by Serat M. Saad and Yuan-Sen Ting. The title is really two promises at once: there's a big catalog of binary stars, but there's also this phrase "agent-callable tools", which tells you that the detection wasn't run by a human with a script — it was run by a language-model agent calling a set of packaged tools.

Jocelyn: And that second part is the genuinely new thing. Saad and Ting took a well-known method by El-Badry and colleagues from two thousand eighteen and they repackaged all the unwritten decisions that make it work — the normalization convention, the pixel masking, the threshold calibration — into nine MCP tool servers plus a written "Skill" that tells the agent what to do and in what order. The paper calls that the procedural know-how that usually only lives in the authors' heads.

Vera: Right. And in "Spectroscopic Binary Detection as Agent-Callable Tools", the forty-one thousand four hundred sixty-six candidates are the proof that this packaging works, not just the end goal. Run over two hundred thirty-eight thousand two hundred five APOGEE dwarfs, the classifier flags seventeen point four percent of the sample as double-lined binary candidates, at an eight point one percent validated false-positive rate — that's about fifteen times more candidates than the previous catalog, which used the same underlying method.

Subrahmanyan: The title's "candidates" is doing careful work there. The authors are explicit that at that false-positive rate, close to forty percent of the flagged systems could actually be single stars, so they release the sample as a candidate list rather than a clean catalog, with Gaia cross-match and multi-epoch flags so users can select a purer subset. That honesty is one of the things I appreciate most about the paper.

Jocelyn: And the authors bring a specific pedigree to this. Serat Saad is at Ohio State, and Yuan-Sen Ting holds positions at both Ohio State and the Max Planck Institute for Astronomy — they've been pushing this agent-based approach to survey analysis. The paper even includes what they call a blind-agent test, where they remove one operating decision from the Skill and let a completely fresh agent try to rediscover it.

Vera: Exactly — and the result is beautifully subtle. When they removed the self-consistent normalization between the model and the data, the agent saw the single-star fit's chi-squared jump by a factor of several, and it recovered the shared normalization on its own. But when they removed the isochrone-tied luminosity weighting, the fit quality didn't change at all, only the recovered mass ratios shifted, so the agent had no internal signal to act on and the decision went unrecovered.

Subrahmanyan: So the title is making a claim about reproducibility: this method can now be handed to any agent, or any human, with the same deterministic core underneath. The catalog with its mass ratios, component temperatures, and eccentricity comparisons is essentially a byproduct of publishing the know-how itself.

Jocelyn: And they say as much in the abstract — they're releasing the tools, the Skill, and the full DR19 catalog with its multi-epoch supplement. Nineteen pages of methodology, a GitHub repository, and a list of forty-one thousand four hundred sixty-six double-lined binary candidates. This is one of those papers where the methods section is the actual discovery.

Paper discussion segment 2: Vera: Welcome back to the second half of the show. We are digging deeper into "Spectroscopic Binary Detection as Agent-Callable Tools: Detecting forty thousand plus Main-Sequence Binary Candidates from SDSS DR19 APOGEE Spectra". The headline number, forty-one thousand four hundred sixty-six binary candidates, needs a lot of context, because the paper says the eight point one percent control false-positive rate implies close to forty percent of them, roughly sixteen thousand stars, are probably single.

Jocelyn: That is a huge asterisk.

Vera: It is, but they handle it honestly: they call the whole thing a candidate list, and they release multi-epoch flags so users can select a purer sample. On the multi-epoch side, sixty-eight point five percent of the multiply-visited SB2 get velocity-confirmed, and they also add five hundred nineteen single-lined velocity variables and flag eight thousand nine hundred eighty-one systems as orbit-ready.

Jocelyn: The orbit-ready count is remarkable, since the DR13 catalog solved only sixty-four orbits. That is not just a bigger sample; it changes what you can ask. The per-visit velocities let them measure eccentricities for the close twins, which is genuinely new.

Subrahmanyan: And they find α twin minus α non-twin equals-zero point two four plus or minus zero point one six, so no significant difference. The injection tests show they would have recovered a +one difference, the size seen for wide twins from Gaia. So this is a genuine null result, not a lack of sensitivity.

Vera: Which suggests the eccentricity excess of twins is not imprinted at birth, but acquired as the binaries widen. That is a concrete piece of stellar formation physics from a method that was originally just a way to clean up parameter biases. And it only works because the multi-epoch data let them track the same system over time.

Subrahmanyan: I keep coming back to the agent experiment, though. The blind-agent test is the most unusual part of the paper, because removing the self-consistent normalization made the median single-star χ2 jump from two point one times ten to the fourth to six point two times ten to the fourth. That is a signal the agent could see and act on, and it did recover the decision.

Jocelyn: But removing the isochrone-tied luminosity weighting left no trace in the fit. The mass ratios shifted by up to zero point two, but without an external reference the agent could not recover the decision. That is exactly the line between procedural know-how and calibrated know-how. Some knowledge only shows up when you have a ground truth to compare against.

Subrahmanyan: And the chosen title, "Spectroscopic Binary Detection as Agent-Callable Tools", is really about that line. Seven of the nine servers and all eight Skill decisions transfer to a new survey unchanged; only the data server and the single-star model get replaced. That is the practical definition of what carries over and what has to be rebuilt.

Vera: The single-star model is the calibrated part, and its limits show up all over the catalog. There are eleven thousand four hundred twenty-four primaries sitting below the four thousand two hundred Kelvin floor, four point eight percent of the searched dwarfs, and they get fit at the floor, which raises the minimum detectable mass ratio for cool stars.

Jocelyn: So the completeness is reported over the (Teff, q) plane, not as a single number. The fifth-percentile recovered q climbs from about zero point six near six thousand Kelvin toward the reliability edge below four thousand seven hundred Kelvin. That is why the median mass ratio of zero point nine one, or zero point eight zero over the zero point two to zero point nine five range, is a recovery distribution, not the intrinsic one.

Subrahmanyan: And the false positives pile up at high q too, with the control stars at a median of zero point nine four. So the twin peak should be read with that contamination subtracted in mind. But the authors are right to keep the twins in, because they carry the signal of a real equal-mass population.

Vera: Exactly. They still keep the near-equal twins because the flagged ones have a median velocity separation of fourteen kilometers per second, and they show genuine line doubling. The Gaia cross-check supports the sample too, with fifty-seven percent of flagged systems at RUWE above one point four versus twenty percent of controls, and nineteen percent with non-single-star solutions.

Jocelyn: Independent anchors, because Gaia never sees the spectra. Next segment, we can talk about how this whole approach transfers to the optical surveys like WEAVE and 4MOST, and whether the Skill survives the change of wavelength.

Paper discussion segment 3: Vera: So this paper, "Spectroscopic Binary Detection as Agent-Callable Tools: Detecting forty thousand plus Main-Sequence Binary Candidates from SDSS DR19 APOGEE Spectra," is really about making the operating know-how of a published method transferable. They took El-Badry's two thousand eighteen decomposition and packaged it as nine MCP tool servers plus a written Skill, so an agent can re-apply it to a new data release without a human re-deriving the normalization, the pixel masks, the multi-start on mass ratio, all the things that usually live in an author's head. The Skill records each decision, the reason, and the failure mode it prevents, which is the part papers normally leave implicit.

Jocelyn: Right, and that know-how made a measurable difference, didn't it? They recovered seventy-four point four percent of the benchmark SB2 at an eight point one percent false-positive rate.

Vera: Exactly, and the open real-data classifier actually outperforms the EB18 reference at every matched false-positive rate, once the thresholds are recalibrated for its own single-star model. But the important part is that they publish the decisions themselves, not just the code, and they document which decisions are enforced by the servers and which are left to the Skill.

Subrahmanyan: I'm struck by their blind-agent test, which actually probes which parts of that know-how an agent can rediscover. When they removed the self-consistent normalization from the Skill, a fresh agent could see the single-star chi-squared jump by a factor of three to more than ten on the controls, and it re-derived the common normalization on its own. But when they took away the isochrone-tied luminosity weighting, the mass ratios shifted low by up to zero point two while the fit quality stayed identical, and the agent had no signal to act on. So the procedural layer is not uniform: some of it leaves a trace, and some of it only shows up against an external mass-ratio reference.

Jocelyn: That distinction seems like the key improvement in how we publish methods, not just for binaries but for any survey analysis.

Vera: Yes, and the paper is explicit that the single-star spectral model is the only survey-specific piece. They show that a network trained on public labels alone recovers only sixty-three percent of the benchmark at the loose thirteen point three percent operating point, while the curated EB18 network reaches seventy-five point six percent there. They trace the gap to the training labels, and fix it by refitting the labels with the network itself, making the model and labels self-consistent, which lifts the catalog classifier to seventy-four point four percent.

Subrahmanyan: The implications for other surveys are substantial. They list LAMOST, WEAVE, 4MOST, and Gaia RVS as places that share the composite-fitting logic but not the wavelength range. The Skill and seven of the nine servers carry over unchanged; you only rebuild the single-star model and swap the data server. They even propose automating that last step as a verify-gated loop with recovery on a labeled benchmark as the objective, which would remove the human from the per-survey path entirely.

Jocelyn: There's also a limitation that points to the next improvement, though: the triple census. They built a three-component extension for DR19 and it didn't work. The problem is that a faint third star changes the spectrum by a few percent while the single-star model trained on public labels misses real spectra at a comparable level, so they blame the training labels rather than the spectra.

Vera: Right, so a triple either gets flagged as an SB2 when two components dominate, or rejected at the acceptance gate when the third component's light dilutes the fit improvement below threshold. Better labels are the planned route to a triple census and to per-system orbits from those eight thousand nine hundred eighty-one orbit-ready systems with enough phase coverage. And releasing the Skill means that failure mode is documented instead of quietly buried in the code.

Subrahmanyan: And the eccentricity comparison gives a nice scientific payoff from the same machinery. With two thousand nine well-sampled systems between six and four hundred days, the close twins have an eccentricity index difference of minus zero point two four plus or minus zero point one six relative to matched non-twins, so there's no excess of the kind measured at wide separations. That supports the picture where the wide-twin eccentricity is acquired during widening, not at birth.

Jocelyn: So for "Spectroscopic Binary Detection as Agent-Callable Tools," the suggestion is that the future of astrophysics methods is writing down the why behind each operating decision, not just the equations, and letting agents run them and audit them. They're also honest that the catalog is a candidate list: at the control false-positive rate, close to forty percent of the flagged systems are probably single. And they make it easy for users who need purity rather than completeness to cut on the multi-epoch confirmation flags in the released supplement.

Vera: Yes, about sixteen thousand false singles out of the forty-one thousand, so it's not a finished census. But the release includes the tools, the Skill, the per-system mass ratios, component velocities, and the Gaia and multi-epoch supplements, which is the raw material for follow-up orbits and for testing the same approach on other surveys. That is a real improvement in how we share analysis methods.

Paper discussion segment 4: Vera: For those tuning in now, we're discussing "Spectroscopic Binary Detection as Agent-Callable Tools: Detecting forty thousand plus Main-Sequence Binary Candidates from SDSS DR19 APOGEE Spectra" by Serat Saad and Yuan-Sen Ting, and we're starting at the very first page. The abstract lays out the headline number right away: forty-one thousand four hundred sixty-six double-lined binary candidates found among two hundred thirty-eight thousand two hundred five main-sequence dwarfs, about fifteen times the two thousand six hundred forty-five identified in DR13, with a median mass ratio of zero point nine one toward near-equal-mass pairs.

Jocelyn: And it immediately flags the caveat — an eight point one percent control false-positive rate, which implies close to forty percent of those candidates could be single stars.

Vera: That's what makes this a candidate list rather than a pure catalog. The authors are refreshingly direct about it, and that honesty is part of the paper's philosophy, because the whole point is that you need to know exactly how a method behaves before you trust its output.

Subrahmanyan: The opening of the introduction gives the scientific motivation: most stars are not loners. Companion frequencies run from nearly one hundred percent for massive O and B stars down to about forty or fifty percent for Sun-like dwarfs, so any large spectroscopic survey is inevitably looking at thousands of unresolved pairs.

Jocelyn: And then the first page outlines the three ways to spot them — astrometric wobble, radial velocity changes, and the spectral signature of two sets of lines blended together. This paper belongs to that third category, but with a twist: it's not just looking for two separate peaks in a cross-correlation, it's asking whether a two-component model fits the spectrum better than a single star.

Vera: Right, that

Conclusion: Vera: So to pull it all together, this paper gives us the largest sample of double-lined spectroscopic binary candidates ever assembled from APOGEE, over forty-one thousand systems, and it does it by taking a method that already existed and packaging the know-how, not just the equations, into tools that an agent can run.

Jocelyn: Right, and that's the part that really stands out to me. They're publishing the operating decisions, the normalization conventions, the masking, the threshold calibration, as a written Skill alongside the code. It means someone else can come along with a different survey and actually reuse the method without having to rediscover all the tacit steps.

Vera: Exactly. And they show that this works, the benchmark recovery actually beats the original method at matched false-positive rates, though they're honest that the catalog is a candidate list, not a clean sample, because close to forty percent of those flagged systems could be single stars once you account for the control false-positive rate.

Jocelyn: But that's why the multi-epoch supplement is so valuable. They confirm nearly seventy percent of the multiply-visited candidates by tracking the velocity changes across visits, and they add five hundred nineteen single-lined variables that the combined spectra can't catch, plus almost nine thousand systems that are ready for orbit fitting.

Vera: The eccentricity result is also a nice payoff, because they find no significant difference between close twins and their non-twin counterparts, which is interesting given the eccentricity excess seen at wide separations. It suggests whatever makes wide twins eccentric is acquired later, during the widening, not at birth.

Jocelyn: And the blind-agent test is a lovely little experiment in its own right, showing that an agent can recover a removed operating decision when its absence leaves a measurable trace in the fit, but not when the effect is invisible without an external reference. That tells us something about what kind of knowledge can be transferred automatically and what still needs a human or a labeled benchmark.

Vera: It really is a paper about the future of how we do astronomy, not just what we found in the data. And so with that, we're wrapping up our look at "Spectroscopic Binary Detection as Agent-Callable Tools: Detecting forty thousand plus Main-Sequence Binary Candidates from SDSS DR19 APOGEE Spectra" from Serat Saad and Yuan-Sen Ting.

Jocelyn: Great paper to dig into, and we'll be back shortly with the next one. Thanks for listening.

Serat M. Saad, Yuan-Sen Ting

The Ohio State University · Max Planck Institute for Astronomy

astro-ph.SR, astro-ph.IM

Submitted: 2026-08-11

Updated: 2026-08-12

Comments: 23 pages, 16 figures, submitted to OJAp

Code: https://github.com/kareemelbadry/binspec

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 75/100

Key concepts

Spectroscopic binary: A binary star system detected through shifts or doubling of spectral lines caused by the orbital motion of the two stars. Double-lined binaries show two sets of lines, allowing measurement of both components' velocities and mass ratios.
Agent-callable tools: A set of packaged software tools (MCP servers) and a written 'Skill' that an AI language-model agent can call to perform a complex analysis. The Skill records procedural decisions and failure modes, making the method reproducible and transferable to new data.
Blind-agent test: An experiment where one operating decision is removed from the Skill and a fresh agent tries to rediscover it. If the decision leaves a trace in the data (like a chi-squared jump), the agent recovers it; if not, it remains hidden, revealing which knowledge is procedural versus calibrated.
False-positive rate: The fraction of flagged binary candidates that are actually single stars. The paper reports an 8.1% false-positive rate, implying about 40% of the 41,466 candidates may be single, so the sample is released as a candidate list with flags for users to select purer subsets.

Terminology

Benchmark — 43× in this paper · explained in 2 episodes
Calibration — 14× in this paper · explained in 7 episodes
Core — 12× in this paper · explained in 2 episodes
Metallicity — 8× in this paper · explained in 3 episodes
Trace — 8× in this paper · explained in 2 episodes
Ablation — 7× in this paper · explained in 1 episode
Model Context Protocol — 5× in this paper
Abundance — 4× in this paper · explained in 1 episode
Agent-callable tools — 4× in this paper · explained in 1 episode
Model Context Protocol (MCP) — 4× in this paper · explained in 8 episodes
Consistency — 3× in this paper · explained in 3 episodes
Fidelity — 3× in this paper · explained in 2 episodes
Alignment — 2× in this paper · explained in 1 episode
Distillation — 2× in this paper · explained in 2 episodes
External Validation — 2× in this paper · explained in 2 episodes
Accuracy — 1× in this paper · explained in 1 episode
European Space Agency — 1× in this paper
Failure Modes — 1× in this paper · explained in 3 episodes
Large Language Models (LLMs) — 1× in this paper · explained in 6 episodes
Self-consistency — 1× in this paper · explained in 2 episodes
Spectral Signature — 1× in this paper · explained in 2 episodes
Transformer — 1× in this paper · explained in 2 episodes
Transformer Architecture — 1× in this paper · explained in 2 episodes

Summary

arXiv:2608.10866v1 [astro-ph.SR] 11 Aug 2026

The paper packages the double-lined spectroscopic binary (SB2) decomposition method of El-Badry et al. (2018b) for reuse on APOGEE spectra, with executable operations as nine tool servers built on the Model Context Protocol (MCP) and operating decisions as a Skill. Run over 238,205 DR19 dwarfs under a fixed driver script, the classifier flags 41,466 SB2 candidates with median mass ratio q = 0.91, about fifteen times the 2,645 identified in DR13, though not at matched purity. The 8.1% control false-positive rate implies that close to 40% are single stars, so the sample is released as a candidate list. Refitting individual visits confirms 68.5% of the multiply-visited SB2 and adds 519 single-lined velocity variables and 8,981 orbit-ready systems. The eccentricities of the best-sampled binaries show no significant difference between close twins (q > 0.95) and matched non-twins, excluding an eccentricity excess of the kind measured at wide separations. An ablation of two operating decisions illustrates that a fresh agent recovers a removed decision only when its absence leaves a measurable trace in the fit.

The paper addresses the problem that unresolved binaries are common in spectroscopic surveys, and their blended light biases inferred stellar parameters. The authors note that "When a single-star model is fit to a composite spectrum, the inferred effective temperature, surface gravity, and chemical abundances are biased, most of all at intermediate mass ratio where the secondary contributes enough flux to change the lines without producing a second set of lines."

The work adopts the forward-modeling decomposition of El-Badry et al. (2018a,b; hereafter EB18), which builds a two-component composite from a single-star spectral model, fits it to the observed spectrum, and compares the result to the best single-star fit. The authors emphasize that the classical SB2 is the large-separation limit of a wider class of composite spectra.

A key motivation is publishing method know-how for agents: A method like the EB18 decomposition is communicated as a paper and, increasingly, as public code. Neither one passes on the operating experience that makes the method run correctly on a new data set. The paper uses Model Context Protocol (MCP) and Skill to publish operating know-how in a form an agent can re-apply to new data.

The authors use the APOGEE arm of SDSS DR19, part of the fifth-generation survey SDSS-V, recorded with the 2.5-m Sloan Telescope and the APOGEE spectrograph, reduced by the Astra 0.6.0 pipeline. Each combined spectrum is sampled on an 8575-pixel logarithmic-wavelength grid spanning 15100.8–16999.8 Å. The spectra are delivered in the rest frame.

The main-sequence dwarf selection from the DR19 the payne apogee star catalog requires:

Surface gravity log g between 4 and 5
Effective temperature Teff between 4000 and 7000 K
Signal-to-noise ratio above 60

This yields 238,205 dwarfs with usable combined spectra. The 11,424 primaries below the 4200 K floor of the single-star model (4.8% of the sample) are retained but fit at the floor.

Two labeled sets underpin calibration:

Benchmark: DR19 spectra of EB18-flagged SB2 that survive quality cuts (2,344 systems as positives) plus 7,866 matched non-flagged dwarfs as controls.
Injection-recovery completeness set: Synthetic secondaries added to real single-star spectra, so injected composites carry the same instrumental residuals, telluric remnants, and label imperfections the classifier must survive.

For external validation, the confirmed sample is cross-matched against the Kounkel et al. (2021) APOGEE cross-correlation catalog and Gaia DR3 astrometry and non-single-star solutions.

The classifier uses a five-label network of Ting et al. (2019), a three-layer perceptron mapping the label vector θ = (Teff, log g, [Fe/H], [Mg/H], vmacro) through two hidden layers of 300 units to a continuum-normalized spectrum. The network has a 4200 K temperature floor.

The training procedure involves a three-step self-consistency approach:

A first network is trained on 4,800 curated dwarfs with pipeline labels (the open reproduction network)
That network refits the labels of 7,179 dwarfs of the underlying control sample left unflagged as candidate binaries (the refit pool), split in half
The catalog network is trained on refit labels of one random half (3,589 stars); the other half (3,590 stars) provides held-out controls for false-positive rate measurement

The authors note: "The refit is anchored, in that the new labels are seeded from and stay tied to the pipeline labels rather than iterated freely, and this is what makes the (label, spectrum) training pairs self-consistent with the model family. Iterating the refit further degrades the classifier."

The binary composite follows EB18: two continuum-normalized single-star spectra are summed with a weight set by each component's band luminosity and shifted by the two component velocities:

fbin(λ) = (w1 f1(λ; v1) + w2 f2(λ; v2)) / (w1 + w2)

where the weight wi = Ri2 Bλ(Teff,i) combines each component's radius with the mean H-band surface brightness. The temperatures, radii, and weights follow from the primary and a single mass ratio q through a MIST isochrone at an assumed common age of 4 Gyr.

The detection statistic is the improvement in fit:

Δχ2 = χ2single − χ2binary
The improvement fraction fimp (Eq. B1 of El-Badry et al. 2018b)

Acceptance is defined by a sliding ladder with rung values recalibrated for the classifier (scaling both coordinates of the EB18 rungs by 0.90), calibrated at a fixed control false-positive-rate target.

The paper records eight load-bearing operating decisions:

Self-consistent normalization: The model and the data are normalized by the same operator, so that a continuum error enters χ2single and χ2binary identically and cancels in Δχ2. This is the single most consequential choice.
A five-label single-star model: A three-label model leaves residual line structure that the binary model can absorb, inflating the false-positive rate.
Pixel masking and a signal-to-noise limit: Persistently bad pixels are masked and per-pixel signal-to-noise is capped at 200 for APOGEE.
A mass-ratio multi-start: The fit is restarted from several trial mass ratios because Δχ2(q) is multimodal.
Threshold calibration at a fixed control false-positive rate: Calibrated against labeled real data rather than model-into-model spectra.
Completeness over the (q, Teff) plane: Reported as a function of mass ratio and primary temperature rather than as a single number.
A positive improvement-fraction requirement: Candidates whose Δχ2 concentrates on a few detector-edge pixels are rejected by requiring fimp to clear its rung.
Isochrone-tied luminosity weighting: The flux ratio is fixed from an isochrone rather than fitted from the mean continuum, because a continuum-based weight biases the recovered q low at the equal-mass end.

Nine MCP servers are exposed (Table 1): data, single star, binary model, isochrone, doppler, broadening, gaia sql, gaia xp, and vision. Seven of nine servers and all eight Skill decisions are reused unchanged across surveys; the data server is rewritten for a new survey's spectra and the single-star model is replaced per instrument.

The agent is a language model primed by the Skill, connected to servers through a plan–call–observe–decide loop implemented in LangGraph. A vision node reads rendered fits and Gaia evidence to flag detections that do not survive inspection, but The full-sample catalog run is deterministic and does not use it.

On the benchmark of 2,344 EB18-flagged SB2 and 7,866 matched single-star controls, at the adopted operating point with recalibrated rung values, the open real-data classifier recovers 74.4% of the benchmark SB2 at an 8.1% validated false-positive rate. At its loosest setting it reaches 82.6% at 12.3%. The EB18 reference network reaches 75.6% at its loosest 13.3% point and only about 55% at the catalog's 8.1% false-positive rate.

Neither classifier recovers every EB18-flagged SB2 because the benchmark positives were flagged on the DR13 spectra, whereas we test on the independent DR19 re-reduction of the same stars.

Run over 238,205 dwarfs, the classifier flags 41,466 SB2 (17.4% of the full sample). The median recovered mass ratio is q = 0.91, or q = 0.80 over the 0.2–0.95 interval where the decomposition is most sensitive, close to the 0.83 of EB18.

Independent Gaia corroboration shows:

77% of the 36,300 systems with parallax signal-to-noise > 5 lie above the single-star main sequence
57% have RUWE > 1.4 (against 20% of controls)
19% have non-single-star solutions (against 4% of controls)
17% appear as SB2 in the Kounkel et al. (2021) catalog

Regarding contamination: "Read at face value, the 8.1% held-out control false-positive rate applies to the single-star population... and implies of order 16,000 falsely flagged singles if the control rate translates directly to the full sample. Those land inside the catalog, where they would make up close to 40% of the 41,466 flagged systems." Chance alignments contribute ≲1% and hierarchical triples mis-fit as two components contribute 3%.

The recovered mass-ratio distribution increases toward equal masses, with a peak at near-equal-mass twins. The falsely flagged benchmark controls concentrate at high mass ratio (median q = 0.94), accounting for close to half of the q > 0.95 bin.
In metallicity, flagged SB2 track the full DR19 dwarf sample closely (median [Fe/H] = −0.09 against −0.11).
The velocity separation has a median v1 − v2 = 11 km s−1 with a tail to 35 km s−1 at the 90th percentile.
The fifth-percentile q climbs from ≈ 0.6 near 6000 K toward the reliability edge below 4700 K, reflecting the detectability floor of the single-star model.
Secondaries populate a cooler main sequence with median near 4,510 K against 5,460 K for primaries.

The multi-epoch fit was run on every catalog SB2 with visit data plus dwarfs with visit-to-visit velocity scatter σv > 1 km s−1, 50,187 unique candidates in all. Key results:

Of the 26,239 catalog SB2 with two or more epochs, the per-visit joint fit prefers the two-component model for 68.5%, with a median primary-velocity change of 16.7 km s−1 across visits.
The confirmation rate rises with epochs: 52.5% at two visits to 71.9% at three, 73.5% at four or five, and 82.0% at six to eight.
519 single-lined velocity variables (SB1 candidates) are found, selected to have Δvmax > 10 km s−1.
8,981 multi-epoch SB2 have three or more visits spanning more than 20 km s−1 in primary velocity, enough phase coverage to anchor a spectroscopic orbit (compared to 64 orbits solved in DR13).

For multi-epoch SB2 with eight or more visits, the authors refit every archived visit and read eccentricity from the shape of the two velocity curves. They keep 2,009 systems with median periods between 6 and 400 days, corresponding to separations of roughly 0.08 to 1.2 au.

The population is modeled as f(e) ∝ e α. Results:

αtwin = +0.15 and αnon-twin = +0.41
After injection-test residual removal: αtwin − αnon-twin = −0.24 ± 0.16
The interval reaches zero at 1.5 standard deviations, so the sample does not separate the two classes.

The authors conclude: Whatever makes the wide twins eccentric is therefore not present at these separations, which supports the picture in which the excess is acquired during the widening rather than at birth.

Two decisions were ablated:

Removing self-consistent normalization (decision 1): "The median single-star χ2 over the controls rises from 2.1×104 to 6.2×104 for a few-percent continuum mismatch and to 2.4 × 105 for a larger one... The control false-positive rate, by contrast, does not rise: it falls, from 12% to 3%." An agent given the high χ2 re-derived the common normalization.
Removing luminosity weighting (decision 8): "It biases the recovered mass ratio low, by up to 0.2 at the equal-mass end, but leaves the fit quality and the acceptance unchanged, so nothing internal to the fit marks the change. Without an external mass-ratio reference the agent has no signal to act on, and the decision goes unrecovered."

The DR19 catalog is about fifteen times larger in count against a twelve-fold growth of the searched dwarf sample from DR13 to DR19. The modest excess reflects the recalibrated classifier's higher benchmark recovery at the catalog operating point. The color–magnitude over-luminosity is a consistency check on a shared forward model, not an independent validation. The genuine external anchor is the Gaia enrichment.

Four inputs set limits:

Component temperatures and mass ratios come from a fixed 4 Gyr solar-scaled isochrone, carrying an age systematic
11,424 primaries at the 4200 K model floor are fit with labels the grid cannot reach
Systems with eight or more visits come from higher-cadence APOGEE fields, not a random draw
EB18 do not publish a control false-positive rate in a form matched to ours

The three-component (triple-lined) extension does not work on DR19 products: A faint third star changes the spectrum by a few percent, and the single-star model trained on the public DR19 labels misses real spectra at a comparable level, so we cannot separate the two star by star.

The open reproduction network recovers only 63.0% of benchmark SB2 at a common 13.3% control false-positive rate, versus 75.6% for the EB18 network. A distillation diagnostic locates the cause: When the same architecture is trained instead on noise-free, label-consistent spectra generated by the EB18 network itself, it recovers 74.3%, so the gap lives in the training target rather than the network.

Ablation results:

Kurucz synthetic models alone: 24% recovery
Real spectra with survey pipeline labels: 54–63%
Refitted labels (self-consistent): 69.5% at original EB18 thresholds
Recalibrated thresholds at 8.1% target: 74.4%

The single-star model is the only part of the system tied to a specific instrument... so it is the only part that does not transfer unchanged.

The paper identifies three categories:

Codified know-how: equations and acceptance thresholds, fully portable
Procedural know-how: order of operations and decisions, portable once written down (what the Skill does)
Calibrated know-how: survey-specific numbers, does not transfer and must be rebuilt

The paper's main results:

41,466 SB2 candidates (17.4% of searched sample), about fifteen times the largest previous APOGEE SB2 catalog
Validated 8.1% control false-positive rate with benchmark recovery exceeding the EB18 network at every matched false-positive rate
Mass-ratio distribution rising toward equal masses (median q = 0.91)
Independent Gaia corroboration (77% over-luminous, elevated RUWE, more non-single-star solutions)
Multi-epoch supplement confirming 68.5% of multiply-visited SB2, adding 519 SB1 candidates and 8,981 orbit-ready systems
No significant eccentricity difference between close twins and matched non-twins (−0.24 ± 0.16)

The catalog was produced without re-deriving the method, by publishing its operating know-how rather than only its equations, packaged as agent-callable MCP tool servers and a written Skill of the decisions that make it work.

The authors conclude: "as language-model agents begin to carry out analyses, we expect the operating know-how of a method, long left implicit in code and expert practice, to become a research product in its own right. Publishing a method for an agent to run will increasingly mean publishing the decisions that make it work, alongside the equations."

The MCP tool servers, written Skill, DR19 SB2 catalog with multi-epoch supplement, orbit posterior summaries, and benchmark star list are released at https://github.com/seratsaad/agent4binary. The catalog is produced by a deterministic science core reproducible under a fixed driver script. The language-model agent uses gemini-2.5-flash for benchmark and worked examples. The EB18 reference network weights and curated training set are not redistributable.

Improvements for AI systems

Based on this paper, here are specific improvements to AI systems and what the improved systems can do:

Improvement: Encode procedural know-how (operating decisions, order of operations, failure modes) as structured, executable tools and natural-language skills, not just code or equations.

Improved system: An AI agent can re-apply a complex scientific method (e.g., binary decomposition) to new datasets without human re-derivation, automatically handling normalization, threshold calibration, multi-start fitting, and quality controls.

Improvement: Ensure the model and data are normalized by the same operator so systematic continuum errors cancel in differential statistics (Δχ2).

Improved system: An AI that detects and corrects normalization mismatches automatically, reducing false positives in spectral fitting and improving robustness across instruments.

Improvement: Test which operating decisions leave measurable traces in fit diagnostics (e.g., χ2) versus those that are invisible without external references.

Improved system: An AI that audits its own pipeline, identifies when a removed step causes detectable degradation, and re-derives missing steps—while knowing when external validation is required (e.g., mass-ratio bias).

Improvement: Recalibrate acceptance thresholds against labeled real data (not synthetic-only) at a target false-positive rate, scaling both coordinates of detection statistics.

Improved system: An AI that self-calibrates detection thresholds per dataset, maintaining consistent purity across surveys and avoiding over- or under-detection.

Improvement: Report detection completeness as a function of mass ratio and primary temperature, not as a single number.

Improved system: An AI that provides uncertainty-aware catalogs with per-object detection limits, enabling reliable population statistics and bias corrections.

Improvement: Use visit-to-visit velocity variations to confirm candidates, classify single-lined vs. double-lined systems, and flag orbit-ready systems with sufficient phase coverage.

Improved system: An AI that prioritizes follow-up observations, automatically identifies which candidates need more epochs, and produces orbit-ready samples for dynamical studies.

Improvement: Inject synthetic secondaries into real single-star spectra (with real residuals, tellurics, label imperfections) to measure completeness under realistic noise.

Improved system: An AI that estimates detection limits and false-negative rates under real observational conditions, improving statistical inference from survey data.

Improvement: Cross-validate detections against independent data (Gaia astrometry, RUWE, non-single-star solutions, other catalogs) to quantify contamination.

Improved system: An AI that automatically cross-validates its outputs across surveys, providing confidence scores and flagging potentially false positives.

Improvement: Train the same architecture on noise-free, label-consistent synthetic spectra to isolate whether performance gaps come from network architecture or training targets.

Improved system: An AI that diagnoses its own model limitations, distinguishing between representational capacity and data-quality issues, guiding retraining strategies.

Improvement: Identify which pipeline components transfer unchanged across surveys (skill, servers) versus which must be rebuilt (single-star model, data server).

Improved system: An AI that automatically adapts to new instruments by replacing only the survey-specific modules while preserving the procedural know-how, reducing deployment time.

What the improved AI system can do overall:

Run a full spectroscopic binary detection pipeline on new surveys with minimal human intervention, producing calibrated candidate catalogs with known false-positive rates.
Self-audit its own decisions, recover missing procedural steps when detectable, and flag when external validation is needed.
Provide physically meaningful completeness and contamination estimates, enabling robust population studies.
Automatically cross-validate against astrometric and photometric surveys, producing multi-messenger-confirmed samples.
Transfer methods across instruments by separating portable know-how from survey-specific calibrations.

Abstract

Unresolved binaries are common in spectroscopic surveys, and their blended light biases the parameters inferred for them. Methods to detect them exist, but applying one to a new data release is limited mostly by operating know-how that is rarely written down. We package the double-lined spectroscopic binary decomposition of El-Badry et al. (2018b) for reuse on APOGEE spectra, with the executable operations as nine tool servers built on the Model Context Protocol (MCP) and the operating decisions as a Skill. Run over the 238,205 DR19 dwarfs under a fixed driver script, the classifier flags 41,466 SB2 candidates with median mass ratio q = 0.91, about fifteen times the 2,645 identified in DR13, though not at matched purity. The 8.1% control false-positive rate implies that close to 40% are single stars, so we release the sample as a candidate list. Refitting the individual visits confirms 68.5% of the multiply-visited SB2 and adds 519 single-lined velocity variables and 8,981 orbit-ready systems. We compare the eccentricities of the best-sampled binaries and find no significant difference between the close twins (q > 0.95) and matched non-twins, which excludes an eccentricity excess of the kind measured at wide separations. An ablation of two operating decisions illustrates that a fresh agent recovers a removed decision only when its absence leaves a measurable trace in the fit. We release the tools, the Skill, and the DR19 catalog with its multi-epoch supplement.