Hi all,
I am posting a self-contained summary of an unusual transient found in public L1 O4a strain. I am an independent researcher without LVK-side access, and I would appreciate help determining whether this is a known instrumental morphology.
This is not a detection claim, a claim of a new glitch class, or a globally calibrated significance claim. The purpose of this post is to ask what instrumental information may be missing from the public data.
Event summary
| Property | Value |
|---|---|
| Detector | L1 |
| Observing run | O4a |
| Analysis window | GPS 1382955232–1382955264 (32 s) |
| Historical catalogue label | GPS 1382955228, corresponding to the start of the padded crop rather than the analysis window |
| Localized feature | GPS 1382955253.17, or +21.17 s from the analysis-window start |
| Frequency | Approximately 28–30 Hz, primarily within 26–42 Hz |
| Morphology | Sharp impulsive onset followed by several seconds of narrowband ringing, with weaker structure near 50–60 Hz |
The public-strain characterization places the in-band peak at 28.0 Hz. A separate measurement gave 28.4 Hz, so the two determinations agree within 0.4 Hz.
The in-band energy is approximately 304 times the mean of 16 adjacent 32 s windows. Replacing the mean with the median as a sensitivity check gives approximately 315 times. These values are only loudness descriptors; they are not calibrated significances because whitening can react strongly to a transient of this amplitude.
What the anomaly detector selected
The anomaly score is produced by Top-k pooling over spectrogram patches. After correcting the diagnostic described below, the selected patches were found to concentrate on the feature itself rather than diffusely across the window.
The selected region spans approximately GPS 1382955253–1382955264 in the 26–42 Hz band. The feature at GPS 1382955253.17 lies at the beginning of this region, and 20 of the 68 selected patches fall within ±1 s of it.
The corrected detector-aware reconstruction classifies this window as ROBUST, with a native anomaly score of 0.6122661233. In this context, ROBUST is a pipeline classification relative to the detector-specific reference distribution; it is not an astrophysical classification.
Corrections made during the investigation
For transparency, three errors affected an earlier diagnostic interpretation:
- A candidate lookup used
np.isclose(gps, target, atol=0.5)without settingrtol=0. At GPS values near 1.38 × 109, NumPy's default relative tolerance produced an effective tolerance of approximately 13,800 s and silently selected a different event about 3.8 hours away. - The Q-transform axes were transposed. The array is ordered as time by frequency, but it had been interpreted as frequency by time. Both dimensions were 256 pixels after resizing, so this produced no shape error.
- The quoted GPS 1382955228 identified the start of the padded crop. The 32 s analysis window actually begins four seconds later, at GPS 1382955232.
These errors produced the incorrect claim that the selected patches were diffuse and that the approximately 30 Hz feature had not driven the anomaly score. That claim is withdrawn. With the correct event, axes, and time convention, the selected patches concentrate on the ringing feature.
I also tested whether a change in the major version of the analysis library could explain numerical differences. Isolated runs using versions 3.0.13 and 4.0.1 produced spectrogram agreement at approximately 5 × 10−10 and identical scores to six decimal places. The library version was therefore not the source of the discrepancy.
H1 cross-detector checks
A raw public-data H1–L1 correlation calculation gave a maximum of 0.0585 at −11.96 ms. Repeated calculations showed that the maximizing lag can move to the edge of the search interval. The value and lag should therefore not be treated as a stable measurement; the result is consistent with no resolved H1 counterpart.
An earlier stored production coincidence check gave an on-source statistic of 0.0716, compared with a time-shift null mean of 0.1970 and a null maximum of 0.2864 over seven admissible shifts. With so few shifts, I do not assign a tail probability to this comparison. It supports only the limited conclusion that no robust H1 counterpart was resolved.
In the subsequent corrected reconstruction, the opposite-detector data were evaluated regardless of the partner window's class or presence in the candidate catalogue. This event still did not exceed the corrected pooled-null threshold used to select cross-detector follow-ups. There is therefore no new corrected cross-detector support for it.
The pooled-null threshold is a diagnostic follow-up rule. It is not a formal globally trial-corrected significance threshold.
Auxiliary-channel limitations
An earlier test on the public L1 auxiliary subset found a maximum coherence of 0.478, below a family-wise threshold of 0.663 derived from time-shift surrogates. The result reproduced exactly when rerun.
However, the public subset available for this check contains only nine L1 channels and does not include a seismometer or accelerometer. It also does not provide the full seismic, suspension, and ASC witness set.
Consequently, “aux-clean” means only that no coupling was resolved in the tested public channels. It does not exclude scattered-light coupling, mechanical excitation, or another instrumental mechanism visible only in unavailable witness channels.
The corrected follow-up used a 65-event PEM shortlist consisting of 9 ROBUST primary targets and 56 AMBIGUOUS diagnostic targets. This event did not enter that shortlist, so there is no new corrected PEM result to report for it.
Catalogue reconstruction and boundary audit
A later provenance audit identified a boundary-condition defect in the historical frozen catalogue. Each 32 s analysis window should receive 40 s of strain: 4 s of context on both sides, followed by whitening and cropping. For 169 of 10,429 frozen detector–GPS windows, the historical input crossed an HDF5 boundary and supplied only 36 s, omitting the right-hand context.
In a controlled replay using the historical representation and historical detector-specific thresholds, restoring the complete context changed 120 class assignments and reversed 97 routing decisions, all from ESCALATE to ROUTINE.
The subsequent end-to-end reconstruction rebuilt both the detector-aware reference representation and the detector-specific calibration. It produced 10,942 corrected candidates. The comparison contains 10,022 shared identities, 407 historical-only identities, and 920 corrected-only identities. Among the shared identities, 1,626 class assignments differ.
Those broader differences cannot be attributed solely to the padding defect because the reference representation and thresholds changed together. The controlled 169-window replay is the direct evidence for the padding effect; the full-catalogue comparison is descriptive.
For the specific L1 window discussed here, the corrected classification remains ROBUST. It does not exceed the corrected pooled coincidence threshold and is not part of the corrected PEM shortlist.
Other limitations
- The publicly released morphology classifications available to me do not cover O4a. A failed catalogue match is therefore a coverage gap, not evidence of novelty.
- I have not been able to check the complete hardware-injection and detector-state information for this time.
- This is a single observed instance in the search. It cannot establish a new glitch class.
- Its instrumental origin is consistent with the available evidence, but remains an inference rather than a demonstrated identification.
Questions for people with detector-side access
- Is the impulsive-onset-plus-approximately-28 Hz ringing shape a known O4a morphology in L1? It appears qualitatively similar to low-frequency scattered light, but I do not have enough witness information to classify it.
- Is there an Omicron trigger associated with GPS 1382955253.17 in L1, particularly in the 26–42 Hz band?
- Does the full seismic, suspension, or ASC channel set show a witness around this time?
- Was L1 in a normal operating state, or was there any known commissioning activity, hardware injection, or instrumental excursion?
The most defensible description I can currently give is: a loud, L1-local, approximately 28 Hz ringing transient, classified as a strong single-detector anomaly, with no resolved H1 counterpart, but still instrumentally unclassified.
If this is a familiar instrumental artifact, identifying it would be extremely useful. It would clarify both this event and the limits of anomaly searches based only on public strain and a restricted auxiliary-channel subset.
Thanks,
Luca Cirfeta
Independent researcher, Rome