// see it running
Ghost Architect™ Local a walkthrough of the fleet, the queue, LLM Status, learning and the reports
Demo estate: every screen is the real product, rendered with demo data
Scripts are off in this viewer, so the screens are shown one after another. Open the file in a browser for the slideshow.
Five machines, four model calls running, 215 tokens a second across the estate, and where the next scan goes. Every count names its unit.
Four machines work as one pool: the stacked bands show how many tokens per second each generated over the last 24 hours, under a dashed line for what all four could generate together. One strip per machine shows when it was busy, idle or not answering, and qm-studio-02's 40 minutes offline before dawn stay a gap rather than being filled in.
Nine scans across six modes, four running at once on different machines. One pool, one order, and you set it: Top, Down, or Move to the position you type on any waiting scan, Pause and requeue on a running pass loop, Kill or Cancel on any row. The machine that holds the scan carries out the change and reports back. The page says which pass a scan is in and never claims a phase it cannot see.
Model and quantization, memory held and what is left, context window, decode and prompt-reading speed today against the prior six days, by prompt size, hour by hour, and the model server's own log scrolling live.
A 19 percent fall in decode speed on the week and two failed calls, each with its evidence. On a laptop hub the drift is thermal throttling before anyone notices scans taking longer.
Ghost reviewer reads a finished report and drafts a verdict on each finding it may judge, quoting the code that shows it: real, false, or the wrong severity. Every draft waits for a person to open the code and accept or reject it, and a HIGH finding is left for a person outright.
Priya Shah bundles her verdict into Learning Request LR-1, Marcus Lee approves it, and Dana Ortiz deploys it as billing-api's rule set v1. Every step is named, timed and kept in the learning log, and a rule set can be rolled back.
Health score with its trend, coverage stated in files and passes, findings verified against the source before they are printed.
A sizing estimate that says it is one: what a full scan would read, how long it would take, where to start.
What a change reaches, and what it would take with it.
Where the code disagrees with itself: contracts, schemas, and the seams between components.
Risk before the sprint starts, from the code as it is today.
A Ghost Partner™ profile turns a Points of Interest scan into an evidence review under the partner's own name: each finding states what the code shows and what would settle it, with no score, cost or fix. The provenance inventory is read straight from the files: every copyright holder, license line and "based on" or "copied from" comment, each with its file and line.
Code patterns known to fail as traffic or data grows, each at a file and line, judged in its context, rechecked against the code and cited to the sources that document it. It reads the code and says so: nothing is run or measured.