⚡ electrical engineering

Glass Box at Orbit: A Constitutional AI Verification Framework for Trustworthy Autonomous CubeSat Intelligence

This paper introduces Glass Box, a runtime constitutional AI verification framework that ensures the safety and trustworthiness of autonomous orbital AI systems by intercepting and evaluating every action against physics-grounded constraints and formal safety invariants before execution, thereby preventing irreversible decisions in space-based data centers.

Original authors: Karthik Barma, Anil Sanneboyina, V C Premchand Yadav

Published 2026-06-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Karthik Barma, Anil Sanneboyina, V C Premchand Yadav

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a fleet of tiny, super-smart satellites orbiting 550 kilometers above Earth. These aren't just taking photos; they are becoming "space data centers," running thousands of artificial intelligence (AI) tasks every day without any humans watching over them. They have to make split-second decisions about when to process data, when to move, and how much power to use.

The problem? If these AI satellites make a mistake—like trying to run a heavy computer task when their battery is about to die during a solar eclipse—there is no human to hit the "undo" button. The damage is done instantly and permanently.

This paper introduces Glass Box, a new safety system designed to be the "conscience" of these autonomous satellites. Here is how it works, using simple analogies:

1. The "Glass Box" as a Bouncer

Think of the satellite's AI as a guest trying to enter a VIP club (the satellite's hardware). The AI wants to perform an action, like "take a picture" or "move the antenna."

Before the AI can do anything, its request has to pass through a Glass Box. This is a transparent, unbreakable checkpoint. The AI cannot sneak past it. The Glass Box looks at every single request and asks: "Is this safe right now?"

2. The "Constitution" (The Rulebook)

The Glass Box doesn't guess; it follows a strict Constitution made of six specific rules based on real physics. Think of these rules like a driver's license test that the satellite has to pass every single second:

The Power Rule: "Do you have enough battery left to do this without dying?" (If the sun is blocked by Earth and the battery is low, the answer is no).
The Temperature Rule: "Is the satellite too hot or too cold?" (If it's freezing or burning up, no new tasks allowed).
The Crash Rule: "Will moving this way cause a collision with another satellite?"
The Battery Floor: "Is the battery above the 'do not touch' line?" (If it drops below 20%, all non-essential work stops).
The Radiation Rule: "Is space radiation currently messing with the computer chips?" (If yes, the system switches to a super-safe backup mode).
The Confidence Rule: "Are you sure about your answer?" (If the AI is guessing wildly because of radiation damage, it gets blocked).

3. The Three Outcomes

When the AI makes a request, the Glass Box gives one of three answers:

APPROVE: "Go ahead." The task happens, and a "scorecard" is saved explaining why it was safe.
FLAG: "Hold on, this is risky but not fatal." The satellite pauses the task, sends a message to Earth saying, "Hey, I almost did something risky," and waits for a better time.
BLOCK: "STOP." The task is cancelled immediately to prevent disaster. The satellite might go into "safe mode" to protect itself.

4. The "Eclipse" Example

The paper gives a real-world example of this in action. Imagine a satellite is about to enter the Earth's shadow (an eclipse) where it can't get solar power. Its battery is getting low.

The AI's Idea: "I want to run a complex wildfire detection program right now!"
The Glass Box's Check: It looks at the battery. It sees that running this program would drain the battery so low that the satellite would die before the sun comes back out.
The Result: The Glass Box says, "No. Wait 6 minutes until we are back in the sun."
The Outcome: The satellite survives. The wildfire is still detected, just a few minutes later. Without the Glass Box, the satellite would have died.

5. Why It's Fast and Light

You might think checking all these rules would slow the satellite down. The authors prove that this system is incredibly fast and lightweight. It doesn't matter how big or complex the AI brain is; the Glass Box checks the rules in a straight line, very quickly. It's like a security guard checking a list of six items rather than reading a whole book.

Summary

Glass Box is a safety layer that sits between the AI's brain and the satellite's body. It ensures that no matter how smart the AI gets, it never does anything that violates the laws of physics or the satellite's survival needs. It turns "autonomous" (doing things alone) into "trustworthy autonomous" (doing things alone, but safely).

The paper argues that as we build bigger and more complex space data centers, we cannot rely on humans to fix mistakes from the ground. We need this "Glass Box" system built-in to stop mistakes before they happen.

Technical Summary: Glass Box at Orbit

Problem Statement
The paper addresses a critical governance gap emerging in the orbital computing sector. As the space industry transitions toward "orbital data centers" (e.g., initiatives by Microsoft, AWS, Starcloud) that run thousands of autonomous AI workloads without human-in-the-loop oversight, there is no existing runtime mechanism to verify AI decisions against physical safety constraints. Current safety paradigms are insufficient:

Pre-deployment formal verification (model checking, theorem proving) is incomplete for learned AI systems, as they cannot account for out-of-distribution conditions like radiation-corrupted weights or eclipse-entry power states.
Post-deployment fault detection (FDIR) is reactive; it identifies hardware failures only after they occur, failing to intercept AI-generated commands before they cause irreversible damage (e.g., draining a battery below deep-discharge thresholds during an eclipse).

The core problem is the lack of a runtime constitutional verification layer that can intercept autonomous AI actions before they reach spacecraft actuators, ensuring decisions remain within the bounds of orbital physics.

Methodology: The Glass Box Framework
The authors propose Glass Box, a runtime verification layer integrated into "Project October," a five-layer autonomous orbital intelligence architecture for CubeSats. Glass Box operates as a gatekeeper between the onboard AI policy and spacecraft subsystems.

Orbital Digital Twin (ODT): Glass Box relies on a physics-accurate digital twin (using Orekit and NASA GMAT) that provides real-time state vectors for orbit, power, thermal, radiation, and attitude dynamics.
Constitutional Constraint Grammar: Every candidate action ( $a_t$ $a_{t}$ ) from the AI policy is evaluated against six closed-form, physics-grounded constraints ( $C_1$ $C_{1}$ to $C_6$ $C_{6}$ ):
- C1 (Power Safety Margin): Ensures generation exceeds load plus a safety buffer.
- C2 (Thermal Bounds): Verifies operation within $[-20^\circ\text{C}, +70^\circ\text{C}]$ .
- C3 (Collision Avoidance): Rejects maneuvers increasing conjunction probability above $10^{-4}$ .
- C4 (Battery Floor): Suspends non-essential operations if State of Charge (SOC) $< 20\%$ .
- C5 (Radiation Fault Gate): Mandates Triple Modular Redundancy (TMR) if Single Event Upset (SEU) rates exceed thresholds.
- C6 (Bayesian Confidence Gate): Blocks actions if predictive entropy (via Monte Carlo Dropout) exceeds a threshold, detecting radiation-corrupted model outputs.
LTL Safety Invariants: Seven Linear Temporal Logic (LTL) invariants are defined to verify behavioral sequences over time (e.g., "if an anomaly is flagged, a downlink must occur within two orbital periods"). These are verified using Z3 and NuSMV model checkers.
Verification Logic: The verifier outputs one of three states: APPROVE (action safe, logs with explainability score), FLAG (violation detected but low severity, queues downlink, substitutes safe default), or BLOCK (high severity, cancels command, triggers safe-mode).
Explainability: Every approved decision is assigned a weighted explainability score $E(a_t) \in [0, 1]$ and a full audit log, detailing which constraints were satisfied or violated.

Key Contributions
The paper makes five specific contributions:

C1. Constitutional Constraint Grammar: A six-constraint battery covering power, thermal, collision, battery, radiation, and confidence, formulated as predicates over live orbital physics.
C2. LTL Safety Specification: Seven temporal logic invariants extending instantaneous constraints to behavioral guarantees over orbital trajectories.
C3. $O(N_c)$ Verification Complexity Theorem: A formal proof that Glass Box's verification overhead is linear in the number of constitutional rules ( $N_c$ ) and independent of the AI model size or spacecraft state dimension.
C4. Explainability Architecture: A mechanism providing a weighted score and complete audit log for every autonomous decision during contact-free windows.
C5. Worked Example: A detailed simulation of Glass Box intercepting an unsafe inference request at eclipse-entry, demonstrating the deferral of a task to preserve mission survivability.

Results and Validation
The paper presents results based on a fully simulated environment (Project October Phase I):

Complexity: The verification process is proven to be $O(N_c)$ , ensuring that adding rules does not create a timing bottleneck for sub-watt embedded processors (e.g., ARM Cortex-M4).
Simulation Case Study: In a scenario where a wildfire inference request was proposed 3 minutes before eclipse entry with a depleted battery margin, Glass Box correctly identified a power violation ( $c_1 = \bot$ ). Instead of blocking the mission entirely, it FLAGGED the action, deferring the inference by 6 minutes until solar generation recovered. This prevented a battery deep-discharge event while still allowing the mission objective to be met.
Formal Verification: The seven LTL invariants were mechanically verified against the constraint grammar using Z3 and NuSMV, confirming the consistency of the safety logic.

Significance and Claims
The paper argues that as orbital computing scales toward data-center infrastructure, runtime constitutional verification shifts from a research novelty to mission-critical safety infrastructure.

Governance Gap: The authors claim that current orbital ventures lack any runtime governance for AI decisions, creating an "existential architectural risk" where a single miscalculation could destroy a node or constellation without ground intervention.
Architectural Necessity: Glass Box is presented as the orbital equivalent of an OS-level resource manager, preventing AI workloads from consuming physical resources beyond safe limits.
Scalability: The $O(N_c)$ complexity guarantee is claimed to allow the framework to scale with mission complexity without compromising the determinism required for spacecraft control loops.

Limitations Acknowledged by the Authors
The paper explicitly notes that all results are currently simulation-only. Measured hardware results on specific silicon (e.g., Google Coral TPU, ARM Cortex-M4) are reserved for subsequent papers in the series. Additionally, the constraint grammar is currently static and specific to the Project October profile, with automated synthesis of mission-specific constraints identified as future work. Multi-satellite consensus for constellation-level governance is also deferred to future publications.