No runs yet. The schedule fires at 08:00.
Under the hood
What is live
The rendering. Every card, row, edge and counter is real DOM built from one script file.
The numbers. Token counts interpolate frame by frame, and cost is computed from published per-million list pricing for Haiku 4.5, Sonnet 4.5 and Opus 4.1.
The controls. Drag the splitter, click a node, open a trace. Your hand always wins over the script.
What is replayed
All of it. One recorded run, r-8814, played back on a deterministic clock. No network, no Math.random, no Date.now. The same pixels every time, which is why it is safe to screen-share on a call.
The visual language is WRITER's, measured from seven captured screens. But every WRITER node in that capture set is green-success, neutral-not-run or blue-selected. There is no failure, degraded, retrying or blocked state anywhere in the reference.
So the amber Chat collector at HTTP 429, the blocked synthesis node, the amber log row carrying the retry line, and the gap note are original work built inside WRITER's design language. Not a recreation of anything WRITER ships. Each amber token is one stated channel rotation away from a sampled mint token, so the derivation can be checked rather than trusted.
What production would need
Five collector agents with their own context windows, dispatched from subagent definition files, plus a Microsoft 365 Graph client with real backoff on 429.
A synthesis gate that refuses to cluster on partial input, and a writer that records what it could not recover instead of interpolating over it.
A human checkpoint that actually blocks the publish step, and a trace store so any row on this page resolves to the call stack that produced it.
Roughly two weeks of build for the pipeline, then the integration work is whatever your tenant demands.