Choose an approved task and enter a short non-sensitive request.
THE VPS AS A REQUEST-DRIVEN MACHINE LOCAL-FIRST
One input.
A system of workers.
Ask once. ORBIT classifies a bounded request, queues work, executes on the VPS, and returns a receipt. Browse transportation from human movement to interstellar fiction while models, tests and tasks remain clearly labeled.
ONE OPERATING LOOP
From request to evidence.
Three bounded tasks can progress; model inference is serialized for CPU safety.
Results include evidence labels and test limitations.
Building and public publishing remain controlled by the existing Factory OS.
QUICK ENTRANCES
Choose a mission.
VPS-NATIVE / PERSISTENT QUEUE
Request console.
Run one of five bounded workflows. Each job receives a private random receipt ID; arbitrary server commands and unreviewed deployments are not accepted.
Execution receipt
NO REQUESTSubmit a request to receive a job ID, current state, and a real result. Your job ID is kept in this browser only.
READY-MADE REQUESTS
Start without prompt engineering.
EVERY ENVIRONMENT / EVERY ERA
Transportation Universe.
Original indexed catalog: human and animal mobility, roads, trains, boats, aircraft, orbital craft, proposed machines, and science fiction. Compare real and imaginary ideas without confusing them.
PHYSICS LAB · LIVE CALCULATOR
Forces, propulsion, energy and rocket delta-v.
Eleven parameterized scenarios, comparison view and inspectable formulas.
Loading…Click a record to inspect or compare
REAL CPU / REAL RESPONSES
Model benchmark lab.
Benchmark fixed structured-JSON instructions across installed Ollama models. Results measure time and exact-format compliance—not overall intelligence or commercial frontier parity.
One inference lane. No hidden providers.
On this two-vCPU host, running many models simultaneously would distort results. Independent I/O tasks remain parallel; inference runs serially.
INSTALLED LOCAL MODEL FAMILY
Full capability and reasoning-mode matrix
Loading the model sweep…
Testing roadmap
Expand across installed small models, reasoning modes where supported, task families, sample sizes, repeated trials, structured evaluation, error analysis and resource quotas. A single JSON test is only a baseline.
Open Frontier-12 local chat ↗DETERMINISTIC IMPROVEMENT MATRIX
One topic.
One hundred questions.
The question engine produces 20 evaluation areas × 5 question patterns, each requiring before-and-after evidence. It does not pretend these questions were generated by a commercial LLM.
QUICK START
Choose the subject of your next engineering review.
UNATTENDED DAILY GENERATION
Latest automatic question pack
Awaiting the scheduled worker.
SOURCE-FIRST / YOUTUBE
Videos into experiments.
Register a YouTube watch URL and receive an ingestion checklist and suggested validation tasks. Transcripts, copyrighted footage and video claims are not automatically fetched or fabricated.
Video source intake
No downloads, no API keys, no paid usage
Creator, publication date, transcript provenance and timestamps.
Translate ideas into tasks, modules and measurable experiments.
Use real system receipts before promoting a claim to production.
ENGINE / LIMITATIONS / GOVERNANCE
System evidence.
Current worker counts, queue status and guardrails. Read-only public state excludes the contents of other users’ requests.
Live worker state
Connecting…
Operational contract
Execution — approved templates and fixed handlers only.
Concurrency — three bounded tasks, one local inference lane.
Publishing — existing owner-approved Factory OS remains canonical.
Requests — capped per-day guest workload; no unrestricted shell, private file access or arbitrary code.
Verification — each job has status and a limited receipt, no fake completion claims.
Research — YouTube intake does not imply transcript ingestion.