People want a number. This site will give them one, and then immediately show the seven numbers that produced it. Methodology is the contract: capture, intelligence, privacy, battery, wearability, look and value, each 0–10 in half-point steps, each with a written justification that has to name the device so the sentence can be quoted alone.
The total is the weighted sum, out of 10, never stored. It is recomputed every build so a round that drifted from the axes cannot ship. Sorting uses full precision; the page prints one decimal place. Two totals that differ only in floating-point dust are treated as a tie, because otherwise the published tiebreak chain would never run.
What the weights are saying
Capture is the largest share, because a summary of a file that missed the table is a literary exercise. Intelligence is next: whether the transcript and assistant are usable, including a standalone agent invoked from the device. Privacy is on-device processing, a visible recording indicator, export and delete — checkable facts, not a policy novel. Battery is continuous-capture endurance and recharge friction. Wearability is whether you keep it on, and whether you can invoke it on the body with one hand. Look is whether it reads as jewellery or as a gadget in a room. Value is hardware price and whether a usable free tier exists, against what you get.
Those percentages are editorial policy about the category. They live in one place so a page cannot quietly re-weight a favourite. Changing them re-orders the list, which is the point, and forces every device file to be re-read, which is the brake. Readers do not need a version stamp on the homepage. They need the table.
Ties are not coin flips
If two shipping devices match, the order is higher capture, then higher privacy, then lower year-one cost (hardware plus twelve months of the encoded plan), then the slug, so the result is deterministic. Capture first is the same thesis as the weight. Privacy second is the bystander. Cost third is the shopper. Alphabetical last is so nobody has to pretend a 1e-16 difference was a judgement.
The ranking desk itself prints hardware price and the plan label, not that year-one lump. A $189 card with 300 starter minutes is a $189 card. Twelve months of Pro is the plan you buy if you outgrow those minutes, not a cover charge to turn the device on.
Only availability: shipping enters the order. Preorders and discontinued pages stay in the corpus.
| Cost | |
|---|---|
| Price | $189 |
| Plan | Starter 300 min, or Pro $99.99/yr |
| Year-one on the encoded plan | $288.99 if you pay for twelve months of the encoded plan. Not the ranking column. |
| Capture | |
| Microphones | 4 |
| Capture mode | Press to record |
| Audio out | None |
| Battery (active capture) | 30 h — 30 h in Enhance mode at the full 5 m pickup range; 50 h only in Endurance mode, which drops the range to 3 m. |
| Speech recognition | Cloud |
| Language model | Cloud |
| Body | |
| Form factor | Recorder |
| Weight | 30 g |
| Materials | Aluminum alloy & Gorilla Glass |
| Charging case | Not published |
| Recording indicator | Yes |
| Connectivity | Bluetooth, Wi-Fi |
| Companion app | iOS, Android, Web |
| Data export | Yes |
Device pages are the canonical reviews. This Notes section is for arguments that span more than one SKU — battery units, consent lights, form factors. A separate “review” post would compete with the URL that should rank for the product name, and would go stale when a spec moved. <SpecTable /> and rank badges resolve at build for the same reason: a correction should not have to hunt through essays.
If a score is wrong, it is a mistake. The sources are at the bottom of the device page. The axes are on /methodology. The order is arithmetic on top of that. Anything that cannot be explained in that stack should not have a position.