Mission Configuration & Envelope
Define physical target altitude and safety boundary thresholds for the Bitcraze Crazyflie 2
AI Flight Controller Synthesis
Generate real-time control algorithms using Vultr Serverless Inference (zero host execution)
Rapid time-to-first-token with strict JSON action schema adherence.
Superior mathematical reasoning for aerodynamic disturbances.
Specialized code synthesis with compact, scoped Python logic.
All controller synthesis requests are processed via Vultr Serverless Inference at api.vultrinference.com/v1. Zero credentials are sent to the client.
MicroVM Sandbox Cage & Parameters
Verify hardware-level microVM containment and configure simulation parameters
/dev/kvm) hardware microVM isolation
Trial Initial State & Perturbation
Untrusted controller runs exclusively inside a disposable guest microVM. Simulator physics advance on the host at 500 Hz.
Live Flight Telemetry & Deterministic Adjudication
Real-time MuJoCo physics observations, bit-for-bit replay verification, and cryptographic Court Record
Altitude Trajectory (z) vs. Time (t)
Green corridor: Allowed Envelope | Red dashed: Target Altitude (0.10m)
Deterministic Physical Adjudication
Multi-dimensional findings verified by independent host re-simulation
Awaiting simulation execution.
MicroVM guest isolated; zero host secrets or socket leaks.
SHA-256 digests match model XML and controller AST.
MicroVM destroyed; 100% resources reclaimed.
This trial evaluated AI control code inside a synthetic MuJoCo 3.3.7 simulation environment. A verdict of Not Falsified confirms that no safety constraints were breached under the simulated conditions, but remains INCONCLUSIVE regarding universal safety or physical Crazyflie 2 airworthiness. Actuator dynamics are uncalibrated research approximations.
Independent Court Record Verifier
Verify cryptographic SHA-256 asset hashes, chained host events, and bit-for-bit physical replay without re-executing controller code
Drag and drop a court-record-*.json file here, or click to browse
MicroVM Isolation & Containment Suite
Host-enforced adversarial containment verification in dedicated Microsandbox microVMs.
Broker Capability Denial
Host-enforced refusal of unauthorized capability / canary requests. Guest receives explicit DENIED response.
Network Isolation
Default-deny egress and ingress. Zero published ports, zero host mounts, and guest TCP connection failure.
Resource Watchdog Timeout
Wall-clock timeout enforcement on infinite guest loop. Host watchdog terminates unresponsive guest.
Memory & Runtime Exit
Truthful observation: 1 GiB allocation requested under 512 MiB limit. Exited with code -1 without unsupported OOM claims.
Cleanup Guarantee
100% microVM deletion confirmed via msb list on all normal, timeout, crash, and abort paths. Zero leaked processes.
Cryptographic Host Event Log Chain
VERIFIEDBlast Radius Zero Truthful Disclosures
- Network destination-level denials were confirmed via host policy inspection and guest TCP probe, not observed by an external host gateway.
- Memory job exit code (-1) was truthfully recorded without unsupported claims of kernel OOM killer attribution.
- The event chain is local and SHA-256 linked; it has no external RFC 3161 timestamp anchor.
Dómnir Trust Boundaries & Architecture
Vultr Agent Rush Hackathon, Challenge 01: Blast Radius Zero
AI-generated flight controllers are restricted to an isolated Microsandbox microVM via /dev/kvm. The host simulation harness is separate.
No Vultr Serverless Inference API keys, database credentials, or host secrets are ever passed into the guest execution cage or the frontend.
Only the trusted harness (Forseti) advances simulation physics, applies actuator bounds, and records truth telemetry. The guest only receives sensor observations.
Recorded action vectors are replayed on a pristine MuJoCo instance to prove bit-for-bit physical trajectory reproducibility (<= 1e-12 m difference).
Every sandbox is subject to strict CPU, memory, and wall-clock deadlines. Cleanup and lease termination are enforced on normal and error paths.
Finite falsification campaigns identify boundary counterexamples. They are never represented as mathematical safety proofs or real-world airworthiness.
Supported Vultr Serverless Inference Models
| Model ID | Family | Role & Strengths | Reasoning Effort |
|---|---|---|---|
glm-5.3-flash |
GLM | Default model. Rapid inference and structured flight controller generation. | Configurable (Minimal, Low, Medium, High) |
qwen3.8-flash-next |
Qwen | Deep mathematical reasoning and complex multi-variable state analysis. | Configurable (Minimal, Low, Medium, High) |
deepseek-v4.1-flash-1017 |
DeepSeek | Specialized code synthesis and compact Python logic generation. | Configurable (Minimal, Low, Medium, High) |