A Multi-Survey Anomaly Search for Unresolved Main-Belt and Inner Solar System Objects

Premise

Single-survey anomalies are cheap. Instrument systematics, photometric pipeline failures, crowding, and geometry extremes all produce per-survey outliers that vanish the moment a second survey is consulted. The discipline of Phase Two was to trust no object that does not survive skeptical follow-up on an independent instrument. The motivating question is narrow and adversarial: which asteroids persistently fail to behave like asteroids across independent surveys?

Method

The pipeline ingests Gaia, NEOWISE, and ZTF photometry, fits per-object photometric models conditioned on geometry and physical priors, and promotes residual outliers through a hard-gated cross-survey filter. Stage-1 consumes photometric history (mag_prev, dt, σ), geometry (α, r, Δ, elongation, phase-scatter), physical priors (H, G, diameter/albedo proxies), enrichment from SDSS, MOVIS, Gaia reflectance, and NEOWISE derived properties, together with survey/band one-hot embeddings. Stage-1 RMSE across all splits is 0.108 mag with no train/val/test overfit. The shortlist, hard-gated set, and strict-residual set are nested: shortlist ⊃ hard-gated ⊃ strict cross-survey residual.

Trusted corpus: Gaia 940,283 rows (99.94%) · NEOWISE 470 (0.05%) · ZTF 108 (0.01%).
Run ID: 20260410T021500Z_noztf_neowise200clean_tranche10 (no hand-selection after fit).

The cross-survey residual flag is load-bearing precisely because the corpus is Gaia-heavy. An object that survives hard-gating despite ZTF and NEOWISE contributing under 0.1% of rows is structurally rarer than its raw rank suggests.

Watchlist — seventeen strict cross-survey residuals

The shortlist of 17 survives multi-survey hard-gating. The top of the watchlist has held stable across multiple reruns — a signal of repeatable structure rather than ranking noise. Cross-rank, final rank, anomaly apparitions, matched MOST images, spectral score, and thermal follow-up state are tracked per object in the working catalogue. Two residuals carry NEOWISE time-series in the active run; the remainder sit in derived_props_only or live_query_no_rows states — not supported by thermal evidence, but not falsified by it either.

005232 — strongest overall

Cross-rank 1, artifact score 29.28, four apparitions, 116 MOST matches. The most dynamically inner object on the watchlist (a = 2.849 AU, e = 0.162, i = 12.22°, period 4.81 yr, H = 12.22, D ≈ 11.56 km, albedo 0.209, taxonomy unknown). Its thermal state is live_query_no_rows: no NEOWISE time-series has yet been resolved. It carries survey_only as its reason code, meaning anomaly strength remains concentrated in a single survey more than is ideal. If forced to pick one object by rank alone, this is the target.

016467 — strongest cross-branch agreement

Cross-rank 3, artifact score 25.42, gaia_slope_z = 2.11. The best candidate when balancing anomaly rank against spectral support (a = 3.143 AU, e = 0.044, i = 15.73°, H = 12.46, D ≈ 21.89 km, albedo 0.057). If forced to pick one object by agreement across diagnostic branches rather than by peak rank, this is the target.

000466 — the biggest late-phase surprise

000466 was a near-miss for several passes. Deeper ZTF image extraction promoted it into the strict cross-survey residual set: cross-rank 4, final rank 16, five anomaly apparitions across two surveys, 287 matched MOST images. It is the largest object on the watchlist — a C-type of roughly 95.5 km diameter (a = 3.352 AU, e = 0.097, i = 19.02°, H = 8.41, albedo 0.092). That does not make it a smoking gun. It makes it the clearest got-stronger-under-pressure case in late Phase Two.

026895 — spectral residual leader

Cross-rank 9, final rank 100, but the highest spectral score in the focused set (1.96, gaia_slope_z = 2.99). Spectrally the most anomalous object on the watchlist, even though its overall anomaly rank is modest. The top spectral-interest objects (026895, 016467, 032523) are not the same as the top overall-rank objects — different diagnostic branches are surfacing different candidates, which is scientifically useful.

000200 — the resolved natural case

000200 had real NEOWISE time-series support (n_obs = 462). After thermal follow-up it moved toward natural_likely with a thermal_optical_mismatch reason code. This is exactly the behavior expected from a skeptical anomaly pipeline: some objects should get less mysterious when more evidence arrives. It is the most important single piece of evidence that the triage layer is functioning correctly — and the reason the remaining seventeen are worth a photon budget rather than a shrug.

Reason codes and structural cautions

survey_only (11/17) · spike_glint_like (7/17) · absolute_offset_pattern (5/17) · band_inconsistency (5/17) · phase_resid_pattern (1/17)

The persistence of survey_only in 11 of 17 residuals is the strongest structural caution on the watchlist. It does not invalidate the objects; it means their anomaly strength remains concentrated in a single survey more than is ideal for a cross-survey claim.

Sensitivity and the pipeline's blind spot

Phase 5 injection-recovery testing shows the pipeline is well-tuned to strong thermal mismatches (100% recovery of injected thermal excess controls) and to strong glint-like transients (37.5% at injection level 3.000), and poorly tuned to weak steady additive optical uplifts (0% recovery at the 0.100 level; 0% for intermittent beacons at 0.300 / 0.100 duty). This limit is stated plainly: the current pipeline cannot see a steady faint additive optical signal, and any public claim must carry that caveat.

False-positive rate estimates on 19 control objects return a stage-1 FPR of 0.158 and an additive-track FPR of 0.000. The stage-1 FPR is the single strongest caution statement attached to the shortlist and must be cited alongside any individual candidate claim.

Promotion criteria

Any of the following would promote a Tier-1 unresolved object to an actionable anomaly claim:

1   Thermal-optical mismatch: NEOWISE time-series places the object outside the expected diameter/albedo locus for its taxonomic prior.
2   Spectral residual inconsistent with any SDSS or MOVIS complex member.
3   Cross-survey glint coincidence: spike-glint-like patterns in both ZTF and Gaia within the same apparition window.
4   Recurrent 3D motif significance after multiple-testing correction (currently q = 0.597 for stable_triangle_shape — not significant).
5   Weak additive optical recovery: any pipeline modification lifting 0.100-level additive injection recovery above zero would justify re-scoring the full shortlist.

Follow-up queue

If the next stage must pick a small set of objects for targeted external characterization, the prioritized list is 005232 (thermal time-series needed) · 000692 (second-strongest overall; weak spectral support) · 016467 (best cross-branch balance) · 000466 (needs independent optical) · 026895 (strongest spectral residual) · 032523 (recurrent across multiple diagnostic branches). Plumbing for MPCORB-based two-body heliocentric propagation already exists in the repository; a 3D unresolved-object product is a packaging layer, not new science.

Conclusion

Phase Two succeeded on its own terms. It did not produce a discovery claim and no object on the current watchlist constitutes compelling evidence of a non-natural source. It produced something more defensible: a reproducible anomaly-search pipeline, a functioning skeptical triage layer, a cross-survey shortlist of 17 objects, one thermally-resolved natural case (000200), one major late-phase promotion (000466), and a clear priority queue. The honest conclusion: yes, there are real anomalies worth follow-up; no, there is not yet a compelling non-natural or technosignature-style detection.

The working console

This report is the bound publication; Artifact-SETI maintains a live console with the shortlist ranking, strict-residual table, injection-recovery tests, orbital context for the top unresolved objects, and reason-code distribution.