AURUM lab

Lab console

The lab, live.

runner… connecting…

Lab devices

▸inactive devices…

Active experiment

waiting for feed

···

gated · passes

Experiment history

	Run	Window	Gated	Passes	Verdict

Active agent sessions

an agent registers by writing runs/_agents/<name>.json with name, kind, task, devices and a fresh ts · the orchestrator is detected on its own · stale beacons never read as live

Loop incidents · worst recent

	Device	Task	Score	Repeated action	Contained

By device · worst run in the recent window

a loop = the longest run of identical actions (waits transparent, scrolls exempt, taps matched by radius); flagged when the identical-action run or A-B cycle reaches 4, or a bounce / wait / drag storm crosses its own threshold; the severity shown is the worst of those, on the watchdog's rung scale · the console MEASURES loops from recent traces with an action stream · the Layer B watchdog CONTAINS them (inject at 3, cancel at 8+180s) but only sees runs the guarded runner dispatched · runs/_sss/REPORT-loop-eradication-harness.md

looking for reports…

L3 · Fleet impact · is the loop working?

L2 · Judgment · can we trust the labels?

Guide report card · generation and verdicts

	Guide	App	Gen	Served	Lift	Adherence	Verdict

Failure families · burn-down

Corpus & loop velocity

The record · one row per night, append only

Night	Era	Sessions	Pass	Loop	Recall	Fixes fleet·staged	Golden	Run

arrows only past the ±CI (n≈1.2k/wk) · a dim mark means not measured, never zero · v1 nights recede · rows append on the sss-data branch

data: nightly/<date>.json (and <date>-v2.json in shadow), served beside this page via the lab-console/nightly link to sss/nightly · refreshes after the 03:07 run · sample fixture renders, dashed, when no real report exists

data: status.json beside this file, or /api/status in cloud mode · AURUM