bamfai/bigbounce-anomaly-catalog

BigBounce — Multi-Survey Autoencoder Anomaly Catalog (Paper 3)

Corrected release description. This dataset backs Golden (2026), Paper 3 — A Multi-Survey Autoencoder Anomaly-Candidate Catalog. It is released CC-BY-4.0. The corrected submission release is frozen at the immutable tag p3-v3.1.161. The frozen inventory is heterogeneous and is not a complete six-survey, independently rerunnable per-object product. Consumers should verify downloaded files against RELEASE_MANIFEST.json (per-file SHA-256 + row counts) and observe the provenance restrictions below.

• Paper versions: v3.1.157 (deprecated PRD development variant) / v3.1.161-apjs (review-of-record)
• Author: Houston Golden <houston@hubify.com>
• Code / model weights: https://github.com/Hubify-Projects/bigbounce
• Release state: the corrected manifest, README, machine-readable-table manifest, and revised provenance audit files are published at p3-v3.1.161. This is not a complete six-survey row-level submission package: native Planck and per-object LAMOST products remain absent.

Count interpretation

Quantity	Value	Source
Validated point-source science product	268,319	DESI + SDSS + geometry-gated NEOWISE after 5″ dedup
Point sources + archival Planck continuity count	268,519	scripts/reproduce_headline_dedup.py; appends 200 non-overlapping released Planck cross-transfer rows
Inclusive continuity set (ACT-excluded)	378,280	pathc_unique_objects.parquet, rows with survey_list excluding act_dr6; includes quarantined Gaia and membership-only eROSITA
Filtered continuity set	377,482	377,282 point-source continuity rows (including failed-exploratory LAMOST aggregate membership) + 200 archival Planck rows; excludes ACT/Gaia/eROSITA
Multi-survey coincidence clusters	637	pathc_multi_survey_matches.parquet

Recompute 377,482: load pathc_unique_objects.parquet (378,480 rows) and count rows whose survey_list contains none of act_dr6, gaia_dr3, erosita_dr1 → 377,482. Excise only act_dr6 → 378,280 inclusive.

Recompute 268,519: run scripts/reproduce_headline_dedup.py (in the companion GitHub repo) against DESI, SDSS, NEOWISE, and the released Planck file. The script returns 268,519 because it appends 200 archival Planck map patches with no point-source overlaps. Subtract those 200 rows to obtain the validated 268,319 point-source product.

Contents & schema

Merged catalog

• pathc_unique_objects.parquet — 378,480 unique objects (8-way 5″ positional dedup). Columns: cluster_id, n_detections, n_surveys, survey_list, ra_mean, dec_mean, best_score, member_ids, best_survey. The survey_list column is authoritative for tier derivation (see count table above).
• pathc_multi_survey_matches.parquet — 637 multi-survey coincidence clusters (canonical, ACT-excluded).

Released per-survey blocks

• desi_dr1_anomalies.parquet — 195,829 (canonical-S per-object scores).
• sdss_dr18_pathc_native.parquet — 77,905 continuity slice (canonical-S).
• planck_cmb_anomalies.parquet — 200 patches from the original 20,000-patch cross-transfer scan (patch_idx < 20,000; observed released score range 0.306–62.999). This is an archival diagnostic, not the native 200,000-patch top-200 and not a validated CMB science tier.
• neowise_anomalies.parquet — 436 raw (419 survive the 80° ecliptic-pole mask used in the headline dedup); canonical-S.

Excluded / quarantined tiers (NOT in the 377,482 continuity count)

• gaia_dr3_anomalies.parquet — 500, quarantined synthetic-placeholder fallback from gaia_expanded.py, not real Gaia DR3 data. Exclude from all counts and science use.
• blocks/erosita_dr1/erosita_dr1_anomalies.parquet — 298, membership-list only. S_BigAE axis is irreproducible (could not be reconciled with the committed raw-score artifact on 16 monotone rescalings + 3 IsolationForest retrains). Use the n=298 membership list ranked by committed raw score; do not use S_BigAE as a continuous science axis.
• act_dr6_anomalies.parquet — 200, quarantined cross-transfer diagnostic.

Products not released per object

• LAMOST DR10: no per-object table is present. Its failed-exploratory contribution appears only in aggregate/dedup continuity accounting, so the 377,282 point-source continuity set cannot be reconstructed from a released LAMOST per-survey table.
• Native Planck: the native 200,000-patch top-200 table, checkpoint, input tensor, and full score bank are absent. Native-run summary artifacts support the analysis-level validation, but no native row-level product is public.

ApJS machine-readable submission package (v3.1.161 verified local bundle)

The six scoped tables below were downloaded from the immutable revision into apjs_submission_v3.1.161/ and verified on 2026-07-14 against their Parquet metadata and frozen SHA-256 values. The tracked apjs_submission_v3.1.161/SUBMISSION_BUNDLE_MANIFEST.json records exact byte sizes, rows, schemas, hashes, roles, and exclusions. The Parquet payloads are gitignored; the immutable tag remains their public source. At journal submission, attach these files with units, null conventions, score definitions, selection tier, and provenance metadata.

Submission role	Exact file	Rows	Required column metadata
Validated DESI point sources	desi_dr1_anomalies.parquet	195,829	tid, ra, dec, score, worst, rB, rR, rZ
Validated SDSS continuity slice	sdss_dr18_pathc_native.parquet	77,905	plate, mjd, fiberid, ra, dec, z, class, anomaly_score; state the fixed-size continuity selection
Geometry-gated NEOWISE sources	neowise_anomalies.parquet	436 raw / 419 retained	source_id, coordinates, epoch/time-span fields, W1/W2 variability features, anomaly_score; state the ecliptic mask
Merged continuity catalog	pathc_unique_objects.parquet	378,480	cluster_id, n_detections, n_surveys, survey_list, ra_mean, dec_mean, best_score, member_ids, best_survey; carry tier/quarantine interpretation from this README
Cross-survey coincidences	pathc_multi_survey_matches.parquet	637	same merged schema; define the 5-arcsec union and ACT exclusion
Archival CMB diagnostic	planck_cmb_anomalies.parquet	200	patch_idx, ra, dec, anomaly_score; label cross-transfer and non-native in table title and metadata

The eROSITA 298-row membership list may accompany the submission only as a clearly labeled diagnostic table with its anomaly_score non-comparability warning. The synthetic Gaia and quarantined ACT tables are provenance exhibits, not science tables. No submission metadata may imply that the absent LAMOST per-object or native Planck products are present.

Reproducibility / provenance

• p3_compute_to_accept/ — sixway_dedup.py, sixway_dedup_artifact.{csv,json}, held_out_rescore.py + result — the six-way dedup + held-out rescore stack.
• pta_real_kde_2026-05-01/ — legacy out-of-scope artifact retained in the frozen inventory; it is not part of the v3.1.161 ApJS paper or submission table package.
• pathc_dedup_summary{,_no_act}.json — dedup summaries.
• r42_results/ — DESI OOD MSE plus legacy out-of-scope Bayesian JSON.
• cutouts/ — 827 top-anomaly image cutouts (128px).

Per-survey score-schema flags

Not every block carries the same score schema. Check before treating any score column as a continuous cross-survey axis:

Survey	score_axis	membership_only
DESI DR1	canonical-S	no
SDSS DR18	canonical-S (native rescale)	no
NEOWISE	canonical-S (post ecliptic-pole mask)	no
Gaia DR3	INVALID synthetic-placeholder axis	n/a — quarantined
Planck CMB released	cross-transfer reconstruction score	no — archival diagnostic
eROSITA DR1	NONE — S_BigAE irreproducible	yes

LAMOST has no released per-object block. The unavailable native Planck run must not be inferred from the released cross-transfer Planck rows.

Verification

import pandas as pd, hashlib, json
m = json.load(open("RELEASE_MANIFEST.json"))
for f in m["files"]:
    got = hashlib.sha256(open(f["path"], "rb").read()).hexdigest()
    assert got == f["sha256"], f["path"]
d = pd.read_parquet("pathc_unique_objects.parquet")
excl = {"act_dr6", "gaia_dr3", "erosita_dr1"}
headline = d[d.survey_list.apply(lambda s: not (set(s) & excl))]
assert len(headline) == 377482

This verifies the filtered continuity count, not a uniformly validated catalog. It does not supply the missing LAMOST per-object table or native Planck row-level artifacts.

Citation

Golden, H. 2026, A Multi-Survey Autoencoder Anomaly-Candidate Catalog, Paper 3 (in preparation). Dataset: https://huggingface.co/datasets/bamfai/bigbounce-anomaly-catalog Code: https://github.com/Hubify-Projects/bigbounce