אם ירצה ה׳
When the Speed of Light Closes the Loop: Golden Dome and the 5ms Kill Decision
Image: LEO edge node running the 7-check gating matrix. The 5ms clock is not decorative. Image generated by Gemini with RenderLens.
On January 27, 2025, an executive order launched Golden Dome for America: a missile defense system meant to cover the entire continental United States against ballistic, hypersonic, and cruise missile threats. Not a region. Not a corridor. All of it. The price tag runs from $175 billion (White House) to $831 billion (CBO), with some lifecycle projections reaching $3.6 trillion. The Missile Defense Agency has already approved over 2,440 vendors through the SHIELD contract. Anduril Industries, Palantir, Scale AI, and others are building the software layer that fuses sensor data and runs command and control across a constellation of thousands of LEO satellites.
That last sentence is where the hard problem lives. Legacy systems like THAAD, Patriot, and Aegis were designed around a human in the loop: a trained operator watching a screen, making a call, pressing a button. Golden Dome cannot work that way. A hypersonic glide vehicle in terminal phase travels at Mach 5 or above and gives you 5 to 15 seconds from track lock to required intercept actuation. A GEO relay path adds 474 ms of round-trip propagation delay, irreducibly, from physics. Any given LEO sensor node has direct downlink to a ground station for roughly 8% of its orbit. You do the math. By the time a human confirmation request reaches the ground and comes back, the window is gone or was never open. The question is not whether to remove the human from the inner control loop. Physics already did that. The question is what governance structure you put in its place.
My companion paper on Anduril's LatticeOS (DOI: 10.5281/zenodo.19266807) laid out the Autonomous Operations Model: assign decisions to tiers by cognitive budget and task value. Golden Dome breaks that model because it adds a second axis the original didn't have: latency budget. For HGV terminal intercept, the latency budget is under 5 ms of compute, on rad-tolerant hardware, in orbit. My new technical note specifies what that actually looks like: a 7-check gating matrix (schema validation, MIO hash integrity, physics envelope plus cross-modal consistency, temporal coherence, a chaotic-quantized physics-informed neural network seeded by a hardware radiation-event counter, geo-fence and local quorum, and a cryptographic trace commit before actuation), a pre-authorized Mission Intent Object that operators sign in advance on human timescales, and a model swap protocol for surviving political or legal severance of any single AI provider. The falsification suite that accompanies it (simulate.py, DOI: 10.5281/zenodo.19368609) tests four attack scenarios against the matrix and runs 50,000-trial Monte Carlo sensitivity analysis across adversary sophistication levels. Combined detection sits at 99.0% against a nation-state adversary and drops to 89.2% against one with sensor calibration access. If the PINN's hardware entropy assumption fails, that number falls to 61.7%.
None of these numbers are validated. Every threshold in the paper, the 5 ms budget, the 0.95 confidence floor, the 80 m temporal coherence detection threshold, is a proposed engineering default. Phase GD-0 hardware characterization on actual rad-tolerant FPGA hardware has to replace the parametric models with measured failure rates before any of this touches an operational system.
The point of publishing it now is to make the residual risks explicit and the governance dials visible while there is still time to argue about them. The dial between engagement availability and false-positive risk is set by mission doctrine.