Photometric classification and out-of-taxonomy detection of ZTF transients with FLARE

arXiv:2610.30656 · astro-ph.IM, astro-ph.HE · Submitted 2026-10-22 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Astrophysics Radio. Generated commentary on the latest astrophysics papers.

Vera: We're looking at the paper: "Photometric classification and out-of-taxonomy detection of ZTF transients with FLARE".

Jocelyn: Here's the gist of the paper. The gist The flare module, which combines photometric classification, anomaly ranking, and learning from confirmed candidates, provides a practical sequence for transient discovery by moving from blind detection to human review to supervised classification How it works The core of the approach is the AppleCiDEr framework's flare module, which is built for the boom alert broker and whose spectroscopic module, astraNet, provides classifications and anomaly scores for spectra.

Vera: This module uses ZTF photometry and public catalogues without spectra or image cutouts, evaluating it on 10 497 spectroscopically confirmed transients and 53 objects from families outside its six-class taxonomy.

Jocelyn: The workflow focuses on three aspects: (i) the information that the external context adds to the light curve, (ii) the recovery of out-of-taxonomy objects by blind ranking, active review and learning from first labels, and (iii) the reliability of both layers under missing context, short observation horizons and transfer to Fritz.

Vera: Classification Granularity The authors evaluate the pipeline at three levels of granularity to assess its performance.

Title and authors: Vera: So we're looking at this paper, "Photometric classification and out-of-taxonomy detection of ZTF transients with FLARE," and it’s about creating a practical sequence for finding new things in transient surveys.

Jocelyn: It seems to take you from a blind detection alert all the way through human review to actually getting a supervised classification, which is really useful for how we operate today.

Subrahmanyan: It’s interesting because it’s building on existing tools that are already used in the alert brokers and spectroscopic modules, like AstraNet.

Vera: Exactly. The core idea here is using ZTF photometry and public catalogues without needing spectra or image cutouts to evaluate things on ten thousand four hundred ninety-seven spectroscopically confirmed transients and fifty-three objects from families outside its six-class taxonomy <ref:2610.30656#pg1>.

Jocelyn: What I find important is how they focus on three main aspects: the context information that external data adds to the light curve, recovering those out-of-taxonomy objects through blind ranking and learning from first labels, and checking the reliability of both layers when context is missing or observation times are short.

Subrahmanyan: That part about recovery through blind ranking is key for finding things we haven't seen before in this specific survey.

Vera: Right. And they look at classification granularity at three different levels—splitting core-collapse supernovae into Type II and stripped-envelope supernovae, or merging superluminous supernovae into the core-collapse class.

Jocelyn: The way they handle those classification metrics is telling; for instance, splitting things lowers the macro F1 by seven points, and merged objects improve every aggregate metric because they are separated from Type II supernovae mainly by their absolute magnitude.

Subrahmanyan: That makes sense cosmologically; separating them by luminosity often separates them by physical process in a way that matters for our models.

Vera: Then they talk about host-galaxy context, incorporating four blocks: Pan-STARRS host photometry and angular offset, a photometric redshift from Legacy Surveys, the Gaia and WISE point-source context, and a Pan-STARRS pre-outburst counterpart.

Jocelyn: The host photometry seems to give most of the gain in classification accuracy; it moves the macro F1 from zero point seven two up to zero point eight three when all four blocks are added.

Subrahmanyan: So incorporating that photometric redshift gives you a pseudo-absolute magnitude, M pseudo equals m peak minus five log ten d L of z photo divided by ten parsecs, which is a way to estimate distance based on the redshift information they have.

Vera: And the point-source block helps with AGN through the W1 minus W2 color and having a Gaia counterpart, while that pre-outburst counterpart separates flares in a nucleus from supernovae near one and marks CVs that Gaia might miss.

Title and authors: Jocelyn: That last part about marking CVs is interesting because it points to objects that Gaia might not even register as point sources in the first place.

Subrahmanyan: For me, that’s important for structure formation studies; if we can separate nuclear flares from nearby supernovae this way, it helps constrain the environment around those events.

Vera: Moving to anomaly detection, they use luna1, an out-of-distribution scoring package with fourteen families of scorers on top of the trained classifier.

Jocelyn: They combine those scores either blindly as the mean percentile rank across all scorers or supervised when confirmed examples are available.

Subrahmanyan: Blind detection recovers a few out-of-taxonomy objects at a one percent false alarm rate, but they show that a simulated review loop can recover eighty-three percent of them after inspecting fifteen percent of the stream.

Vera: That learning from confirmations to new classes is where it gets practical; for TDEs, retraining with them as a new class reaches eighty-five percent detection, while for novae, just using the head on the frozen features is enough.

Jocelyn: It really suggests a practical sequence: rank the unlabelled candidates for review and learn from their features as confirmations accumulate.

Subrahmanyan: So, when you look at the bigger picture, this sequence implies that we shouldn't just rely on one single classification method but build up knowledge iteratively based on what we find.

Vera: And then there’s the deployment aspect; it runs in production on Fritz, which is the ZTF instance of SkyPortal.

Jocelyn: The alarm reported there is calibrated as a false alarm rate per unit time based on the photometry that the service receives, which is essential for high-cadence surveys.

Subrahmanyan: That calibration method shows how important it is to tie your operational metrics directly to the input photometry because those things change depending on what you feed in.

Vera: They also generate classconditional Mondrian conformal prediction sets for follow-up schedulers, providing these sets to guide resource allocation by reporting the nonconformity one minus p c is calibrated separately.

Jocelyn: That sounds like a very direct way to tell a scheduler exactly where they should spend their time looking for something new.

Subrahmanyan: It’s about optimizing the observation time based on the predicted uncertainty of each class, which is vital when we have limited telescope time across many targets.

Vera: And they conclude that both families benefit more from learning in feature space than from combining the anomaly scores.

Title and authors: Jocelyn: That’s a clear statement about where the real power lies for discovering new things in these transient streams.

Subrahmanyan: So, to wrap up on this paper, "Photometric classification and out-of-taxonomy detection of ZTF transients with FLARE," it shows a practical way to move from raw alerts to confirmed classes by using context and iterative learning.

Vera: It’s really about making sure that when we get an alert, we know what kind of object it is quickly, even if it doesn't fit our standard bins.

Jocelyn: And the deployment on Fritz shows that this framework works in production, reporting its false alarm rate per unit time calibrated on the actual photometry it receives.

Subrahmanyan: From a theoretical side, this suggests that for structure formation studies, having a robust way to flag objects outside the known classes could help us identify unexpected physical processes happening in galaxies.

Vera: The main limitation they point out is that the size and selection of the samples are still an issue because the rarest families only contain one to six objects, which is too few to establish a general photometric detection limit.

Jocelyn: And they also mention that the review loop assumes perfect labels at a prevalence of three point three percent, which is higher than what you usually get in a raw alert stream.

Subrahmanyan: So, while it’s a good tool for discovery, the limitation is that we still need more confirmed examples to build up those rare classes robustly enough for general predictions.

Vera: They also noted that the early-time study selects objects by their eventual light-curve quality, and when transferring to service photometry, both false alarm rates and prediction set coverage change based on the input.

Jocelyn: So, it’s not a fixed thing; what you feed into the system really dictates its performance.

Subrahmanyan: It reinforces that any model we build for cosmology needs to be sensitive to these kinds of observational biases because they affect what we actually see in the sky.

Vera: So, if you take this whole process—ranking, reviewing, learning—it gives us a solid path toward shorter timescales between an initial alarm and a fully trained class.

Jocelyn: It’s about making that path as short as possible for those objects we are hunting for.

Subrahmanyan: That’s what matters when you're trying to catch rare events before they fade or evolve too much in the universe.

Vera: So, that’s the sequence: rank the unlabelled candidates, review them, and learn from their features as confirmations accumulate.

Jocelyn: And that framework is what they call FLARE for ZTF transients.

The paper's summary: Vera: So, this paper is about taking those ZTF alerts—those initial light curve detections—and using this whole system, FLARE, to sort them out and figure out what kind of transient we're looking at without needing a full spectrum right away.

Jocelyn: Exactly. It’s moving from just seeing something bright in the sky to actually having a label for it, even if that label is one of these new, weird objects that don't fit our usual categories.

Subrahmanyan: From my side, this is about building a practical pipeline where the information you already have—the photometry—can do a lot of heavy lifting in deciding if you need to go get a telescope time for a spectrum.

Vera: They’re using external context blocks—like host galaxy data and redshift estimates from Legacy Surveys—to give the classifiers much better information about the light curve.

Jocelyn: That extra context seems to be what really boosts that classification accuracy, moving it up from a baseline score to something much more reliable.

Subrahmanyan: And that's important because when you’re looking at structure formation, knowing if an object is in a specific environment tells you a lot about the physics happening around it.

Vera: They also have this anomaly layer on top of the classifier using a package called luna1, which scores how much something looks like it belongs in any of the known families versus being totally out there.

Jocelyn: And that’s where they get some interesting numbers; blind detection recovers a small fraction of those weird objects, but when you use a simulated review loop to inspect just fifteen percent of the stream, you recover eighty-three percent of them.

Subrahmanyan: Eighty-three percent recovery after only looking at a small subset is significant because it means this system can actually flag things we’ve never seen before without needing immediate spectral confirmation for every single alert.

Vera: The authors conclude that the best way to use this is a sequence: you rank the unlabelled candidates for review first, and then you learn from those initial confirmations as they accumulate.

Jocelyn: That means we don’t have to wait for perfect labels on everything before we start building our knowledge base; we can use the learning process itself to guide which objects deserve that precious telescope time.

Subrahmanyan: That iterative approach makes sense for cosmology because it allows you to start constraining models with what you *do* know while simultaneously flagging the most interesting new data points for follow-up.

Vera: The deployment on Fritz is real-time, and they report the alarm as a false alarm rate per unit time based on the exact photometry that service receives—which shows how much you actually have to calibrate your metrics against.

Jocelyn: It’s crucial for operational work because it tells you exactly how often you should expect a false alarm based on what the system is seeing right now, not some old average.

Subrahmanyan: So, the implication here is that this isn't just a classification tool; it's a way to dynamically build our catalog by prioritizing observations based on where the uncertainty is highest.

Vera: The main limitation they point out is still the size and selection of the samples; those rarest families only contain one to six objects, which means you can’t really set a general photometric detection limit yet.

Jocelyn: And they also admit that their review loop assumes perfect labels at three point three percent prevalence, which is way higher than what you usually get when you look at the raw alert stream.

Subrahmanyan: That tells us that while this method is powerful for finding rare things, we still need more confirmed examples across all families before we can trust its predictions for the whole sky.

Vera: It also showed that whether you learn in feature space or combine anomaly scores, learning in the feature space ends up being better for those new classes.

Jocelyn: So for any follow-up strategy, focusing on the features of an object rather than just mixing anomaly scores seems to be the stronger path forward.

Subrahmanyan: That connects back to structure formation; if we can reliably learn from these features, it helps us understand the physical mechanisms driving those transients in different galactic environments.

Vera: This whole sequence—rank, review, learn—is a practical way to shorten that time between seeing an alarm and having a trained class for an unknown object.

Jocelyn: It’s about making sure we get that path as short as possible so we can catch those rare events before they evolve too much in the universe.

Subrahmanyan: That efficiency is what matters when you're trying to catch things that are inherently transient and evolve rapidly over cosmic timescales.

The paper's improvements: Tom: So, we're looking at how they suggest improving this FLARE system, moving beyond just what they did in the main paper to show where things can go next.

Vera: They’re focusing on integrating that catalogue context more deeply into the light curve feature space so that the classifiers actually make a better decision about which alerts get a spectrum.

Jocelyn: That means if the defining information isn't in the raw light curve, this method can use external data to help pick it up, which is huge for finding things that look weird.

Subrahmanyan: From a theoretical standpoint, that context integration means we can start constraining physical parameters earlier in the process of classifying these transients.

Vera: They also mentioned using classconditional Mondrian conformal prediction sets for follow-up schedulers; this gives them specific sets of objects to watch based on the predicted uncertainty of each class.

Jocelyn: That’s a direct tool for resource allocation, telling a scheduler exactly where they should spend their time looking for new things rather than just scanning blindly.

Subrahmanyan: It’s about optimizing telescope time based on the predicted uncertainty of each class, which is vital when you have limited telescope time across many targets.

Vera: They also explored differentiating between learning in feature space versus combining those anomaly scores, concluding that learning in feature space actually works better for building new classes.

Jocelyn: That suggests that instead of just mixing the anomaly scores together, focusing on how the features themselves change is a more productive way to grow the classifier.

Subrahmanyan: For structure formation studies, if we can reliably learn from those features across different environments, it helps us understand the physical mechanisms driving those transients in different galactic settings.

Vera: The authors also discussed how they handle calibration when deploying this on Fritz; they said the alarm should be reported as a false alarm rate per unit time calibrated directly on the photometry the service actually receives.

Jocelyn: That emphasizes that you have to be very careful about what you calibrate against, because those things change depending on what input data you are feeding into the system.

Subrahmanyan: So, it reinforces that any model we build for cosmology needs to be sensitive to these kinds of observational biases because they affect what we actually see in the sky.

Vera: They also highlighted that the deployment on Fritz showed this module works in production, and they keep that consensus as its alarm, using p-values computed separately for groups of objects with a similar number of detections.

Jocelyn: It means even when things get messy with many detections, the system is still making a calculated call based on the data it sees.

Subrahmanyan: That operational stability is what we need if we want to use these tools to map out large-scale cosmic structures consistently over time.

Vera: The main limitation they flag is that the size and selection of the samples are still an issue because the rarest families only contain one to six objects, which is too few to establish a general photometric detection limit.

Jocelyn: And they also noted that their review loop assumes perfect labels at three point three percent prevalence, which is higher than what you usually get when you look at the raw alert stream.

Vera: It also showed that the early-time study selects objects by their eventual light-curve quality, and when transferring to service photometry, both false alarm rates and prediction set coverage change based on the input.

Jocelyn: So it’s not a fixed thing; what you feed into the system really dictates its performance in real time.

Conclusion: Vera: So we’re wrapping up on this paper, "Photometric classification and out-of-taxonomy detection of ZTF transients with FLARE," by summarizing what they did and where this research leads us.

Jocelyn: Basically, they showed how to take a raw light curve alert, use external context to classify it better than just the light curve alone, and then use an anomaly layer to flag those truly weird objects.

Subrahmanyan: It’s about building a practical sequence for transient discovery by moving from blind detection all the way up to a supervised classification.

Vera: The main takeaway is that you rank unlabelled candidates for review first, and then you learn from their features as those confirmations accumulate before trying to assign them a formal class.

Subrahmanyan: For structure formation, this iterative approach allows you to start constraining models with what you do know while simultaneously flagging the most interesting new data points for follow-up.

Vera: The deployment on Fritz showed that this module works in production and reports its alarm as a false alarm rate per unit time calibrated directly on the photometry it receives.

Jocelyn: That means we have a real operational metric, so you know exactly what the system is seeing right now without relying on some old average.

Subrahmanyan: This calibration method shows how important it is to tie your metrics directly to the input photometry because those things change depending on where you are in the sky.

Vera: The authors also pointed out that while this sequence is powerful, their biggest limitation is still the size of their samples; those rarest families only contain one to six objects, which means you can’t really set a general photometric detection limit yet.

Jocelyn: And they also admitted that their review loop assumes perfect labels at three point three percent prevalence, which is way higher than what you usually get when you look at the raw alert stream.

Vera: They also showed that whether you learn in feature space or combine those anomaly scores, learning in feature space ends up being better for building those new classes.

Jocelyn: So for any follow-up strategy, focusing on how the features of an object change seems to be a more productive way to grow the classifier.

Subrahmanyan: That connects back to cosmology; if we can reliably learn from those features across different environments, it helps us understand the physical mechanisms driving those transients in different galactic settings.

Vera: This whole sequence—rank, review, learn—is a practical way to shorten that time between seeing an alarm and having a trained class for an unknown object.

Jocelyn: It’s about making sure we get that path as short as possible so we can catch those rare events before they evolve too much in the universe.

Subrahmanyan: That efficiency is what matters when you're trying to catch things that are inherently transient and evolve rapidly over cosmic timescales.

Vera: So, this paper, "Photometric classification and out-of-taxonomy detection of ZTF transients with FLARE," gives us a solid path toward a shorter timescale between an initial alarm and a fully trained class.

Jocelyn: It’s about making that path as short as possible for those objects we are hunting for.

Subrahmanyan: We need more robust samples, but this framework provides the best way to use the data we have right now to prioritize our observations across the whole sky.

University of Minnesota · Aristotle University of Thessaloniki · California Institute of Technology · Stockholm University · NSF Institute on Accelerated AI Algorithms for Data-Driven Discovery (A3D3) · IPAC, California Institute of Technology · Caltech Optical Observatories, California Institute of Technology · Department of Astronomy, Stockholm University · Cahill Center for Astrophysics at California Institute of Technology · Center for Data Driven Discovery, California Institute of Technology

astro-ph.IM, astro-ph.HE

Submitted: 2026-10-22

Updated: 2026-10-22

Comments: 21 pages, 10 figures

Code: https://github.com/asasli/luna

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

Importance score: 84/100

The gist: The gist The flare module, which combines photometric classification, anomaly ranking, and learning from confirmed candidates, provides a practical sequence for transient discovery by moving from

Key concepts

Flare Module: The core system that combines photometric classification, anomaly ranking, and learning from confirmed objects. It creates a practical workflow for finding new transients by moving from initial blind detection to human review and eventually supervised classification.
Host-Galaxy Context: Incorporating external data like host photometry, photometric redshifts, and Gaia/WISE information into the analysis. This context significantly boosts classification accuracy by providing crucial information about the transient's environment.
Anomaly Detection Layer: A layer using luna1 scorers to flag objects that are out-of-taxonomy based on their features. It allows for blind detection of new types, recovering a significant portion of these objects after a simulated review loop.

Summary

The gist The flare module, which combines photometric classification, anomaly ranking, and learning from confirmed candidates, provides a practical sequence for transient discovery by moving from blind detection to human review to supervised classification

How it works

The core of the approach is the AppleCiDEr framework's flare module, which is built for the boom alert broker and whose spectroscopic module, astraNet, provides classifications and anomaly scores for spectra. This module uses ZTF photometry and public catalogues without spectra or image cutouts, evaluating it on 10 497 spectroscopically confirmed transients and 53 objects from families outside its six-class taxonomy. The workflow focuses on three aspects: (i) the information that the external context adds to the light curve, (ii) the recovery of out-of-taxonomy objects by blind ranking, active review and learning from first labels, and (iii) the reliability of both layers under missing context, short observation horizons and transfer to Fritz.

Classification Granularity

The authors evaluate the pipeline at three levels of granularity to assess its performance. Splitting core-collapse supernovae into Type II and stripped-envelope supernovae lowers the macro F1 by seven points, and stripped-envelope objects reach only F1 = 0.32 because they are divided almost evenly between SN Ia and SN II classes. Merging SLSNe into the core-collapse class improves every aggregate metric because the SLSNe are separated from Type II supernovae mainly by their absolute magnitude.

Host-Galaxy Context and Distance Chain

The model incorporates four external context blocks: Pan-STARRS host photometry and angular offset, a photometric redshift from Legacy Surveys, the Gaia and WISE point-source context, and a Pan-STARRS pre-outburst counterpart. The host photometry gives most of the gain in classification accuracy, with the macro F1 rising from 0.72 to 0.83 when all four blocks are added. The photometric redshift supplies a pseudo-absolute magnitude, Mpseudo = mpeak − 5 log10[dL(zphot)/10 pc]. The point-source block improves the AGN through the W1 − W2 colour and the presence of a Gaia counterpart, while the pre-outburst counterpart separates a flare in a nucleus from a supernova near one and marks CVs that Gaia misses.

Anomaly Detection and Recovery

The anomaly layer operates on top of the trained classifier using luna1, an out-of-distribution scoring package with fourteen families of scorers. The scores are combined either blind, as the mean percentile rank across the scorers, or by a supervised combiner when confirmed examples exist. Blind detection recovers few out-of-taxonomy objects at a 1% false-alarm rate, while a simulated review loop recovers 83% of them after inspecting 15% of the stream.

Learning from Confirmations to New Classes

The paper measures what the first confirmations of a new family are worth when used in a combiner over anomaly scores, in a head on frozen features, or as a new class of the classifier. For TDEs, retraining with them as a new class reaches 85% detection, while for novae, the head on the frozen features is enough. The results suggest a practical sequence: rank the unlabelled candidates for review and learn from their features as confirmations accumulate.

Deployment and Service Calibration

The module runs in production on Fritz, the ZTF instance of SkyPortal, where its alarm is reported as a false-alarm rate per unit time calibrated on the photometry that the service receives. The transfer to the photometry of the service shows that calibration has to be made on that photometry and an alarm of a service which re-analyses its sources is best reported as a false-alarm rate per unit time.

Conclusion

The results suggest a practical sequence: rank the unlabelled candidates for review, and learn from their features as the confirmations accumulate. The deployment on Fritz shows that the module operates in production, and the transfer to the photometry of the service shows that calibration has to be made on that photometry and an alarm of a service which re-analyses its sources is best reported as a false-alarm rate per unit time. The framework is expected to be very useful for the Rubin era, where the path of an unknown object from the first alarm to a trained class has to be short. The main limitations come from the size and selection of the samples. The rarest families contain one to six objects, too few to establish a general photometric detection limit. The review loop assumes perfect labels at a prevalence of 3.3%, much higher than in a raw alert stream, where the first rounds would return fewer confirmations. The early-time study selects the objects by their eventual light-curve quality, and the transfer to the photometry of the service shows that both false-alarm rates and prediction-set coverage change with the input. The deployed module keeps this consensus as its alarm, with p-values computed separately for groups of objects with a similar number of detections. The weekly threshold is used to send sources for expert review, and the quarterly threshold does not achieve its target. The final population on which the method can produce new results is the 7 631 BTS objects that never received a spectrum. The objects whose prediction set excludes their own most probable class and the top of the anomaly ranking form a target list for spectroscopy, once the false-alarm threshold has been recalibrated on classified objects of matched magnitude. The paper is available at https://github.com/applecider-ml/flare. The outof-distribution scoring package used by the anomaly layer is available at //github.com/asasli/luna. The authors acknowledge support from the National Science Foundation with grant numbers PHY2117997, PHY-2308862 and PHY-2409481. The UMN authors acknowledge support from the National Science Foundation with grant numbers PHY2117997, PHY-2308862 and PHY-2409481. The LINCC Frameworks is supported by Schmidt Sciences, a philanthropic initiative founded by Eric and Wendy Schmidt, as part of the Virtual Institute of Astrophysics (VIA). The paper makes use of data products from the Wide-field Infrared Survey Explorer, which is a joint project of the University of California, Los Angeles, and the Jet Propulsion Laboratory/California Institute of Technology, and NEOWISE, which is a project of the Jet Propulsion Laboratory/California Institute of Technology. This research made use of Astropy, LightGBM, Optuna, and the light-curve feature-extraction library. The AI disclosure statement indicates that the authors used Claude (Fable 5.1 and Opus 5.5) and ChatGPT (OpenAI) to assist with editing and restructuring portions of the text for clarity.

The gist The flare module, which combines photometric classification, anomaly ranking, and learning from confirmed candidates provides a practical sequence for transient discovery by moving from blind detection to human review to supervised classification

How it works

Improvements for AI systems

Integration of catalogue context into light curve feature space allows classifiers to decide, in practice, which alerts receive a spectrum, improving detection for objects where the defining information is absent from the light curve.
Implementation of an anomaly layer using luna1 scorers allows the system to perform blind detection of out-of-taxonomy objects at a 1% false-alarm rate, recovering few out-of-taxonomy objects and achieving 83% recovery after inspecting 15% of the stream via a simulated review loop.
Deployment on Fritz enables real-time operation where the system reports an alarm as a false-alarm rate per unit time on the stream, calibrated on the photometry that the service receives, which is crucial for operational use in high-cadence surveys.
The system can generate classconditional Mondrian conformal prediction sets for follow-up schedulers, providing classconditional (Mondrian) conformal prediction sets to guide resource allocation by reporting the nonconformity 1−pc is calibrated separately.
The framework allows for a practical sequence where the system can rank the unlabelled candidates for review, and learn from their features as the confirmations accumulate, which suggests a methodology for prioritizing spectroscopic follow-up.
The system can differentiate between learning in feature space and combining anomaly scores, concluding that Both families benefit more from learning in the feature space than from combining the anomaly scores.

Abstract

Photometric classifiers decide which of the million nightly alerts of time-domain surveys receive one of the few thousand spectra taken each year, and they are the first place where a transient outside the trained classes can be noticed. We present flare, the photometric classification and anomaly-detection module of AppleCiDEr, evaluated on 10 497 spectroscopically confirmed ZTF transients in six classes and 53 objects from families outside the taxonomy. Catalogue context raises the macro F1 score, with the largest gains where the defining information is absent from the light curve. Blind detection recovers few out-of-taxonomy objects at a 1% false-alarm rate, while a simulated review loop recovers 83% of them after inspecting 15% of the stream. The first confirmations of a new family are most useful in the feature space: with the same eleven labels, a head on the frozen features detects 68% of the tidal disruption events hidden from the training against 57% for a combiner over the anomaly scores, and retraining with them as a new class reaches 85%. The module runs in production on Fritz, the ZTF instance of SkyPortal, and its alarm is reported as a false-alarm rate per unit time on the stream, calibrated on the photometry that the service receives