☰

Babamul

/objects/ztf/ZTF19acbzgog

Design proposal, not a live page. The chrome, the card grid, the “ML Scores” tiles, the band colours and every token here are taken from boom-astro/boom’s own frontend source; the FLARE numbers are real outputs of the shipped bts6 models. Nothing was screenshotted and no BOOM classifier score is invented.

ZTF19acbzgog

ZTF

IAUAT2019rsi RA00:42:23.53 Dec+41:11:52.2 detections30 peak mag17.91 spectroscopicclassical nova — outside the taxonomy

Lightcurve

ZTF g, r

ML Scores

FLARE ▾

CurrentTemporal

FLARE

SN Ia

50%

SN CC

44%

SLSN

AGN

TDE

BOOM classifiers

drb

–

btsbot

–

acai_h

–

sgscore1

–

Tiles follow the widget's own rule: filled block, three-bucket colour at 0.7 and 0.4, integer per cent. BOOM's own classifiers are blank here because this page was rendered offline from the FLARE models alone.

FLARE · photometric classification

bts6

prediction set

{SN CC, SLSN}

90% ± 10%

target coverage ± the error the set admits (α = 0.10, class-conditional)

measured out of fold, SN Ia n = 7159	90.6% ± 0.3%
credibility max p — matches any class at all	0.143
confidence 1 − second p	0.875
p-value floor 1/(n_cal+1)	0.003

conformal p-values set = {p > α}

SN Ia0.024

SN CC0.143

SLSN0.125

AGN0.053

TDE0.083

CV0.085

is it an anomaly?

30.7% (23–41%)

probability it is out of taxonomy, at the benchmark’s own 3.1% base rate

likelihood ratio prior-free	13.7×
novelty p-value 1 in 315 known-class objects is this extreme	0.0032

anomaly energy · own 181-column space, red mark = 1% false-alarm budget

percentile against the known-class stream 99.7

ANOMALY · the argmax class is excluded by its own prediction set and the energy clears the 1% false-alarm threshold — sent to the review queue

Cross-matches

PS1_DR2 (host)1 match

sep [arcsec]	2.13
rKronMag	17.88
iKronMag	18.80
g-r	nan

AllWISE1 match

sep [arcsec]	6.10
W1	12.76
W1-W2	-0.55

Gaia_DR30 matches

LS_DR10_PHOTOZ0 matches

PS1_DR2 (at position)0 matches

ZTF25abhoijcSN Ia (BTS)

FLARESN Ia

set (α=0.10){SN Ia}

coverage90% ± 10% target · measured 90.6% ± 0.3%

credibility0.457

P(anomaly)2.1% (2–2%) · LR 0.6×

ZTF22abkfhuaTDE (BTS)

FLARETDE

set (α=0.10){TDE}

coverage90% ± 10% target · measured 91.2% ± 3.4%

credibility0.750

P(anomaly)1.5% (1–2%) · LR 0.5×

What the current UI could and could not show. The six class scores drop straight into the existing “ML Scores” card — but only after mapAlertClassifications in ClassificationsV2.tsx stops being a hard-coded whitelist (acai_*, btsbot, drb, reliability, sgscore1, LSPSC), which silently drops every other key. The prediction set, the error it admits and the anomaly energy have no home in the widget at all: it draws one number per classifier with a three-bucket colour and no threshold lines, so the FLARE card above is a new component.

Pipeline fit. BOOM already runs its models as ONNX in enrichment workers (ORT_DYLIB_PATH, results written back to MongoDB) and crossmatches static catalogues per alert, which is exactly what FLARE consumes: PS1 host and pre-outburst counterpart, Legacy Surveys photo-z, Gaia, AllWISE. The classifier is LightGBM and converts to ONNX; the anomaly layer reads its own narrower 181-column space.