Constellation Gemini

Inspiration

Language models can explain a recovery plan beautifully while still missing one small rule. In orbital infrastructure, logistics, finance, health, or energy, that hidden mistake can make an otherwise convincing answer unsafe to execute.

Constellation Gemini explores a stricter pattern:

A language model interprets. A specialized engine searches. Independent code verifies. Only verified evidence permits action.

What it does

Constellation Gemini is a proof carrying mission control prototype for a simulated orbital AI network.

A deterministic incident removes a ground station, isolates two orbital compute nodes, and introduces an urgent workload. The operator describes the recovery objective in natural language.

1. Gemini 3.5 Flash interprets the request and converts it into a structured mission intent. If a missing choice changes which plan should win, the workflow pauses for one material clarification.
2. HexStellar Cortex searches the connected combinations. Cortex is not an LLM. It receives an explicit computational contract containing obligations, conflicts, costs, forced choices, and forbidden resources.
3. Independent Python code tries to reject the result. It replays the timeline, recomputes coverage and resource constraints, checks provenance fingerprints, and returns a concrete counterexample when a rule fails.
4. The sandbox remains locked unless the current mission version, canonical input fingerprint, plan fingerprint, and verification fingerprint all match.

The visual experience shows the same threat and clock twice. In the failure view, SAT 07 is struck and mission routes break. In the checked plan, the satellite performs an illustrative avoidance maneuver and workloads are rerouted. The orbital motion is explicitly illustrative. The executable evidence is the formal schedule, Cortex receipt, Python verification report, event log, and downloadable replay bundle.

How we built it

The frontend uses React, TypeScript, Vite, Three.js, and React Three Fiber for the synchronized cinematic globe, timeline, evidence room, and accessible fallback.

The backend uses Python, FastAPI, Pydantic, and a central auditable state machine. Google ADK coordinates the interpretation stage, while Gemini 3.5 Flash produces structured intent in live mode. The same workflow has an explicitly labeled degraded fixture mode so a service outage never becomes a fabricated live result.

The public Cortex HTTPS adapter performs contract analysis before the solve, follows accepted asynchronous jobs without resubmitting them, handles bounded retries, and preserves the assurance returned by Cortex exactly.

The independent verifier does not import the HexStellar client and does not call Gemini or Cortex. A judge can download the replay ZIP and run the same verification offline.

The cloud architecture targets Cloud Run, Pub/Sub, Cloud Tasks, Firestore, Cloud Storage, Secret Manager, and Cloud Logging. The repository also includes Terraform, tests, threat boundaries, claims and limitations, and a judge guide.

Challenges

The hardest challenge was not producing another impressive AI answer. It was maintaining a chain of custody from human intent to structured rules, selected plan, independent recomputation, and sandbox mutation.

We also had to make failure honest. If the requested meaning cannot be represented from available mission state, the workflow abstains before Cortex instead of inventing missing data. If Cortex is unavailable, the UI offers a transparent local deterministic simulation rather than relabeling it as live. If verification fails, the plan cannot apply.

The cinematic comparison was another challenge. Collision and avoidance share the same deterministic clock, debris path, camera language, and encounter point so viewers can understand exactly what changed.

Accomplishments

• A first time viewer can see the problem, the checked correction, and the evidence in one continuous story.
• The clarification is material and changes the canonical mission fingerprint.
• Cortex receives a formal combination contract rather than an open ended prompt.
• Every accepted plan is independently replayed in the original software domain.
• A failed rule produces a specific witness instead of a generic error.
• The full evidence package can be downloaded and checked without network access.
• The release passes 67 Python tests with 90 percent total coverage and 25 frontend tests, plus strict TypeScript, Ruff, ESLint, production build, and replay verification.

What we learned

Responsible agentic action needs separation of duties. Natural language understanding, combinatorial search, verification, and mutation should not be collapsed into one probabilistic step.

We also learned that verification must remain narrow and explainable. In this project, verified means the committed checker passed the declared software rules for this exact deterministic scenario and plan. It does not mean real spacecraft are safe, every mission is understood, or a universal speedup has been proven.

What's next

The next step is to generalize the mission intermediate representation beyond the committed golden scenario, add larger evidence backed workload fixtures, and deploy the complete event driven architecture with durable cloud artifacts. The same proof carrying pattern can then be evaluated in logistics, critical infrastructure, and other domains where a plausible plan is not enough.

Intellectual property and boundaries

Constellation Gemini was created for this hackathon. HexStellar Cortex and its public API and client predate the project and are integrated as an external computational platform. No proprietary Cortex engine, private runtime, internal prompt, activation artifact, or trade secret implementation is included.

Constellation Gemini is a simulated research prototype inspired by public orbital infrastructure research. It is not affiliated with Project Suncatcher and does not control real satellites.

Built With

• cloud-run
• cloud-storage
• cloud-tasks
• fastapi
• firestore
• gemini-3.5-flash
• github-actions
• google-adk
• google-cloud
• hexstellar-cortex
• opentelemetry
• pub/sub
• pydantic
• python
• react
• react-three-fiber
• secret-manager
• terraform
• three.js
• typescript
• vertex-ai
• vite