Skip to content

Limitations

Selection and coverage

  • The baseline and scaled benchmarks are quality-filtered validation samples, not population-representative surveys.
  • The bounded search covers 250 deterministically selected targets from a pool of 100,347 and cannot support occurrence-rate inference.
  • Scale-up and candidate search use four Kepler products per target even where more products exist.
  • The fixed BLS domain excludes periods outside 0.5–50 days and durations outside the configured grid.

Detection and preprocessing

  • Savitzky-Golay detrending can attenuate or reshape signals when assumptions about variability and duration are violated.
  • BLS favors periodic box-like events and is less suitable for strong transit timing variations or non-box-like signals.
  • Gaps, edge effects, residual systematics, stellar activity, dilution, and harmonics can create or suppress peaks.

Machine learning

  • RF/CNN metrics describe the constructed candidate-level benchmark, not all Kepler light curves.
  • The RF threshold was selected for manual-review recall and is not Bayesian calibration.
  • Model scores reflect similarity to the training construction and must not be read as P(planet | data).
  • The CNN result is a compact research baseline, not an exhaustive architecture or hyperparameter search.
  • CNN early stopping uses each held-out fold as validation, and the RF review threshold is selected on the same out-of-fold predictions used to report its operating-point metrics. Target grouping prevents cross-target leakage but does not provide an independent nested model/threshold evaluation.

Independent validation

  • Segment circular shifts do not reproduce every instrument systematic, stellar process, or contaminating blend.
  • Empirical FAP is conditional on preprocessing, sampling, null construction, BLS grid, and finite draws. It is not a VESPA-style astrophysical false-positive probability.
  • At 1,000 shuffles, the minimum reportable plus-one FAP is 1/1,001.
  • Missing external evidence is not positive evidence. A non-match in TESS, Gaia, or ExoFOP can reflect coverage, sensitivity, cadence, or catalog timing.
  • No new spectroscopy, high-resolution imaging, or pixel-level physical follow-up was performed.

Software and operational constraints

  • Public archive availability and schemas can change after the frozen record.
  • Full runs are storage-, network-, and compute-intensive.
  • Windows has the strongest complete-workstation validation; hosted CI covers the deterministic core and installed wheel on Ubuntu, Windows, and macOS rather than full TensorFlow/MLflow training on every platform.

These constraints are part of the result. They should remain visible in any publication, redistribution, competition entry, or downstream project using SXS outputs.