Production proposal. Every number on this page is a
genuine output of the shipped models on real ZTF objects — including one the
detector misses. What a live deployment would add is throughput, not different
numbers.
ZTF19acbzgog
· 10.59808 +41.19782 · 30 detections
over 76 d · peak 17.91
Anomaly
Anomaly. Energy clears the 1% false-alarm budget and the prediction set excludes the argmax class.
Light curve
Classification 6 classes · 190 features
SN Ia50%p 0.024
SN CC44%p 0.143
SLSN1%p 0.125
AGN0%p 0.053
TDE1%p 0.083
CV4%p 0.085
{SN CC, SLSN}
90% ± 10%coverage the set is calibrated for — it may
miss the true class one time in ten, by construction
measured out of fold, SN Ia
90.3% ± 0.3%
n = 7159
credibility does it match any class at all
0.143
confidence is the runner-up excluded
0.875
The most probable class is not in the set: its p-value is below the 10% error level, so the classifier is contradicting itself.
Anomaly own 181-column space
31.1%(nan–nan%)
probability this is outside the six-class taxonomy, at the
benchmark's 3.2% base rate
likelihood ratio carries no prior — rescale
for your stream
13.7×
novelty p-value 1 known-class object in
315 is this extreme
0.0032
energy
+2.06
percentile against the known-class stream99.7
Evidence context at alert time, no spectrum
Gaia point source
none
nothing persistent at the position above the Gaia limit
WISE W1−W2
-0.55
no mid-infrared AGN excess
Pre-outburst counterpart
none
nothing at this position in Pan-STARRS before the transient
Host offset
2.22″
displaced from its host
Photometric redshift
none
no distance proxy available
ZTF22abkfhua
· 350.84908 +10.68551 · 48 detections
over 97 d · peak 18.08
Follow-up
Follow-up. Singleton set for a class whose defining property only spectroscopy resolves.
Light curve
Classification 6 classes · 190 features
SN Ia0%p 0.001
SN CC1%p 0.016
SLSN0%p 0.042
AGN0%p 0.053
TDE99%p 0.667
CV0%p 0.021
{TDE}
90% ± 10%coverage the set is calibrated for — it may
miss the true class one time in ten, by construction
measured out of fold, TDE
92.6% ± 3.2%
n = 68
credibility does it match any class at all
0.667
confidence is the runner-up excluded
0.947
Anomaly own 181-column space
2.8%(nan–nan%)
probability this is outside the six-class taxonomy, at the
benchmark's 3.2% base rate
likelihood ratio carries no prior — rescale
for your stream
0.9×
novelty p-value 1 known-class object in
3 is this extreme
0.3179
energy
-0.81
percentile against the known-class stream68.3
Evidence context at alert time, no spectrum
Gaia point source
0.18″ away
a persistent source sits at the position — nuclear or stellar
WISE W1−W2
0.02
no mid-infrared AGN excess
Pre-outburst counterpart
0.20″ · r = 18.6
Pan-STARRS saw something here years before the outburst
Host offset
0.20″
on the nucleus
Photometric redshift
none
no distance proxy available
How to read it. The prediction set is the operational
output, not the argmax: it is calibrated so that it contains the true class
90% of the time, and the 10% it admits is stated rather than hidden. A set
with two classes is the model declining to choose, which is the correct
hand-off to spectroscopy. The anomaly probability is separate machinery —
scored in its own representation, calibrated against the benchmark's
out-of-taxonomy objects — and it is quoted with the base rate it assumes,
because the same likelihood ratio gives a very different probability in a raw
alert stream than in a spectroscopically confirmed sample.