About & Methods

Rubin Anomaly Platform — early-stage research tool, not a vetted catalog

What this is

An automated pipeline that polls the Vera Rubin Observatory (LSST) public alert stream via the Lasair and Fink brokers, computes photometric features from flux light curves, scores each object for anomalousness, and cross-matches against archival catalogs to identify candidates with no known counterpart.

The goal is to surface transients that current broker classifiers cannot confidently identify — objects that may warrant follow-up, or that represent data-quality artifacts worth investigating.

Pipeline stages

Stage	What it does
Ingest	Polls Lasair-LSST and Fink-LSST REST APIs for new diaSource detections. Stores raw alert JSON and extracted flux fields (psfFlux in nanojanskys, midpointMjdTai, band).
Quality gate	Removes solar-system objects (ssObjectId ≠ 0), near-zero-length dipole fits (dipoleLength < 0.15″), PSF/centroid measurement failures, and detections with \|SNR\| < 3. Objects with fewer than 3 passing detections are deferred rather than scored.
Features	Extracts light-curve features per band using the `light-curve` library: flux mean, standard deviation, reduced chi-squared (Γ), weighted mean magnitude proxy, and inter-band color proxies where multi-band coverage exists. Feature set is versioned alongside the model.
Anomaly score	IsolationForest (scikit-learn) trained on the current feature corpus. Score is normalized to [0, 1]; higher = more anomalous relative to the corpus. The model is re-trained nightly as the corpus grows. Model name and version are recorded with every score row.
Cross-match	Cone-searches four catalogs at the object's RA/Dec: Gaia DR3 (stellar parallax), SIMBAD (spectroscopic IDs), AllWISE (IR counterparts), and TNS (reported transients). Results cached; query failures recorded distinctly from genuine no-matches and excluded from novelty classification.
Novelty	Assigns novel_candidate, known_object, or ambiguous based on cross-match evidence. An object is promoted to novel_candidate only when all three required catalogs (Gaia DR3, SIMBAD, AllWISE) have been checked and returned no match.

How to read the leaderboard

Column	Meaning and caveats
Score	IsolationForest anomaly score, normalized 0–1. Measures how unlike the rest of the corpus this object's feature vector is. Not an astrophysical significance. A high score on a sparse (few-epoch) light curve is less trustworthy than the same score on a rich light curve — check the Tier and n_epochs on the detail page.
Γ (reduced χ²)	Reduced chi-squared of a constant-flux model fit. Γ > 1 means the flux varies significantly; Γ ≈ 1 is consistent with constant. Always shown with ±σ. The relative error (σ/Γ) is a quality signal: above 50% the value is poorly constrained. Computed from the r-band light curve when available; falls back to the densest band.
Status badge	novel candidateNo counterpart found in Gaia DR3, SIMBAD, AllWISE, or TNS. ambiguousAn IR/astrometric counterpart exists but is unidentified, or the cross-match cone had multiple sources. known object Matched to a cataloged source (Gaia parallax, SIMBAD type, or TNS report).
Tier	Data richness tier based on passing detection count. Rich ≥ 20 epochs, Moderate 8–19, Sparse 3–7. Sparse-tier objects have fewer features computed and scores are less stable.
n_epochs	Number of detections that passed the quality gate. This is the sample size underlying every feature and score on this page.

Known limitations

These limitations are real and acknowledged. Do not cite results from this platform as scientific findings without independent verification.

Commissioning data only. Rubin is in its commissioning phase. Alert data volume, cadence, and quality are not representative of survey operations. The corpus is small (~hundreds of objects as of mid-2026).
Sparse-tier dominance. Most objects are sparse-tier (3–7 passing detections). Scores derived from few-epoch light curves are sensitive to noise and less reliable.
No time-variability features yet. The current feature set does not include rise time, decay slope, or period. Periodic or fast-evolving transients may not score anomalously.
IsolationForest is an unsupervised baseline. v1 does not incorporate labeled training examples of known transient types. It scores relative to the current corpus — if the corpus is mostly one type of object, rare-but-known objects of another type may score high.
Cross-match completeness. Catalog queries occasionally time out. When they do, the result is excluded from novelty classification (not counted as a no-match), but the object remains in limbo until the query succeeds on a later run.
No spectroscopic confirmation. Every novel_candidate here is a photometric anomaly that lacks an archival counterpart — not a spectroscopically classified new transient class.

Verification approach

Every interpretive claim on this platform traces to a computed field — no astrophysical narrative is generated from background knowledge about a specific object. Cross-match status, feature values, and epoch counts are derived from API responses stored in the database. If a field is uncertain or unavailable, the display says so explicitly.

Pipeline run history (start time, stages completed, query failure counts) is stored in the pipeline_runs table and will be surfaced in a future status page.

Data sources

Alert stream: Lasair-LSST and Fink-LSST (Vera Rubin / LSST alert brokers)
Astrometry: Gaia DR3 via astroquery
Spectroscopic IDs: SIMBAD via astroquery
Infrared: AllWISE via astroquery / VizieR
Reported transients: Transient Name Server (TNS)

Contact

Solo project by a CS/astronomy student. Source code and issue tracker on GitHub. Not affiliated with the Rubin Observatory, Lasair, or Fink teams.

← Back to leaderboard