Verification of Impact-Generated Illumination in LSST
Looking for the flash of someone else's survival
Somewhere out there, perhaps around a star we can see on a clear night, a civilisation may have faced what we will one day face: an asteroid on a collision course with their world. And perhaps they did what we did in September 2022—they hit it.
When NASA's DART spacecraft struck Dimorphos, it produced a flash visible from Earth. A brightening of 2.29 magnitudes. A plume of ejecta carrying more momentum than the impactor itself. A signature written in light.
VIGIL asks: if we can do this, and if the physics is universal—and it is—then someone else may have done it too. And if they did, the Vera C. Rubin Observatory might see it.
This is not a claim that we've found anything. This is a framework for looking.
VIGIL is a detection pipeline designed to identify anomalous transient events in LSST data that match the photometric signature of kinetic impactor planetary defence operations. It rests on a simple, testable hypothesis: that a Variational Autoencoder trained on known transient classes will produce high reconstruction error when presented with something genuinely novel—something its latent space was never taught to encode.
The preliminary results are encouraging. On synthetic DART-like events injected into ZTF archival data, the VAE achieves 0.89 ROC-AUC, outperforming traditional softmax entropy by +0.08. At 1% false positive rate, it recovers 83% of injected events.
These are not extraordinary claims. This is engineering.
We have exactly one data point: DART. One successful kinetic impactor experiment in the history of human civilisation. Everything we know about what these events look like comes from a single rubble-pile asteroid struck at a geometry optimised for our observation.
This is the N=1 problem, and we cannot statistics our way out of it.
What we can do is expand the parameter space through physics. The Los Alamos CTH hydrocode library provides 2,000 simulations spanning impact scenarios DART never sampled: failed deflections where β < 1, fragmentation events, oblique impacts, monolithic targets. We're not pretending we know what an alien planetary defence operation looks like. We're acknowledging that we don't, and searching broadly.
| Metric | Value |
|---|---|
| VAE ROC-AUC (preliminary) | 0.89 ± 0.02 |
| Baseline comparison (softmax entropy) | 0.81 ± 0.03 |
| Performance gap | +0.08 |
| Recall at 1% FPR | 83% |
| Domain adaptation recovery (ADDA) | 90% |
| Human review efficiency gain | 3.2× measured |
If VAE reconstruction error shows no improvement over softmax entropy baselines after optimisation, the hypothesis fails. If reconstruction error doesn't correlate with the physical parameters of injected events, the hypothesis fails. If we can't recover 75% of synthetic DART-like events, we stop.
These aren't rhetorical commitments. They're gate conditions. Phase 1 deployment proceeds only if all five validation criteria are met. Currently, four show preliminary pass; one (spectroscopic access) is achieved.
Ten million alerts per night. That's what Rubin will generate. Within that torrent of data—supernovae, variable stars, asteroids, cosmic ray hits, instrumental artefacts—there may be something we've never seen before. Probably not evidence of intelligence. Almost certainly not. The universe is vast and mostly empty and profoundly indifferent to our curiosity.
But the physics that governs a kinetic impactor on Dimorphos is the same physics that would govern one anywhere. Momentum transfer, ejecta dynamics, photometric evolution—these are not local phenomena. They are consequences of conservation laws that hold everywhere we've ever looked.
If someone else has faced what we will face, and done what we can do, the light from that moment is propagating outward at 299,792,458 metres per second. Some of it may reach a 8.4-metre mirror in Chile.
We probably won't find anything. That's fine. A null result after three years of operation still gives us λ_upper < 0.5 events/year within 30 AU at 95% confidence—a publishable constraint on planetary defence technosignatures. Science progresses either way.
But imagine if we did.
- Detection method: VAE reconstruction error for out-of-distribution anomaly detection
- Empirical anchor: DART mission (β = 3.61, Cheng et al. 2023)
- Parameter expansion: 2,000 CTH hydrocode simulations including β < 1 scenarios
- Domain adaptation: ADDA protocol for ZTF → LSST transfer (90% AUC recovery)
- Spectroscopic follow-up: Las Cumbres Observatory (secured), Gemini/VLT/Keck (in negotiation)
- Dual-use science: Natural collision monitoring, methodology development, rate constraints
| File | Description |
|---|---|
VIGIL_v27_White_Paper_MasterPapers.docx |
Full technical white paper (v27.0) |
generate_vigil_whitepaper.js |
Document generation script |
Aaron Garcia
aaron@garcia.ltd
"The universe is under no obligation to make sense to you."
— Neil deGrasse Tyson
But we're under an obligation to look.
Version 1.0.27.0 | January 2026 | PUBLIC