bshepp/staring-into-the-void-runs

Staring Into the Void — cloud validation runs

Monte Carlo validation artifacts for bshepp/staring-into-the-void, a pipeline applying persistent homology to ensembles of forced-photometry light curves. Each runs/<TIMESTAMP>/ folder contains the NPZ null distribution, a JSON results manifest, the run log, and diagnostic plots from one Hugging Face Jobs run of scripts/hf_jobs/null_sweep.py.

⚠️ Provenance notice (added 2026-07-15)

Both runs published to date (runs/20260424T025543Z, N=1,000 and runs/20260424T033345Z, N=10,000) were intended to run their attenuation sweeps on real ZTF RR Lyrae fetched via ALeRCE, but a bug in null_sweep.py (an oid index-vs-column mismatch) made the real-data path unreachable, and the script silently fell back to synthetic periodic sources. Each run's own .log file records this:

No real RRL data available; falling back to synthetic periodic sources.

Consequences for interpreting these artifacts:

• Null calibrations (NPZ, histograms) are valid. Null ensembles are noise-only synthetic data by design, so the data-source bug does not affect them.
• Attenuation sweeps are synthetic-only. The plots titled "Attenuation sweep on 50 RRL sources" are mislabeled — the 50 sources are synthetic sinusoids, not ZTF RR Lyrae. Do not cite these as real-data validation.
• The attenuation results are additionally unreliable independent of data source: the swept ensembles (150 sources) were compared against a null built from 100-source ensembles, and the statistic family includes H0 summaries that scale deterministically with ensemble size, which guarantees spurious "detected, p=0" outcomes. This ensemble-size artifact is being fixed in the pipeline.

These runs are retained unmodified for transparency. Corrected runs (hard failure instead of silent fallback, data provenance stamped in the manifest, size-matched nulls) will be published alongside them, and this card will be updated when they land.