LIVE INDEX / REV 2026.10.06 / EVIDENCE MOVES
The manifesto asks what a personal AI machine should feel like. This site keeps score as the evidence changes. The index does not shop; it maps.
A number we measured with a stated setup, date and method.
It happened in real use, even if it was not an instrumented benchmark.
The manufacturer says so. Useful input; still somebody else's result or promise.
A reasonable architectural conclusion. Not a direct test.
Plausible or interesting; not trusted yet.
We tried it and it did not behave as expected. Failures stay in the record.
Seven machines / three page types / one evidence language
The lab: web surface, local heavy work, 24 GB VRAM ceiling, unusually strong lived-experience evidence.
The pocket candidate: 300 g-class, dNPU converter dependency, big claims awaiting our parcel.
192 GB cup, modest straw, expandable storage headline that needs translation.
Repairable skin, soldered heart, clustering that is real but still bills in ROCm evenings.
Two memory pools and a placement engine: the software is the spec, if it works as described.
GPU over network, but three different promises: API, SDK and virtual-driver mode.
A memory-first personal AI appliance where editable memory, bounded actions and revocation decide the score.
The live picture moves. The annual report freezes a date.
PROSUMER.REPORT will not invent a second formula for awards. It will freeze this index, methodology and evidence on a stated date and publish the winners that fall out of the same system.