A proposed anomaly-detection method for planetary-defence technosignatures — v28.0, August 2026
A method proposal, and nothing more. Nothing here has been built, run, or measured — no code, no trained model, no dataset, no telescope time, no results. Every number is a target or a published literature value, never an outcome.
The idea: the Vera C. Rubin Observatory's LSST will issue ~10 million transient alerts per night. Broker classifiers built on softmax distribute probability across their known classes, so a genuinely novel event can be assigned moderate confidence spread across familiar categories rather than being flagged as unrecognised. A search for the unprecedented — such as the impact-like signature of an extraterrestrial planetary-defence event — needs a detector that measures unfamiliarity directly.
The proposed detector is a variational autoencoder: it learns to reconstruct known transient light curves, and reconstructs an unfamiliar one badly, giving an unbounded novelty score. Using reconstruction error this way is an established technique (Villar et al. 2021); the proposal is its application to a previously unsearched target class, anchored on the real DART deflection (β = 3.61).
The paper's value is a falsifiable pilot with pre-committed gates: inject synthetic impact-like light curves (across a wide parameter range, not overfit to DART's single example) into real archival survey data, and measure whether reconstruction-error scoring recovers them better than the softmax baseline on identical data. A VAE that doesn't beat the baseline falsifies the claim and the idea should be dropped. Either way the pilot produces something publishable — a measured limit on how well anomaly methods recover impact-like transients, useful to time-domain astronomy regardless of technosignatures.
This repository previously held a document ("VIGIL") written as though the programme had been executed — with preliminary results, an integrated simulation library, validated software, a secured Las Cumbres Observatory Letter of Support, and submitted Gemini/VLT/Keck applications. None of that existed. Those were artefacts of AI-assisted drafting; the document has been withdrawn and purged from this repository's history. This version claims only what an idea can honestly claim, and no external person or organisation has been contacted regarding this work.
See Searching_LSST_for_Deflections_v28.pdf.
Someone who is an observational astronomer or ML researcher, willing to build the injection-recovery test and say whether the reconstruction-error advantage is real on survey data. If it isn't, that answer is worth publishing too.
A companion paper (Planetary Defence Events as Technosignatures) addresses the physics of what would be detectable; this paper addresses how you would search. They are complementary and kept separate.
MIT License — see LICENSE.txt.
Aaron Garcia aaron@garcia.ltd