Agent Lab Vol 3 #04 is an ambient dashboard: one quiet screen with the handful of things that matter today, read on your own machine with no key and no network.
The fixture is invented — including the declaration of what deserved a slot. There is no corpus of a person's calendar, bills, tablets and unread counts to download, and no honest way to make one out of somebody else's life. So 36 profiles, 102 signals, 336 hourly refreshes and every one of the 2,031 declared pairs were written by hand. The declaration is a judgement, not a measurement, and the page says so above every number.
What is measured is what a ranker does with them.
Page: https://dev48.infy.uk/agentlab/vol3-04-ambient-dashboard.html
The ceiling belongs to the screen, not the ranker
More things matter at once than there are slots on 179 of 336 refreshes. So:
| coverage 1.0000 at six slots | unreachable |
| a perfect ranker scores | 0.8577 |
| slots owned by three standing obligations | 2.67 of 6, for the whole fortnight |
If you do not compute that ceiling first, you will spend a week tuning a ranker toward a number no ranker can reach.
Four of the six self-checks cannot fail
Every one of 216 configurations partitions the tokens, so a conservation check is satisfied by construction. Four of the six shipped checks are that kind.
The two that can fail are the interesting ones, and they go red on the better configurations:
every_tile_is_livepasses 336 of 336 while the same run serves 1,400 slot-hours of values over four hours old.
Two null results
- The best ranker misses all 336 stale-backup pairs and all 226 full-disk pairs — because a deadline decay is undefined for a thing with no deadline. It is not a tuning failure; the scoring function has nothing to say about them.
- The rung that is nothing but freshness spends 343 of its 350 weather slot-hours on a signal nobody declared.
Python standard library only. No dependency, no API key, no network. 130 pytest, 6,116 verifier asserts, 157 in-page checks, 0 failures.
Top comments (0)