Anomalous light curves in TESS Sector 103

Data

Mission	TESS, Sector 103 (Year 8)
Product	2-minute SPOC light curves, PDCSAP flux
Data release	DR141 (release notes DRN141)
SPOC version	spoc-5.0.126-20251212
Light curves	12,995 targets

Only data with published release notes are used, so every input can be cited. Sectors 105–107 already have light curves on MAST but no release notes yet: they are left out of this experiment and noted for a second version.

Protocol

The pipeline runs in fixed phases. Every run records the hash of its parameters and the version of each input file.

1. Development sample, then frozen parameters. All thresholds are chosen on a sample of 3,000 targets, drawn with a fixed seed and stratified by TESS magnitude. The parameters are then frozen, and a single final run covers all 13,000 targets without further changes.

2. Cleaning. Cadences with quality flags or non-finite values are removed, the curve is split at gaps longer than half a day, and each segment is normalised by its median. No detrending and no sigma-clipping: they would remove the signals we are looking for. "Impulsive outlier" cadences are kept, because a short real event can carry that flag, and are judged later.

3. Features. For each star: variability and shape of the flux distribution, a Lomb–Scargle periodogram, a box search for transit-like dips, and the strongest isolated dip and flare. The model inputs are expressed in units of each star's own point-to-point noise, so a faint star that is pure noise looks the same as a bright star that is pure noise.

4. Ranking. Isolation Forest and Local Outlier Factor score every star; the rank is the mean of the two percentile ranks. For each star the three features furthest from the median are listed, so the reason for its rank can be read.

5. Artefact rejection, as an explicit phase. Criteria are applied in a fixed order and each one is counted in two ways: how many stars it removes after the previous ones, and how many would fail it on its own.

   criterion	rejects when
   undefined flux	the median flux is close to zero after background subtraction
   contamination	less than half of the flux in the aperture belongs to the target
   centroid systematic	the flux follows the image centroid
   background systematic	the flux follows the sky background
   event at an edge	the event driving the score sits within 0.1 d of a data gap
   impulsive outlier	at least half of the event is flagged as an impulsive outlier

6. Catalogue check. Each candidate is matched against SIMBAD, VSX, the TESS Objects of Interest and the TESS eclipsing-binary catalogue. It is labelled known (a catalogued class explains the event), known with a different class, or absent (no variability class in any catalogue). Events that coincide in time with events on other stars on the same detector are flagged as possible systematics.

7. Persistence. For candidates observed in other sectors (SPOC light curves only, up to Sector 104), the same features are computed there. A candidate with no other sector is labelled persistence not verifiable: it stays on the list and is reported separately.

What this experiment can and cannot say

The output is a ranked list of candidates for follow-up. An absent label means only that the four catalogues above give the star no variability class at the time of the query. It is not a confirmation that the behaviour is new or astrophysical: that needs inspection of the pixels, other sectors and, where possible, other telescopes. The report states the number of candidates in each label and the known limits of the method.

Records

Each run is stored on VDX Core as an execution of the project, with the parameters, the dataset version and every candidate attached as a result with its plot. The code and the report are versioned together in the project repository.

Summary

The final run covered all 12,995 Sector 103 light curves with the parameters frozen on the development sample (parameter fingerprint 7c11bba8442014f5). It produced a list of 50 candidates for follow-up. Eleven of them have no variability class in any of the four catalogues queried. None of the 50 is a confirmed new variable: each one is a candidate until someone inspects the pixels and checks other observations.

step	stars
Light curves in Sector 103	12,995
Kept after cleaning	12,919
Pass every artefact criterion	5,952
Candidates (top 50 of the survivors)	50

Data: DR141, SPOC spoc-5.0.126-20251212. Catalogues were queried on 2026-10-26.

Artefact rejection

Criteria are applied in order. Sequential counts the stars a criterion removes after the earlier criteria have run. Independent counts the stars that would fail that criterion if it were applied alone.

criterion	sequential	independent
undefined flux	3,502	3,502
contamination	2,396	5,708
centroid systematic	11	255
background systematic	79	97
event at an edge	979	1,707
impulsive outlier	0	4

Catalogue labels

event	known	known, different class	absent	total
isolated dip	25	3	9	37
flare	4	7	2	13
total	29	10	11	50

Persistence was checked on the 50 candidates. In 41 the event appears again in at least one other sector, in 6 it does not appear in any other sector, and 3 have no other sector to check. No event lines up in time with events on other stars of the same detector beyond what chance would produce. Twelve candidates have a catalogued variable star within 42 arcseconds, which is two TESS pixels.

Candidates absent from the catalogues

rank	TIC	event	SIMBAD	persistence	notes
1	183562384	dip	HD 4292 (spectroscopic binary)	4 of 4 sectors
4	292292315	dip	HD 170312	3 of 8	also in the development sample
6	231632822	dip	HD 201172	1 of 3	centroid shift during the event
10	370324073	dip	TYC 8829-111-1 (spectroscopic binary)	5 of 6
14	464846595	dip	HD 138585	2 of 2
39	260889990	flare		4 of 5	partial contamination
49	299271875	dip	HD 145453	3 of 4
73	281597526	flare		not persistent (0 of 7)	rotational variable 21″ away
85	469974264	dip	no SIMBAD object	2 of 3	also in the development sample
93	39177520	dip	HD 150168	1 of 4	eclipsing binary 23″ away
94	8419033	dip		not verifiable	rotational variable 26″ away

The rank is the star's position in the combined anomaly ranking of all stars. The ranking was computed before artefact rejection, which is why the ranks go above 50.

A persistent dip in a star with no variability class is the most direct lead for follow-up. TIC 183562384 and TIC 370324073 are spectroscopic binaries, so an eclipse is the obvious first explanation to test. For the candidates with a catalogued variable within two pixels, that neighbour has to be ruled out as the source before anything else.

Limits

• Model dependence. Isolation Forest and LOF agree only weakly (Spearman ρ 0.27, and 3 stars shared in their top 50s). On the development sample the agreement was ρ 0.53 with 10 shared. The list therefore depends on the choice of model, and other anomalous stars sit just below the cut. Refitting the models on the survivors of artefact rejection keeps 38 of the 50.
• One sector. Each event is found in Sector 103 alone. Persistence uses only SPOC light curves up to Sector 104. Sectors 105–107 are left for a second version.
• Resolution. TESS pixels are 21″ wide, and this run does no pixel-level centroid analysis. A nearby star can therefore produce an event in the target's light curve.
• Catalogues. Only SIMBAD, VSX, the TOI list and the TESS eclipsing-binary catalogue were queried. Gaia DR3 variability, ASAS-SN and ZTF were not. An absent label may change once they are.

Records

The run is stored on VDX Core as execution 2 of the project: the 50 candidates are attached as results, each with its plot. The full report, the code and the parameters are versioned together in the project repository. The candidate table can be downloaded below.