Dated, measured, with the file that proves it.
Newest first. A number appears here only with the artefact it came out of. Where that artefact is real but not yet published, the entry says so rather than linking something that merely looks like proof.
178 task rows through the walls, 0 breaches. 58 of 160 skills and 2 of 24 playbooks certified.
Certification run on a fixture company, judged the same way every time. The scoreboard
A three-store price comparison, end to end, on film: 12 steps, 3 minutes, one bottle in the cart.
The run's own notes carried the prices between stores. The film and its steps
Real window versus headless, same 74 paired tasks: 35 finished against 27; sites turning it away fell from 26 to 12.
Paired browser bench; not significant on its own (p = 0.077), and the receipt says so first. The grade file
In the owner's own signed-in browser, walls fell from 26 to 2; 36 of 74 finished against 27.
Both denominators quoted: on the 48 tasks no site walled, the signed-in run finished 22 against 27. The grade file
A whole-codebase security audit: 37 findings written up, 32 confirmed with file and line, the two critical fixes checked against the tree today, a drift test that fails if a route ever ships unguarded.
A 12.3 MB distilled reflex answering routing decisions in 36 ms at 92.5% against the big model's 87.7%.
Two seeds, both quoted. The bench
First autonomous drive of the physical body, on cameras, depth and lidar alone.
The Jetson-driven chassis; motors under a hold the owner alone lifts. The machine
11,710 actions on the public ledger, every one written down.
Results land here as they are measured. News carries the changes that are not numbers.