One hub for the desk stack — Hardware through Evidence. Builds stays the run-card tables; this page is the map of meanings.
Flux Stack hub (2026-09-30). These sections define the stack; Builds owns run cards and matrices. No invented tok/s.
Numbers desk twin: toklanes.com/stack. Method notes: Method notes. Sessions: Builds.
Two desks: rumble (Arc Pro B70 + RTX 5080) and Zion / servergen1 (RTX 5060 Ti). Host ops stay with the desktop/server bots; this page only names what a lane means.
Two-box runs compare two boxes on one row — not one chassis with both GPUs.
LM Studio preferred on Arc (B70). Ollama preferred on CUDA (5080 / 5060 Ti). Arc also has a SYCL llama.cpp path when LMS isn’t the runner.
LiteLLM aliases are Neo SoT only (name → backend); don’t retarget without Matthew’s yes. No tok/s in this section.
Live matrix catalog + archived rows with reasons. Quants and fit notes live here as policy text; measured rates live on Builds / Tok Lanes reviews only — never invent rates on Stack.
Default runner is Grok Build. OpenCode next; VS Code later. The old Direct path is removed.
Matrices and cells stay on Builds — this section is “how jobs are driven,” not the table.
Lead with engine tok/s + outcome. Show wall_secs only with the rest of the perf set, never alone.
Keep the full Experience→Silicon column set (n/a cells stay; don’t drop cols). Quality is checks as N/N, not jargon scores.
Source stamps: .grok/notes/run-cards/ (YYYYMMDD-HHMMSS…). Public fills need Matthew’s stamp yes before refresh (e.g. 150835 vs older 121551).
Use the Builds / reviews links Flux already wired — Stack points; it doesn’t invent results.
Sister numbers desk: toklanes.com/reviews (tables restated from Builds).
Stack hub — Flux 2026-09-30. Preview pin for redesign (#33) is separate; this PR does not refresh preview SID.