The cockpit
One screen to see every artifact, bench, and the warm model's live telemetry — the operator's home base.
Live telemetry rail
Always-on GPU, temperature, unified-memory, and throughput readouts — each with a fixed-window peak-bar chart — so you watch the Spark's envelope while a model runs.
Leaderboard
Live bench-anchored rankings that fold in every chat and compare run, with Spark/OpenRouter source badges — served from a leak-proof public mirror, never your prompts or completions.
Efficiency frontier
Quality versus throughput on one chart with the Pareto skyline in orange — where you decide which quant is worth shipping.
Models browser
Every artifact you can run, filterable by kind and license, one click from chat or compare.
Model detail
Positioning, quant economics with the sweet-spot row, known drift, and a per-model efficiency curve — the full card before you commit GPU.
Chat against any lane
Talk to the warm resident model, an on-demand local GGUF, or a hosted lane — markdown, reasoning, and live tok/s in one composer.
Eval prompts + reference scoring
Pull the exact bench a model was measured on, autofill the composer, and auto-score the answer against gold without leaving chat.
Compare — any vs. any
Duel two lanes side by side with a deterministic rubric score and telemetry-style metric cards — quality, tok/s, TTFT, tokens, cost — each over a session sparkline. Local-vs-local, local-vs-hosted, your call.
Command palette
Hit ⌘K and jump anywhere — fuzzy-search every model, article, and lane, or fire a chat or compare without touching the mouse.
The Lab
A living board of what's shipped, what's next, and what's being explored, with a built-together timeline mined from the commit log.
Guarded lane lifecycle
Launch and tear down serving lanes from the cockpit with a teardown-first confirm, an envelope pre-flight, and an anchor on warm — one resident model in 128 GB, enforced visibly.
Measured benches in the eval drawer
Registered benches replay their measured packets — system contract and reasoning control included — so a row picked in chat scores against exactly what the tracked receipts measured.
Vertical-proof cards on Cortex
A model lane's whole promotion case on one screen: generator preflight, corpus pack with recall gates, routing costs per config, and the publish receipt with its gate chips.
Your first run — the welcome
A dedicated, self-dismissing welcome screen greets your AI Researcher, grounded in your real corpus numbers, with three prompts to try — one of which teaches honest refusal in a single turn. Re-reachable any time from the top bar.
Bring your own cloud key
Your Advisor runs locally with no cloud key. Add an OpenRouter key only if you want to compare against frontier cloud models — the catalog and spend tile stay hidden until you do. Your key is written to a private file on your box; Orionfold never sees it.
Guided console onboarding
One command brings the whole stack up: a preflight matrix, a free-key capture if you need one, a named download manifest, and orientation cards that teach the product while the model pulls — ending on a warm cockpit. curl-to-running, narrated.
