TraviaTechPie Review

TL;DR — Apple published an eleven-page privacy paper for Audio Intelligence, and it is the only document that tells you two things worth knowing before you buy. Live Rewind is not really a watch feature: press the crown with your phone out of range and the transfer “will fail with an error,” and the fifteen seconds are discarded. And activating it fires a chime from the watch speaker that plays even on silent, even with headphones connected. The paper is also where you learn that raw audio — not a transcript — is what crosses from your wrist to your phone for Siri Recap. None of it is hidden. It is just eleven pages deep instead of on the page you were reading.

The Story

The interesting question about a wrist microphone that can replay the last fifteen seconds is usually is the thing always on? I assumed Apple had dodged it, because the product page and the support document both dance around the word “always.”

Apple did not dodge it. It answered in a document called the Audio Intelligence Privacy Overview, dated September 2026, eleven pages, linked from nowhere you would naturally land. The answer, verbatim:

> “The S11 chip uses a lightweight AI model to determine whether speech is occurring nearby. This model doesn’t transcribe or record anything. It simply detects whether a conversation has started. If a conversation is detected, audio flows into a protected buffer inside the Secure Exclave on the S11 chip.”

So: yes, something is listening continuously, and what it is doing is a binary speech/no-speech decision that never becomes text. That is a clear, specific, falsifiable architectural claim, and it is more than most of this industry offers.

Which makes the rest of the paper the story. Because while Apple was answering the question everyone asks, it also documented several things nobody has asked, and those are the ones that change whether you should buy the watch.

What’s Actually in the Box

Apple’s footnote, which appears identically on the spec page, the product page, and both newsroom releases, defines the bundle:

> “Audio Intelligence includes Live Rewind, Siri Recap, Sound Recognition, and faster Shazam and requires Apple Watch Series 12 or Apple Watch Ultra 4.”

Four things, not three, and one of them is not new:

• Live Rewind — double-press the Digital Crown, get the last 15 seconds as text. You can save it or ask Siri about it.
• Siri Recap — rolling summaries of the conversations you have during the day, in the new Siri app. They auto-delete after seven days unless you save them.
• Sound Recognition — detects “important sounds like sirens, alarms, and doorbells,” and notifies you “even when your iPhone is not with you.”
• Faster Shazam — Shazam has been on Apple Watch for years. What changed is that it “instantly detects the music playing around you” and populates the Smart Stack widget “all without a tap.”

Note the shape of that last one. Same feature, converted from something you invoke into something that runs on its own. MacRumors ran it under the headline “Three All-New ‘Audio Intelligence’ Features,” which is understandable and wrong — it is four, one of which is an old feature that stopped waiting to be asked.

The hardware requirements are not uniform, and the split is the first clue about what runs where:

Every feature lists an iPhone, but they are not the same kind of requirement. Sound Recognition and Shazam need a phone that is merely paired — an iPhone 11 clears the bar, and Apple explicitly says Sound Recognition alerts you when the phone is elsewhere. Live Rewind and Siri Recap need an Apple Intelligence phone, and the cutoff has one oddity: Apple’s list includes the iPhone 17e but leaves out the 16e.

Older watches running watchOS 27 get none of it. This is a hardware gate, not a software one.

Live Rewind Is an iPhone Feature Wearing a Watch

This is the part that should be in every review and is in none of them.

Apple’s support page does list iPhone models as a requirement, so the dependency is not concealed. What no page outside the privacy paper tells you is what happens when the phone is not there:

> “After activating, the immediately preceding 15 seconds of buffered audio is sent from the Secure Exclave of the Apple Watch to iPhone. If iPhone is not within wireless range of Apple Watch, the transfer will fail with an error, and the audio will be immediately discarded.”

The watch holds the buffer. The watch does not do the transcription. Your phone does. Press the crown on a run, on a walk, in a meeting you left your bag for — anywhere your phone is not — and the feature does not degrade gracefully. It throws an error and deletes the fifteen seconds you were trying to recover.

Compare that to Sound Recognition, which the same document says works “even when your iPhone is not with you,” and which for the deaf and hard of hearing is the feature that matters most. Apple built one Audio Intelligence feature that is genuinely independent and one that is not, and the independent one is the accessibility feature. That is the right way round, and it is worth saying so.

But if you are buying a Series 12 for Live Rewind, understand that you are buying a two-device feature. The watch is the microphone and the display. The phone is the computer.

Raw Audio Does Leave Your Wrist

Here is Apple’s public privacy claim, the one on the marketing pages:

> “These features do not create or store audio recordings, and raw audio used for processing is completely inaccessible, even to Apple, and is deleted after processing.”

Every word is true. And almost everyone will read “inaccessible” and “deleted after processing” as “it stays on the watch and dies there,” which is not what it says. From the paper:

> “For Siri Recap, the Secure Exclaves on Apple Watch and iPhone establish a secure, encrypted channel during the secure pairing process. Then, raw audio is encrypted and transferred from the Secure Exclave on Apple Watch to the Secure Exclave on the paired iPhone.”

Raw audio — not a transcript, not a summary, the audio — travels from your wrist to your pocket every time Siri Recap decides a conversation is happening.

Apple’s support article does mention the hop, and it is worth being precise about how. It says “the Secure Exclave on Apple Watch encrypts the detected audio and transmits it to the Secure Exclave on the paired iPhone.” Detected audio. The paper is the document that says raw, and that describes the encrypted channel, the pairing mechanism that constrains where it can go, and what happens when the transfer fails. Three documents, three degrees of specificity, and the product page — the only one most buyers will ever open — carries none of it.

To be clear about the engineering: this is well built. The audio is encrypted end to end between two hardware compartments neither operating system can address. Apple added an audio-verified pairing mechanism on top of Bluetooth specifically so that “ambient data from the Secure Exclave on Apple Watch can only be sent to the Secure Exclave of the paired iPhone, and no other device.” The keys are device-bound, time-bound, rotate, and expire — and if the transfer cannot complete, “the encryption keys expire and the encrypted audio is permanently deleted.” Find My invalidates the pairing keys immediately.

That is a better design than “it never leaves the watch” would have been, because the watch does not have the silicon to transcribe speech well and Apple chose not to pretend otherwise. But it is a different sentence from the one on the product page, and the difference is the whole point of reading the paper.

The Sentence Written for a Courtroom

In the introduction, this:

> “Audio Intelligence features do not create audio recordings that can be accessed by the operating system, apps, the user, or Apple. There is no recording to share, forward, or produce if requested by any party, because no recording exists.”

Read the second sentence again. “Produce if requested by any party” is not consumer language. That is discovery language — the vocabulary of subpoenas, warrants and civil litigation. Apple is telling courts, in a document aimed at customers, that there is nothing to compel.

It is a legitimate and in my view admirable thing to design for. Architecture beats policy: a company that cannot produce your audio does not have to be trusted not to. But it is striking to see it stated that plainly, and it tells you which risk Apple’s lawyers were thinking about when this feature was specced.

The Chime You Cannot Turn Off

I had assumed Apple shipped no bystander mechanism. The opposite is true, and the mechanism is unusually aggressive:

> “When activated, an audible chime plays from the speaker on your Apple Watch, even if your Apple Watch is on silent or you have headphones connected. A full-display animation and microphone indicator appear on the Apple Watch display, and a distinct double-press gesture is required to activate Live Rewind.”

Silent mode does not suppress it. Routing audio to headphones does not suppress it. Apple deliberately removed every ordinary way a user could quiet their own device, in order to make the feature impossible to use discreetly.

Think about what that means in practice. You cannot Live Rewind a lecture without the room hearing it. You cannot Live Rewind in a meeting without your colleagues knowing. That is a real product constraint, chosen on purpose, and it is the single most consequential fact about how this feature will feel to own — and it appears on page nine of a PDF rather than anywhere near the buy button.

Siri Recap took the opposite approach and is the weaker case for it. It runs in the background with no chime and no indicator. Apple’s argument is that the output is not a recording and not a verbatim transcript:

> “By design, Siri Recap does not create a recording, does not produce a verbatim transcript, and does not identify and attribute speakers. The output is a brief, high-level summary, comparable to notes a person might write after a conversation.”

That is a reasonable defence and it is not nothing — a summary of a conversation is categorically different from a recording of one. It is also still a machine-generated account of what the people around you said, produced without their knowledge, and in two-party-consent jurisdictions the legal analysis of “notes a person might write” has not been done yet.

Where Your Conversation Actually Goes

The paper includes a stage-by-stage map for Siri Recap, which is worth compressing:

1. Speech detection — S11 Always On Processor. Raw audio “processed on-device and discarded in real time.” 2. Buffered audio on the watch — Secure Exclave, inaccessible to watchOS. 3. Buffered audio on the phone — encrypted in transit, decrypted only inside the iPhone’s Secure Exclave. 4. On-device transcription and condensing — speech recognition plus an on-device language model that produces “a condensed version that is less than half the length of the original transcript,” stripping tone, filler and redundancy. Then “the raw audio is immediately and permanently deleted and no longer exists on either Secure Exclave.” A safety model screens for harmful terms. 5. Private Cloud Compute — the condensed text is encrypted and sent up, “subject to daily usage limits.” Apple Foundation Models write the title and key points. “The results are immediately deleted from Private Cloud Compute.” 6. The Siri app — end-to-end encrypted across your devices, seven-day auto-delete unless saved.

Two things in that chain deserve flagging.

Apple also sends context alongside the text: Now Playing data, calendar data, “high-level location labels such as home, work, and school,” city and state, and point-of-interest categories like “grocery store” or “park.” Precise location is excluded. That is a defensible list for improving a summary, and it is more metadata than “an encrypted transcript” implies.

And the retention question I thought was unanswered is answered — the summarization “is then encrypted and sent back to iPhone. The results are immediately deleted from Private Cloud Compute.” It just isn’t answered anywhere a buyer will read.

Sound Recognition Is a Six-Year-Old Accessibility Feature

Sound Recognition is the oldest thing in the bundle. It shipped in iOS 14 in 2020, buried under Accessibility, detecting alarms, sirens, doorbells and crying babies for people who cannot hear them. What is new is that the watch does it natively, in the Secure Exclave, without an iPhone.

Google’s Sound Notifications also arrived in 2020, with the flashlight strobe and an on-screen alert as its main outputs and the option to relay to a compatible Wear OS watch. Either way the detection ran on the phone and the watch was a display. Apple’s runs on the watch. On the underlying capability Apple is late; on this specific implementation it is not behind.

Two practical notes. Samsung’s Hearing Health measures ambient noise levels for hearing protection — dosimetry, not semantic recognition, so it is not a comparison. And Apple’s long-standing guidance for Sound Recognition on iPhone is blunt: “Don’t rely on your iPhone to recognize sounds in circumstances where you may be harmed or injured, in high-risk or emergency situations, or for navigation.” That warning is in the iPhone user guide. The Apple Watch support article for Audio Intelligence carries no equivalent line, and no accuracy or false-positive rates have been published for the watch implementation. If you are deaf or hard of hearing and evaluating this as a safety system, that is the most important paragraph in this article — and the absence of the warning on the watch page is not a reason to assume it does not apply.

The Part Where It Costs Money

Buried in the newsroom footnotes:

> “Certain Audio Intelligence features that rely on server-side models are subject to daily usage limits, including but not limited to Live Rewind and Siri Recap.”

And in a separate footnote, this one covering Apple Intelligence and Siri broadly rather than Audio Intelligence specifically:

> “Expanded access to such features will be available for a fee in the future.”

The privacy paper independently confirms the limit — the Private Cloud Compute stage is described as “subject to daily usage limits” — without quantifying it either. So the cap is real, it is documented in two places, and its size is stated in neither.

Credit where it is due: Apple disclosed a paywall in advance rather than introducing one quietly later. But if you are buying a Series 12 partly for Live Rewind, you are buying a feature whose free tier has an unstated ceiling.

What This Means If You Wear One

• Live Rewind needs your phone in range, every single time. No phone, no feature, and the audio is discarded rather than queued. If you wanted this for phone-free runs, it does not do that.
• It also makes a noise you cannot mute. Chime plus full-screen animation, overriding silent mode and headphones. Discreet use is not a supported scenario, by design.
• “Raw audio is inaccessible” is true; “raw audio stays on the watch” is not. For Siri Recap it moves to your phone, encrypted, Exclave to Exclave. Good architecture, different claim.
• Siri Recap sends more than a transcript. Condensed text plus calendar, Now Playing, and coarse location context go to Private Cloud Compute.
• Do not treat Sound Recognition as a safety system. Apple’s own guidance says not to, and no accuracy data exists for the watch version.
• The other people in the room get a chime for Live Rewind and nothing for Siri Recap. Apple thought hard about bystanders in one feature and made a policy argument in the other.

The Takeaway

Three documents describe this product, and they get more honest as they get harder to find. The product page says raw audio is inaccessible and deleted after processing, and leaves you to infer that it never moves. The support article admits the hop and calls what moves “detected audio.” The privacy paper calls it raw audio, names the channel, and is the only one of the three that tells you Live Rewind dies without your phone and announces itself with a chime you cannot silence. It is also the one at a URL you will never stumble onto.

That is the pattern worth naming, because it is now the third time in one launch — after the Readiness score with no published validation and a 60x sampling claim with no denominator — that the interesting part of the story was a document rather than a feature. The difference this time is that the document exists and is good. Apple did the hard engineering, wrote it down honestly, and then filed it where nobody making a purchase decision will find it.

If a company is going to build ambient microphones into a wrist computer, this is close to how you would want it done. It would just be better if the eleven pages describing it were the thing you read before buying, rather than the thing you find afterwards.

Source: Apple — Audio Intelligence Privacy Overview (September 2026) · Apple Support — Audio Intelligence on Apple Watch

Photo: Luke Chesser / Unsplash

TL;DR — The Series 12’s headline sensor claim is “60x more frequent background heart rate readings than Series 11.” Apple has never published how often the Series 11 took a background reading, so the number has no denominator you can check. Apple also describes the physical change in exactly one recurring phrase — “larger, power-efficient green LEDs” — publishes no component counts, and quietly dropped the generation number it used to print on the optical sensor. Meanwhile its own white paper says the five-second readings you see on the watch face reach HealthKit only every 30 seconds. The hardware almost certainly did improve. Apple just replaced the description of it with a ratio.

The same S11 chip runs the watch’s new ambient microphone features, and Apple documented those far more thoroughly — the Audio Intelligence privacy paper spells out what the mic is doing.

The Story

There is a specific way to read a launch page, and it is to notice which claims are structural and which are relative.

A structural claim tells you what a thing is: this sensor has these emitters, at this wavelength, in this arrangement. You can verify it, or at least someone with a screwdriver can. A relative claim tells you how a thing compares to something else: 60 times more often, 25 percent longer, twice as bright. It is verifiable only if both ends of the comparison are public.

The Series 12’s sensor is described almost entirely in relative claims. That is what makes it worth a careful read — not because Apple is lying, but because the shape of the disclosure tells you what Apple wanted the conversation to be about.

One Sentence of Hardware

Here is essentially everything Apple has said about what physically changed, from the newsroom release:

> “With larger, power-efficient green LEDs on the optical heart sensor, Apple Watch Series 12 now measures heart rate every five seconds, all day long.”

The product page adds a second formulation of the same idea — “Redesigned with more capable optical and electrical sensors and larger, more power-efficient green LEDs, the sensing system unlocks deeper health and fitness insights” — and the tech specs reduce it to a list: “second-generation electrical heart sensor, always-on optical heart sensor, blood oxygen sensor, and temperature sensor.”

That is the complete public description of the hardware. Larger green LEDs, more power-efficient, second-generation electrical sensor. No count of emitters. No count of detectors. No wavelength figures, no sampling architecture, no arrangement.

Note also how the credit gets distributed. In the sentence that introduces five-second sampling, the cause is optical: with larger, power-efficient green LEDs, the watch now measures every five seconds. A few paragraphs later the silicon gets it instead — “With the new Health Sensing System and S11 chip, Apple Watch Series 12 delivers higher-frequency heart rate and heart rate variability (HRV) measurements.” The same chip is also credited with more accurate step counts and with the new Audio Intelligence features.

Both components are named, and nowhere does Apple say which one made the frequency increase possible, or in what proportion. That is not a contradiction. It is the absence of a claim — and it happens to be the one claim that would tell you whether this is a sensor upgrade or a compute upgrade, which is the same question as whether any of it could ever reach a Series 11 in software.

60 Times More Than What

The number everyone repeated is this one, from the product page:

> “60x more frequent background heart rate readings than Series 11”

Two things about it.

First, it is not in the newsroom release. The press release contains the 24x figure for HRV; the 60x figure for background heart rate lives on the marketing page only. If you see it attributed to Apple’s announcement, that attribution is wrong.

Second, and more importantly: Apple has never published how often the Series 11 read your heart rate in the background. Not in its specs, not in its support documents, not in the developer documentation. The Series 12 end of the ratio is public — every five seconds. The Series 11 end never was.

So the claim is unfalsifiable in the plain sense of the word. You cannot check it, and neither can a reviewer, because one of the two numbers required to check it has never existed publicly. You can work backwards and infer that the Series 11 must have sampled around every five minutes, but you are reconstructing Apple’s denominator from Apple’s ratio, which is not verification. It is arithmetic performed on an assumption.

This is not unique to Apple, and it is not fraud. It is just a category of claim that reads like a specification and functions like an advertisement.

Apple Stopped Counting

Here is a small detail I find more telling than the big number.

On Apple’s own comparison page, the Series 11’s optical sensor is listed as “Third-generation optical heart sensor.” The Series 12’s is listed as “Always-on optical heart sensor.”

The generation number is gone. Not incremented — removed. Apple counted to three across previous models and then, on the generation it describes as the biggest sensing change in years, stopped counting and substituted a capability word.

Meanwhile the electrical sensor did get a number: “second-generation.” So this is not a house-style change. Apple is still willing to number a sensor. It stopped numbering this particular one at the exact moment it redesigned it.

I do not know why, and I am not going to pretend the reason is sinister — “always-on” is genuinely the more meaningful descriptor for a buyer, and generation numbers are a weak signal anyway. But it is a strange place to lose your counter, and it removes the one remaining anchor a reader had for placing this sensor in a lineage.

Five Seconds on the Screen, Thirty in the Database

This is the part almost nobody covered, and the source is Apple itself.

Buried in the methodology of Apple’s heart rate accuracy white paper is an explanation of why Apple needed an internal logging tool to read its own watch during the study:

> “This data is a single value (not an average) reported every 5 seconds — in watch face complications for background heart rate, and in the Workout app for workout activities. The internal app, which logs the same data surfaced to user, was required because background data is only published to HealthKit every 30 seconds to manage device memory.“

So there are two resolutions, not one. The watch face and the Workout app show you a value every five seconds. HealthKit — the database every third-party app reads, the thing your data exports from, the layer any other developer builds on — receives background heart rate every 30 seconds.

That is a six-fold gap between what the watch displays and what the watch stores. Both numbers are real, and Apple’s reason is a plain engineering one: writing a sample every five seconds all day would eat memory. Fine. But it means the 60x improvement is 60x at the glass, not throughout.

If you use a third-party training or sleep app, the resolution it can see is the 30-second one. If you export your health data, that is what you get. The higher-frequency stream feeds Apple’s own features — Readiness, Recovery HRV, the vitals view — and reaches everyone else downsampled.

I want to be careful here: 30 seconds is still an improvement if the Series 11 wrote less often than that, and Apple hasn’t told us what the Series 11 wrote either. The point is not that Apple degraded anything. It is that “60 times more often” is a statement about a display layer, and the number that governs your actual data is a different, lower one that appears only in a methodology paragraph of a PDF.

What People With Screwdrivers Think They See

At the time of writing no retail unit had been torn down — iFixit had not opened one. What exists are hands-on observations from the announcement event, and they are worth reporting precisely because they are the only structural information available — and because they are frequently misquoted.

Engadget’s hands-on describes “[…] a ring of 12 lights compared to the four LEDs laid out in a diamond shape on last year’s device.” That is a count of emitters, made by eye, at a press event.

The 5K Runner describes “[…] eight windows with double photodiodes in each, roughly sixteen photodiode elements in total.” That is an estimate of detectors, hedged with “roughly,” and not sourced.

These two numbers get cited against each other as though the sources disagree. They do not. One is counting lights and the other is counting sensors, and they are different components. What they have in common is more interesting than the discrepancy people imagine: Apple published neither figure. Not the LED count, not the photodiode count, not the arrangement. Every structural number in circulation about this sensor originates outside Apple.

The same applies to two other claims I would avoid repeating. That the photodiodes are arranged in a ring is a press description, not an Apple statement. That the Series 11 used infrared LEDs for background readings and the Series 12 switched to green appears in one outlet and nowhere in Apple’s materials — Apple says “green LEDs,” and says nothing about what the Series 11 used.

What This Means If You Wear One

• The sensor change is real; the size of it is not public. Apple confirmed larger, more efficient green LEDs and a second-generation electrical sensor. Everything more specific than that is somebody’s estimate.
• Treat 60x as marketing, not as a spec. Not because it is false, but because the number it is measured against has never been published. There is nothing to check it with.
• If you rely on third-party health apps, your resolution is 30 seconds, not 5. That is from Apple’s own white paper. The five-second figure describes what the watch shows you.
• Wait for the teardown if the structure matters to you. Every component count you have read so far was made by looking at a demo unit across a table.

The Takeaway

Apple made a defensible engineering decision and then described it in the least checkable way available.

You can see the logic. Component counts invite comparison shopping against Samsung and Garmin on a spec axis Apple does not want to compete on, and they age badly. A ratio — 60 times, 24 times — is vivid, is probably true, and is inert. Nobody can argue with it, because nobody can compute it.

What you are left with is a sensor that is almost certainly better and is described almost entirely in numbers you cannot verify, sitting under a score that has no published validation, measured by a study that Apple designed and ran itself. Each of those three links in the chain is individually reasonable. It is only when you lay them end to end that you notice there is no point along the way where an outsider gets to check the work.

Source: Apple Newsroom — Introducing Apple Watch Series 12 · Apple Watch Heart Rate Accuracy Study, September 2026

Photo: Nik / Unsplash

TL;DR — Apple’s nine-page heart rate study is better than it had to be. It enrolled 1,460 people at five sites, used a chest-strap ECG as reference, and reported its own losses. It also discloses three things worth sitting with: Apple read its own watch through an internal logging app no competitor was given, Apple chose how Samsung’s untimestamped background data would be scored, and the Daily Living protocol — the part underpinning the “60 times more often” marketing claim — had no pre-specified sample size at all. None of that makes the result wrong. It does change what the result means.

For the hardware behind those numbers, see what Apple did and did not disclose about the sensor.

The Story

Apple published one new health white paper for the Series 12 launch, and it is not the one the four new algorithms needed. It is a heart rate accuracy study, nine pages, dated September 2026.

The headline is easy to repeat: across 159 activity-by-measure comparisons against six competitors, Apple Watch was more accurate in 139.

The paper is worth reading properly, though, because it is a genuinely unusual document. Most vendor benchmarks are written to be un-checkable. This one hands you the knife. Apple wrote down its exclusion criteria, its statistical model, its data quality thresholds, the exact reason it needed a special tool to read its own watch, and the two comparisons it lost. A company writing pure marketing does not include the sentence “the comparison device’s accuracy was higher in 2.”

So the interesting question is not whether Apple cheated. It is narrower and more useful: given exactly what Apple describes, what does the 139 actually measure?

The Setup

Participants wore an Apple Watch on one wrist and a competitor on the other, with a Polar H10 chest strap ECG as the reference standard. Which wrist got the Apple Watch was randomized, so the dominant hand carried it about half the time — a small detail that matters more than it sounds, because wrist motion is one of the main things that breaks optical heart rate.

The comparison set: Garmin Forerunner 970, Google Pixel Watch 4, Huawei Watch 5, Samsung Galaxy Watch 8, WHOOP 5, and Oura Ring 5 on an index finger. Those brands account for roughly 90% of the worldwide smartwatch category, according to Omdia shipment data cited in the paper — not Apple’s own estimate.

Activities were chosen to break things on purpose. Outdoor runs, treadmill intervals, outdoor cycling, HIIT, strength training, 15 to 20 minutes each. Apple names the failure modes it was hunting: cadence lock in running, loaded forearm grip in strength work and cycling, rapid wrist motion and fast heart rate transitions in HIIT. There was also a separate Daily Living protocol — rest, desk work, meal preparation, walking, about 20 minutes each — to test background heart rate.

Five sites: two near Cupertino, plus San Diego, Austin, and Selangor, Malaysia. Apple ran Cupertino and Austin itself; a third-party contract research organization ran San Diego and Selangor, and 59% of the analyzed participants came from those CRO sites.

Of 1,460 enrolled, 1,254 contributed at least one paired measurement. Twenty percent had Fitzpatrick skin tones V–VI, which is the demographic optical sensors historically fail and wearable studies historically underweight. Mean age 39, mean BMI 26. Thirty-five percent female — described as “representation across sex,” which it is, though it is a roughly two-to-one male skew.

One exclusion criterion is worth quoting because nobody writes this by accident. Alongside beta blockers, equipment allergies, unstable cardiovascular conditions, and tattoos or large moles at the sensor site, Apple excluded people working in tech media.

Everyone Got a Different Pipe

Here is the first thing that changes how you read the result.

Apple could not just ask each device for its data, because each manufacturer exposes data differently. So Apple picked a method per device. For most competitors it used a Bluetooth Low Energy stream from the manufacturer’s own heart rate app into a logging app on a separate iPhone — because, as the paper puts it, “not all manufacturers write high-fidelity data to HealthKit, or permit such data to be exported from their companion apps.”

For the Apple Watch, Apple used an internal data logging application.

The stated reason is specific and checkable: background heart rate is only written to HealthKit every 30 seconds, to conserve device memory. But the value the watch actually shows you — in a watch face complication, in the Workout app — is a single reading every 5 seconds. To evaluate the number the user sees, Apple needed a tool that captures the number the user sees. HealthKit would have understated its own watch.

That logic holds. It is also true that only one company in this study had the ability to build that tool.

And this is where the framing of the whole paper matters. Apple states its selection rules up front, and there are three: the data stream had to be visible to the user in the product experience, it had to carry timestamps, and where a device offered more than one qualifying stream, Apple took the highest frequency available. The second and third rules are the ones that quietly do the work — a device that surfaces a number but will not tell you when it was taken cannot be scored the normal way, and a device that exports a slower stream than it displays gets scored on the slower one. The first rule is a deliberate and defensible choice — arguably the more honest consumer question is “how accurate is the number on my wrist,” not “how good is the raw diode.” But it has a consequence Apple does not spell out:

Part of what this study measures is how much data each company lets out of its own ecosystem.

If a competitor’s sensor is excellent but its app only exports a coarse stream, this methodology scores the coarse stream. That is a real thing that affects real users. It is not the same claim as “Apple’s sensor hardware is more accurate,” and the paper’s conclusion — “Apple Watch offers users the most accurate heart rate sensing among the wearable devices evaluated” — sits right on the seam between the two.

The Samsung Problem

The Galaxy Watch 8 case is the sharpest illustration, and to Apple’s credit it is disclosed in full.

For daily-living background readings, Samsung’s export gave no per-sample timestamps. It gave a start time, an end time, a minimum, a maximum, and one additional heart rate value whose derivation Samsung does not document.

So Apple had to decide what that mystery value should be compared against. It tested three options: the reference at the interval’s start, the reference at the interval’s end, and the mean of the reference across the interval. The interval mean fit best, so that is what Apple used.

Read that again. Apple selected the scoring rule for a competitor’s data by testing which of three candidate alignments matched the reference most closely — the best-fit option rather than an arbitrary one — and then still reported that Samsung lost.

That is the fair-minded reading, and I think it is the right one. But the underlying situation stands regardless of intent: in one head-to-head comparison, one competitor’s grading method was chosen by the other competitor, because the first one shipped an undocumented number.

The Actual Scoreboard

The 139 is real. It is also not the whole line.

Across 159 activity-by-measure comparisons: Apple better at 99% confidence in 139, indeterminate in 18, competitor better in 2. Ignore statistical significance and just count which way the point estimate leaned, and Apple is ahead in 155 of 159. On the overall workout and overall daily living protocols, Apple beat every device on both RMSE and MAE.

That is a strong result by any standard. Two details are worth pulling out of the figures, though.

The Pixel Watch 4 beat the Apple Watch. Twice — RMSE and MAE, in the rest condition of the Daily Living protocol. Apple reports it and adds that the margin was under 0.5 bpm. Pixel Watch 4 also produced an indeterminate result against Apple in cycling, the only workout activity where Apple did not win outright against everyone.

Now notice the asymmetry in how those are written. Apple’s wins are reported as statistically significant, direction only. Apple’s single loss is reported with a magnitude attached — “by less than a margin of 0.5 bpm.” Both statements are accurate. But the effect sizes for the 139 wins live only inside Figures 1 through 3; there is no table of numbers in the text. You are told Apple won and by how little it lost.

For calibration: the study was sized to detect differences in mean absolute error of roughly 2.5 bpm in running and cycling, up to 3.7 bpm in HIIT. So “statistically significant” here is not synonymous with “large enough to feel.” Whether a 2 bpm edge changes anything about your training is a separate question from whether it is real.

The Least-Designed Part Is the Part Marketing Needs Most

This is the finding I did not expect.

Apple pre-specified its sample size: a minimum of 50 participants per workout activity per comparator device, with 80% power, α = 0.05, and a Bonferroni adjustment for the six device comparisons. That is a properly powered design.

And then one clause: sizing “pertains only to workouts and not Daily Living.”

The Daily Living protocol had no pre-specified sample size. It is still analyzed with the same linear mixed model and the same 99% bootstrap confidence intervals, so the reported results are not casual. But it was not designed to a target in advance.

Daily Living is background heart rate. Background heart rate is precisely what the Series 12 marketing claim is about — readings 60 times more often than Series 11, the finer-grained passive data that Readiness and Health Age are built from. The single most-promoted capability of the new sensor is validated in the half of the study that got no power analysis.

What Apple Chose to Tell You

I want to be even-handed, because the easy version of this article is cynical and the easy version is wrong.

Apple disclosed, without being obliged to:

• that it used a tool on its own device that no competitor had, and exactly why
• that it picked the alignment method for a competitor’s undocumented data
• that it lost two comparisons, and to whom
• that 18 more were indeterminate
• that the Daily Living arm was not power-sized
• that its participant pool was 65% male

Each of those is a stick handed to a critic. Most vendor white papers contain none of them.

The gaps that remain are structural rather than sneaky. The protocol was approved by review boards within Apple — the paper names groups covering research ethics, safety, privacy, and data governance, but no external institutional review board. There is no trial registration. The paper is not peer-reviewed, and the underlying data is not available, so nobody outside Apple can re-run the analysis with different reasonable choices and see whether 139 becomes 120. Those are the normal conditions of vendor research, not misconduct — but they are the reason vendor research and independent research are not interchangeable, no matter how carefully the vendor writes.

What This Means If You Wear One

• Heart rate is the one Series 12 capability with fresh evidence behind it. Everything else in the new health stack shipped without a paper. If you are choosing based on documentation, this is the documented part.
• Read the claim as “the number on the screen,” not “the sensor.” That is Apple’s own stated framing, and it is the useful one for a buyer — but it means a competitor’s poor export policy counts against its score.
• Do not translate “more accurate” into “meaningfully different for you.” The design could detect gaps around 2.5 to 3.7 bpm. If you are training by heart rate zones, ask whether a couple of bpm crosses a boundary you care about. Usually it does not.
• The background-data claim is the softest part. Workout accuracy was designed, powered, and won convincingly. Passive all-day accuracy was measured with less advance design — and it is the input to the scores Apple is promoting hardest.

The Takeaway

Grading your own homework is not automatically dishonest. It becomes dishonest when you hide the rubric. Apple published the rubric: which tool read which device, how a competitor’s ambiguous data was handled, where the power analysis applied and where it did not, and which comparisons it lost.

What you are left with is a well-run study that answers a narrower question than its conclusion sentence implies. Apple Watch reports heart rate to its user more accurately than six competitors report heart rate to theirs, under conditions Apple designed, using an extraction method only Apple could build for itself, and with the strongest evidence in workouts rather than in the passive background data its new features actually run on.

That is still a good result. It is just a different sentence than “Apple has the most accurate heart rate sensor,” and the gap between those two sentences is where all the interesting reading is.

Source: Apple Watch Heart Rate Accuracy Study, September 2026

Photo: Nik / Unsplash

TL;DR — Apple Watch Series 12 launched with a redesigned sensor array and four new health algorithms: Readiness, Health Age, Movement Evaluations, and a Longevity tab. Apple published exactly one new validation paper alongside them — a heart rate accuracy study — and that paper does not test heart rate variability or nighttime data. Readiness runs on heart rate variability and sleep. Apple has published validation papers for sleep apnea, hypertension, sleep stages, blood oxygen, and arrhythmia. Four of those five needed a regulator’s signature. Blood Oxygen never did — Apple shipped it as an unregulated wellness feature and published a paper on it anyway. That is the standard Apple set for itself, and the four new features do not meet it. The same launch also demonstrates the cost of the alternative: hypertension notifications, the regulated feature, is unavailable on the new hardware in 33 countries because clearances do not transfer.

If you are weighing an upgrade rather than the algorithms themselves, the full Series 12 vs Series 11 comparison goes through the spec sheet line by line.

The Story

Apple published one new health white paper for the Series 12 launch. It is nine pages, dated September 2026, and it is titled Heart Rate Accuracy Study.

It is a serious document. Apple enrolled 1,460 people across five sites in Cupertino, San Diego, Austin, and Selangor, Malaysia, and analyzed data from 1,254 of them. Twenty percent had Fitzpatrick skin tones V–VI, which matters enormously for optical heart sensors and is a demographic that wearable validation studies have historically underweighted. The reference standard was a Polar H10 chest strap ECG. Apple ran the watch against six competitors — Garmin Forerunner 970, Pixel Watch 4, Huawei Watch 5, Galaxy Watch 8, WHOOP 5, and Oura Ring 5 — across outdoor running, treadmill intervals, cycling, HIIT, strength training, and ordinary daily living. Across 159 activity-by-metric comparisons, Apple Watch came out ahead in 139.

That is a real study, and the result is probably real too. Hold that thought, because the interesting part is not what the paper contains.

The paper does not evaluate heart rate variability. It does not evaluate nighttime or sleep data.

Now look at what Apple built on top of the same sensor. Readiness is the headline software feature of Series 12 — a 0-to-10 score with four verdicts (Recover, Pace Yourself, Ready, Go For It) that updates through the day. Apple’s description of the inputs is one sentence long: recent activity, training load, vitals, and your sleep score. It does not say what sits inside “vitals,” and it does not say how the four are weighted. Separately, Apple says overnight vitals now includes recovery HRV measurements, and that this generation splits heart rate variability into two readings — Recovery HRV and Overall HRV — sampled up to 24 times more often than on Series 11. Put those two statements side by side, which Apple does not do for you, and the score looks like it leans substantially on HRV and sleep.

So the flagship feature runs on HRV and sleep, and the only validation paper Apple published measures neither.

Four Algorithms, No Documentation

Readiness is not alone. Series 12 shipped with three more inference-heavy health features, and none of them has a technical paper:

Apple says Readiness was developed using data from the Apple Heart and Movement Study, in collaboration with Apple’s own exercise physiologists and physicians. That is the entire disclosure. No subject count, no study duration, no agreement statistic against any reference, no error bars.

And there is a smaller tell in the same launch. Movement Evaluations includes a new VO2 max test run through the iPhone camera — it requires an iPhone, an Apple Watch, and either AirPods Pro 3 or a third-party heart rate monitor working together, not the watch on its own. The document Apple still points to for VO2 max estimation, Using Apple Watch to Estimate Cardio Fitness with VO2 max, carries a creation date of May 2021. A new measurement method arrived; the supporting paper did not move.

Apple Used to Do This Differently

This is not a company that dislikes publishing. Apple has a standing library of health validation papers, and it is reasonably thorough:

• Sleep Apnea Notifications — September 2024
• Hypertension Notifications — September 2025
• Estimating Sleep Stages — October 2025
• Blood Oxygen app — October 2022
• Arrhythmia detection

Line those up against the four new features and a pattern falls out immediately.

Four of those five are features a regulator had to approve. Atrial fibrillation detection, sleep apnea notifications and hypertension notifications are cleared medical device functions. Hypertension notifications went through the FDA as a 510(k), cleared under product code SFR as a Class II device. Sleep apnea went the same route. When you file a 510(k) you are assembling an evidence package regardless; publishing a consumer-facing summary of it costs you almost nothing.

The fifth one is the interesting one. Blood Oxygen was never FDA-cleared. Apple deliberately shipped it as a general wellness feature in 2020, outside the regulatory perimeter, explicitly not for medical use. That was a legitimate path, not a trick — classify a measurement as general wellness and no clearance is required. Then Apple published a paper on it anyway, two years later.

That matters, because it means the standard the four new features are failing is not a regulator’s. It is Apple’s own. Nobody obliged Apple to document Blood Oxygen. Apple documented it because that is what the company did when it shipped a new physiological measurement.

None of the four new features is a regulated claim either. A recovery score is not a diagnosis. A “health age” is not a medical finding. A flexibility rating from a phone camera is not a clinical assessment. In the United States these sit inside the FDA’s general wellness policy, which carves out low-risk products that promote a healthy lifestyle without claiming to diagnose or treat. No submission, no evidence package, no obligation to publish anything.

Blood Oxygen sat in that same carve-out and got a paper. Readiness, Health Age, Movement Evaluations and the Longevity tab did not.

So the regulatory line is not a loophole Apple invented — it is the real boundary, and every wearable company lives on both sides of it. But the boundary does not explain the change. Apple used to publish on both sides of it. What changed is that the papers stopped at precisely the moment the new health features stopped needing anyone’s signature.

The Same Launch Shows Why

Here is the part that turns a pattern into a strategy.

Hypertension notifications — Apple’s most heavily validated recent health feature, with a published paper and an FDA clearance — does not work on Series 12 or Ultra 4 in 33 countries. Apple’s own watchOS feature availability page states it plainly: the feature “may not be available on Apple Watch Series 12 or Ultra 4 in your region, as additional regulatory clearances for these devices are in process.”

The reason is the sensor. Apple redesigned the optical and electrical heart hardware, and a clearance granted for one sensor configuration does not automatically carry to another. So the regulated feature has to queue up again, country by country, on the new hardware. Buyers in Canada, Australia, Japan, Singapore, India, Brazil, Korea and two dozen other markets upgrade to the newer watch and lose a health feature the older watch had.

Readiness, meanwhile, shipped everywhere on day one.

There is also a business reason the unregulated side is worth building out. At the same event Apple announced a Quest Diagnostics blood panel — more than 50 biomarkers for $119, ordered through the Health app, US only, described as arriving later in 2026 and not live yet. Health Age already takes optional lab values as an input. Neither company has disclosed how the money moves, so this is a shape rather than an accusation — but the shape is a wellness score that asks you for lab work, and a paid lab order sitting one tap away inside the Health app. That is a loop Apple can close on its own surface. A regulated diagnostic sends you to a doctor instead.

That contrast is the whole argument in miniature. The regulated path buys you credibility and costs you years and geography. The wellness path costs you nothing and ships instantly. Apple built a sensor overhaul whose entire justification is finer-grained physiological data, and then routed that data almost entirely into features that no regulator will ever examine.

Frequency Is Not Accuracy

There is one more thing worth separating out, because Apple’s own language is careful about it in a way the coverage has not been.

Apple’s claim for the new sensor is that Series 12 takes background heart rate readings 60 times more often than Series 11, and HRV readings up to 24 times more often. Read those again. They are claims about frequency. Apple did not claim the readings are more accurate, and the accuracy paper it did publish is about heart rate, not HRV.

The two multipliers also describe two different signals, and they are worth keeping apart. Heart rate moves from an interval Apple’s general support documentation puts at roughly five minutes, varying with activity, down to about five seconds — that is the 60×. HRV moves from spacing on the order of a couple of hours down to roughly five minutes — that is the 24×.

Neither of those baselines appears on a Series 11 product page. Apple has never published a per-model background sampling interval; the numbers above come from general support documentation and from working backwards through Apple’s own multiplier. That is a reconstruction, not a disclosure, and it means “60 times more often” cannot be checked against anything Apple has actually stated.

The distinction matters because of how a recovery score works. Sampling an unvalidated signal more often does not make the signal more true. It makes the derived number more stable and more confident-looking. A Readiness score that updates all day off 24× the HRV samples will feel far more authoritative than one computed from a single morning reading — and the underlying question of whether wrist HRV tracks recovery well enough to justify a 0-to-10 verdict is exactly as open as it was before.

Nobody Documents the Score. Apple Documents Nothing.

It is worth being precise about how Apple compares here, because the honest version of the comparison is narrower than the easy version.

Nobody documents their composite recovery score. WHOOP does not publish how Recovery is calculated. Oura does not publish how Readiness is calculated — and yes, Oura’s flagship metric has carried that exact name for years, which makes Apple’s choice of word a small statement in itself. Garmin does not publish how Body Battery is calculated. All three are black boxes at the top, and so is Apple.

The difference sits one level down, at the components. WHOOP’s menstrual-cycle work went through peer review in npj Digital Medicine, and its Healthspan feature has a published white paper behind it — not a journal article, but a document you can read. Oura’s sleep staging algorithm, OSSA 2.0, was validated against polysomnography in Sleep Medicine, 96 subjects and over 400,000 epochs. Garmin’s physiological metrics rest on the publicly documented Firstbeat research base. Each of them has published something about the parts, even where the whole stays closed.

Apple has both layers for its regulated features and neither for its new ones. For Readiness, Health Age, Movement Evaluations and the Longevity tab there is no composite documentation and no component documentation. That is the actual gap — not that Apple is uniquely opaque about the score, but that it is the only one of the four with nothing underneath it either.

What This Means If You Wear One

None of this makes Readiness useless. A directional signal built from real physiological data, tracked against your own baseline, can be genuinely helpful — the same way a bathroom scale is useful even though it cannot tell you your body composition.

The practical guidance is narrower than that:

• Treat Readiness as a trend, not a measurement. Its value is in how it moves relative to your own history. The absolute number has no external reference to be right or wrong against.
• Heart rate is the part that was actually tested. If you are buying Series 12 for a specific capability, continuous heart rate is the one with a fresh nine-page study behind it and a documented win over six competitors.
• Check the regulated features against your country before upgrading. Hypertension notifications is the current example, and it is a real downgrade in 33 markets. Apple lists availability by region, and that page is worth reading before an upgrade rather than after.
• Health Age deserves the most skepticism, and for a practical reason. Compressing VO2 max, resting heart rate, sleep and HRV into one “biological age” is the largest inferential leap of the four. The concrete problem is not that the number might be wrong — it is that you cannot see which input moved it. If your Health Age climbs two years, nothing tells you whether that was sleep, cardio fitness, or a bad week of resting heart rate, so there is nothing in it you can act on.
• Movement Evaluations is a camera measurement, so read it like one. Flexibility, strength and balance scored from phone video will shift with lighting, camera angle, distance and what you are wearing. Compare sessions you recorded the same way in the same place, or do not compare them at all.

The Takeaway

The story of this launch is not that Apple built a worse sensor. By its own testing, it built a better one, and the testing looks credible.

The story is where the output goes. Apple spent a hardware generation on physiological measurement and then pointed that measurement at four features sitting deliberately outside the regulatory perimeter — while its most rigorously validated feature was stuck at the border in 33 countries. The papers stopped appearing at exactly the point where nobody was requiring them. Apple’s own Blood Oxygen paper is the proof that it did not always need to be required.

If you want to know how much a health feature has been checked, the most reliable signal is not the marketing. It is whether Apple published something you can read.

Source: Apple Heart Rate Accuracy Study, September 2026

Photo: Simon Daoudi / Unsplash

TL;DR — 1X unveiled a new hand for its NEO humanoid: 25 degrees of freedom (22 in the fingers and palm, 3 at the wrist), tendon-driven, with the motors sitting in the forearm instead of the hand. The spec that actually matters isn’t the DOF count — it’s the gear ratio. Most robot joints run 100:1 to 200:1 reduction. 1X runs roughly 5:1 to 15:1, which makes every joint natively force-controlled and fully backdrivable. That’s not a better hand. It’s a hand that flipped the trade-off: it gave up peak grip strength to buy force transparency. And the second-order effect lands somewhere unexpected — on the training data.

The Story

Every humanoid company eventually shows you a hand. It’s the money shot of the demo reel — fingers curling around a mug, threading a cable, doing something delicate enough to make you forget you’re watching a machine. And the number they put on the screen is almost always the same kind of number: degrees of freedom. Twenty-two. Twenty-five. Look how many joints.

Here’s the problem with that number. It tells you what the hand can move. It tells you nothing about whether the hand can feel.

1X’s new hand for NEO is worth looking at precisely because the company put a different number in front. According to 1X’s own technical writeup, the hand has 25 degrees of freedom — 22 fully actuated in the fingers and palm, plus 3 at the wrist — driven by tendons, with the motors housed in the forearm rather than crammed into the hand itself. Fine. Impressive, but not unusual. What’s unusual is the transmission: gear ratios of roughly 5:1 to 15:1.

If that means nothing to you, stay with me, because it’s the whole story.

A gearbox trades speed for torque. Small motor, big reduction, lots of force — that’s why industrial robot joints commonly run high-reduction harmonic drives, often in the 100:1 to 200:1 range. It works beautifully in one direction. The catastrophe is what it does in the other direction. Push back on a joint like that from the outside and the force has to fight its way backward through the entire gear train, against friction that doesn’t care which way you’re going. Reverse efficiency on a high-ratio harmonic drive typically lands in the 30-to-50 percent band, and drops below 40 percent in plenty of configurations. Some of them need three or four times more torque to backdrive than to drive forward.

The practical consequence is brutal and rarely stated plainly: if the gearbox burns 5 Nm just overcoming its own internal friction, the joint is blind to every contact force below 5 Nm. Not “imprecise.” Blind. The signal doesn’t attenuate — it never arrives.

And there’s a second penalty that’s arguably worse. Reflected inertia — how heavy the motor feels to the outside world — scales with the square of the gear ratio. Go from 10:1 to 150:1 and the apparent inertia at the joint doesn’t grow 15 times. It grows 225 times. A finger geared that hard isn’t a finger anymore. It’s a small hydraulic press that happens to be finger-shaped.

So what does a hand like that actually do when you tell it to pick up a wine glass? It executes a position command. “Close to this angle.” It has no meaningful idea how hard it’s squeezing, because the only channel through which that information could travel is clogged with friction and inertia. It finds out the glass was slipping the same way you’d find out — from the sound.

This is the trade 1X made, and it’s important to name it honestly as a trade. Dropping to single-digit and low-double-digit ratios means every one of those 25 joints is, in 1X’s phrasing, natively force-controlled and fully backdrivable. Force flows outward to the object; contact information flows back along the exact same mechanical path. No load cell bolted on as an afterthought, no inference from motor current fighting through a gear train. The transmission itself is the sensor. Call it “force transparency” — the hand and the world can feel each other.

The published numbers are modest in a way that tells you this was a real choice, not a free lunch. Peak torque of 3.5 Nm at the thumb CMC joint, 2.6 Nm at the finger MCP joints, distal flexion forces up to 45 N, and 17.75 Nm at the wrist. Positioning accuracy of ±0.2 mm. IP68 rating with food-safe materials — which is 1X telling you where it thinks this robot lives, and the answer is your kitchen sink. 1X also says finger assemblies were tested through millions of cycles and wrist joints beyond 2 million under high load.

Forty-five newtons at the fingertip is roughly five kilograms of pinch force. That’s enough for a mug, a plate, a bag of groceries. It is emphatically not enough to crush anything, and that’s the point — but it’s also a ceiling, and a real one. You cannot open a badly stuck jar with a hand tuned for transparency. Somewhere in a spreadsheet at 1X, someone decided that mattered less than knowing when a glass is sliding. My read is they’re right, but it’s a bet, not a fact.

The other layer here is the tactile stack. 1X says the fingertip skin senses normal force, contact location, and shear — and shear is the interesting one. Shear is the sideways sliding force, and it’s the physical signature of a grip beginning to fail. Detect shear early enough and you catch what roboticists call “incipient slip,” which lets the hand tighten before the object is gone rather than after. But notice the dependency: shear sensing is nearly useless if the joints can’t respond with a fine-grained force adjustment. Sensor and transmission only work as a pair. Bolt great tactile skin onto a 150:1 gearbox and you’ve built a hand that knows precisely how it’s failing and can’t do anything about it.

One more thing that’s easy to skip. 1X claims it makes the whole stack in-house — motors, electronics, embedded sensors, tendon materials, polymer skins, hand-specific firmware — on a dedicated line it says can produce up to 10,000 hands annually. That vertical integration claim is the part I’d flag hardest, and I’ll come back to it.

The Takeaway

Here’s where this gets bigger than one hand, and it’s the reason I wanted to write about it at all.

I’ve argued before that the real bottleneck in robotics isn’t hardware, it’s data — that the field’s center of gravity moved from “who builds the best body” to “who collects the most experience.” I still think that’s right. But this hand exposes something I underweighted: hardware sets the ceiling on what the data can contain.

Think about how humanoid manipulation data actually gets made. A human teleoperates the robot, the robot records what happened, and that recording becomes a training demonstration. Now run that through a non-backdrivable hand. What got recorded? Joint angles. Timestamps. Camera frames. What did not get recorded, because the mechanism was physically incapable of measuring it? How hard the fingers were pressing, and how that pressure changed moment to moment as the object shifted.

You can train a very large model on a very large pile of that data and it will never learn force control. Not because the model is too small or the dataset is too short, but because the variable was never in the file. Hardware that cannot feel force generates data that cannot teach it.

That reframes the vertical integration story. Everyone reads “we make our own motors” as a cost-and-supply-chain claim, and partly it is. But a company that controls the transmission, the skin, and the firmware controls what columns exist in its dataset. That’s a compounding advantage that doesn’t show up on a spec sheet, and it’s the same structural logic behind LG deciding to build the body and license the brain — except pointed the opposite way. LG bet the brain is the commodity. 1X is betting the body determines what the brain can ever learn.

Worth noting the differentiator is narrower than the headlines suggest. Tendon drive and forearm-mounted motors are not 1X exclusives — Tesla’s Optimus V3 patents describe a tendon-driven hand with actuators moved into the forearm, routing cables through the wrist — 22 DOF in the fingers and palm plus a 2-DOF wrist, against 1X’s 3-DOF wrist. The forearm is where a lot of the industry is going, for the obvious reason that it’s where your own grip muscles are. So the tendons aren’t the story. The gear ratio is. That’s a narrow claim, and narrow claims are usually the true ones.

Now the cold water, because there’s plenty.

Low gear ratios cost you more than peak force. Tendon systems stretch, fray, and wear, and they demand far more sophisticated control than a stiff geared joint — the compliance that makes the hand safe also makes it harder to command precisely. 1X reports millions of test cycles, but as the technical writeup at Embodied Global points out, there’s no published third-party data on mean time between failures, contact lifetime, or field repair cost. Lab cycles under controlled load and a year in a real kitchen are different tests.

Moving motors to the forearm solves a packing problem and creates a mass problem. Every actuator you relocate there adds weight and inertia to a segment that swings at the end of a long lever. Lighter fingers, heavier arm. There’s no free space in a humanoid — you move a problem, you don’t delete it.

The 10,000-hands figure is a stated internal capacity, not a demonstrated run rate. Manufacturing capacity announcements have a long and unimpressive history of describing a line that could theoretically run at that volume if everything downstream cooperated.

And the largest caveat isn’t about the hand at all. NEO is in early access at $20,000 outright or $499 a month, and reporting on the launch — TNW’s coverage among others — notes that unfamiliar tasks still route through a teleoperated “Expert Mode,” meaning a remote human. A hand that feels contact does not give you long-horizon autonomy. It doesn’t give you failure recovery, the thing that actually separates a demo from a product: knowing the grasp failed, figuring out why, and trying a different approach without being told. That’s a software problem, and better hardware only moves the starting line.

Which is why the most interesting robot in any given year is usually not the most impressive one. I’ve made the case that the machine quietly winning Physical AI is a cabinet on wheels in a hospital hallway — boring, narrow, and actually deployed. That’s still true. NEO’s hand won’t beat the cabinet on usefulness for a long time.

But it’s doing something the cabinet can’t, and it’s worth being precise about what. It’s generating a category of data that most of the industry’s hardware physically cannot capture. If dexterity really does follow a scaling law, then the companies that spent this era collecting force-annotated demonstrations will have a dataset nobody can retroactively reconstruct. You can always collect more video. You cannot go back and measure a contact force that no sensor was there to record.

The DOF race was always a bit of a vanity metric. Counting joints is easy, and it photographs well. Whether the hand can feel what it’s touching is harder to put on a slide — and it’s the number that decides whether any of the joints matter.

This article is for informational purposes only and is not investment or purchasing advice.

Photo: Franck V. / Unsplash

TL;DR — “NSF certified” is not a specification. Certification is granted claim by claim, contaminant by contaminant, so a filter certified to NSF/ANSI 53 might reduce exactly one health contaminant and nothing else. Korea’s mandatory test is even narrower: four items, six for reverse osmosis, and everything beyond that is optional and paid for by the manufacturer. That’s why a purifier advertising “48 certified items” and one advertising “certified” can sit on the same shelf. Meanwhile RO systems typically deliver an efficiency rating of 10–20 percent, meaning most of the water goes down the drain, and filter costs over a rental contract routinely exceed what the hardware is worth. Buy the claim list, check the drain ratio, and price the filters before the machine.

The Story

Here’s a question worth sitting with: what exactly does “NSF certified” tell you about a water filter?

The intuitive answer is something like “it’s been tested and it works.” That answer is wrong in a specific and expensive way, and once you see why, the entire category rearranges itself.

NSF/ANSI certification isn’t a grade. It’s a claim registry. A manufacturer picks which contaminants it wants to advertise, pays for testing on those specific contaminants, and if the product passes, that contaminant gets listed. Nothing else does. A filter “certified to NSF/ANSI 53” — the health effects standard — may have been tested for lead reduction and only lead reduction. It carries the same certification language as a filter tested for lead, cysts, VOCs, mercury, asbestos, and PFOA/PFOS. Same phrase on the box. Wildly different products.

This is not a loophole. It’s the intended design. The standards exist so claims can be verified individually, which is genuinely useful — but it means the standard number alone is close to meaningless as a purchase signal. What matters is the list underneath it.

So let’s be precise about what each standard actually measures, because they measure different things and people routinely treat them as a ladder where 53 beats 42 and 58 beats everything.

NSF/ANSI 42 is aesthetic effects. Chlorine taste and odor, chloramine, particulates, iron, manganese, zinc. Things you notice. Explicitly not a health standard. A filter certified only to 42 is a taste filter, and that’s a legitimate product — it just isn’t protection from anything.

NSF/ANSI 53 is health effects. Lead, cysts like Giardia and Cryptosporidium, VOCs, mercury, asbestos, chromium-6, arsenic V. Also where PFOA/PFOS claims live. This is the standard that matters if you’re worried about what’s actually in the water rather than how it tastes — but remember, per-claim.

NSF/ANSI 58 is reverse osmosis systems. Note the wording: it certifies a system, not a cartridge. It carries a mandatory TDS reduction test plus optional claims for fluoride, nitrate, hexavalent chromium, lead, arsenic, PFAS. RO systems also contain pre- and post-filters typically rated under 42 or 53, so an RO unit’s real coverage is a composite of several listings.

NSF/ANSI 401 is emerging contaminants — 15 of them, requiring at least 95 percent reduction. Pharmaceuticals (ibuprofen, naproxen, atenolol, carbamazepine, meprobamate, phenytoin, trimethoprim), the hormone estrone, and industrial compounds (BPA, DEET, nonylphenol, TCEP, TCPP, linuron, metolachlor). Note what’s not here: PFAS claims are certified under 53 or 58, not 401, despite 401 sounding like the “modern contaminants” standard.

You can verify any of this yourself, which is the part I’d emphasize. NSF runs a public listings database searchable by model number, and it shows each certified claim as a separate line item. If a product page says “NSF certified” and the database shows two claims, you now know something the marketing copy was designed to prevent you from knowing.

Korea’s version of this problem is even sharper, and almost nobody explains it

If you’re buying in Korea, the NSF framework is background. The binding rule is domestic, and it’s structured the same way — mandatory floor, optional everything else.

Korean water purifiers require quality certification under the Drinking Water Management Act (먹는물관리법) before they can be sold, administered by the Korea Water Technology Certification Institute. Performance testing splits into four parts, and the split is the whole story:

Mandatory purification performance testing covers free residual chlorine, color, turbidity, and chloroform. Four items. For reverse osmosis systems, hardness and nitrate nitrogen are added, bringing it to six.

Optional purification performance testing is everything else — drawn from the 45-odd items in the national drinking water quality standards — and it happens only when a manufacturer chooses to advertise a removal claim and pays to have it verified.

Four items. That’s the floor. Every legally sold water purifier in Korea has cleared chlorine, color, turbidity, and chloroform, and legally, nothing more is required.

This explains a marketing pattern that otherwise looks like noise. When Chungho Nais announced its premium RO ice purifier had passed 48 KC-certified items, the breakdown was six general performance items plus 42 special performance items — including boron, which the company noted was then unique among Korean household purifiers. That’s not 48 versus some competitor’s 40. That’s 48 versus a legal minimum of four or six. The gap between a heavily certified machine and a minimally certified one is enormous, and both are sold as “certified.”

There are two other tests worth knowing. Leaching safety testing checks 51 items for substances migrating out of components that touch the water — a real concern in a plastic device holding water at temperature. And effective purification volume testing determines how much water a filter can treat before it stops working, evaluated at 25, 50, 75, 100, and 120 percent of the manufacturer’s claimed volume, with chloroform reduction of at least 80 percent required to certify.

That last one deserves a moment. The filter replacement interval on your purifier isn’t a maintenance suggestion. It’s the point at which certified performance was last demonstrated. Past it, you’re outside the tested envelope, and the machine gives you no indication because chloroform breakthrough has no taste. Typical intervals assume roughly 10L per day for a four-person household — heavier use shortens them.

The number RO makers leave off the spec sheet

Now the second thesis, and this one costs money every single day.

Reverse osmosis forces water through a semi-permeable membrane. What passes through is permeate — your drinking water. What doesn’t becomes concentrate, carrying the rejected contaminants, and it goes down the drain. The ratio between those two streams is a design characteristic of the system, and it’s the spec most likely to be absent from the page you’re buying from.

The EPA’s WaterSense specification for point-of-use RO systems puts hard numbers on it. A typical POU RO system, EPA writes, sends five gallons or more down the drain for every gallon of treated water it produces. Expressed as NSF/ANSI 58’s efficiency rating — the percentage of intake water that reaches you as permeate — typical systems land between 10 and 20 percent, and EPA assumes 15 percent as the market average for tank-equipped residential systems.

Run that forward. A household producing about 1,000 gallons of drinking and cooking water a year at 15 percent efficiency sends roughly 5,667 gallons to the drain annually. EPA set the WaterSense threshold at 30 percent efficiency, which cuts that to about 2,333 gallons — saving around 3,300 gallons per household per year.

Sit with the shape of that. At typical efficiency, over 80 percent of the water entering the machine is discarded. Not filtered, not stored, not used. Sent to the sewer so a membrane can stay clean.

Two clarifications, because this gets overstated in both directions. The reject water isn’t pure waste in a technical sense — flushing the membrane surface is what prevents concentration polarization and keeps the membrane alive. And efficiency depends heavily on line pressure; higher inlet pressure pushes more water through and improves recovery, which is why permeate pumps and booster pumps exist. But the structural fact holds: an RO system is a machine that throws away most of its input, and the ratio is knowable and rarely published.

This is the actual reason tankless direct-flow purifiers took over the Korean market, and the reason is usually explained backwards. The common story is hygiene — no storage tank means no reservoir sitting at room temperature growing a biofilm, which is true and does matter. But direct-flow designs also pair with nanofilter and hollow-fiber membranes rather than RO membranes, and those filter at pore sizes that don’t require a reject stream at all. No RO membrane, no concentrate, no drain ratio. The hygiene argument sells the machine; the water-and-throughput argument is what made the engineering work.

The tradeoff is real, though, and it runs in the direction most Korean marketing won’t state plainly. RO’s pore size — on the order of 0.0001 microns — is why it can strip dissolved inorganics like nitrate, hardness, boron, and heavy metals that a nanofilter simply passes. Nanofilters work largely by electrostatic adsorption, capturing negatively charged microbes and particulates while letting dissolved minerals through. That’s marketed as “retains minerals,” and for taste it’s a fair point. But minerals passing through and nitrate passing through are the same physics. If your concern is dissolved contaminants, RO is the technology that addresses them, and direct-flow is not. If your concern is chlorine, taste, odor, and particulates from aging plumbing — which, for most Korean households, it is — direct-flow does that job without discarding four-fifths of your water.

What Korean tap water actually needs removed

Which brings up the framing error underneath most purifier shopping in Korea: importing an American mental model about treating questionable source water.

Korean tap water is already treated to a legal drinking standard. Under the Drinking Water Quality Standards regulation, tap water must satisfy limits across microbiology, inorganic and organic substances harmful to health, disinfectants and disinfection byproducts, aesthetic substances, and radioactivity. Total trihalomethanes are capped at 0.1 mg/L, chloroform at 0.08 mg/L, bromodichloromethane at 0.03 mg/L. Free residual chlorine must not exceed 4.0 mg/L, and treatment plants deliberately maintain a residual — a minimum of 0.1 mg/L at the tap — because that residual is what prevents recontamination between the plant and your kitchen.

Read that last sentence again, because it reframes the purchase. The chlorine in your tap water is a feature, not a defect. It’s there on purpose, doing a job, right up until the moment the water leaves the tap. After that it’s just an unpleasant taste and a small source of disinfection byproducts.

So the honest job description for a purifier on Korean municipal supply is narrow: remove residual chlorine and its taste, remove the disinfection byproducts indexed by chloroform, remove particulates and metals that enter downstream of the treatment plant through aging building plumbing and pipework, and reduce microbial risk from that same downstream stretch. That’s a job a well-certified carbon-and-membrane direct-flow unit does completely.

Notice that this is almost exactly the mandatory test list — chlorine, color, turbidity, chloroform. That’s not a coincidence. The mandatory items were chosen to match the realistic threat profile of treated municipal water. The optional 42 items exist for people with a specific reason to want them.

Which is the correct way to decide: buy optional claims when you have a reason. Well water or a private source, a building with documented lead plumbing, an area with nitrate from agricultural runoff, a household with infants where nitrate matters more. Absent one of those, paying an RO premium in both money and discarded water to remove dissolved solids that were within legal limits when they arrived is buying protection against a problem you don’t have.

The cost structure, which decides more purchases than anyone admits

Same pattern as the home printer, where the hardware is a delivery mechanism for consumables. With purifiers, Korea’s rental model makes it stranger, because the recurring cost is bundled and therefore invisible.

The standard structure: a monthly fee, a mandatory contract term of three to seven years, ownership transferring at term end. Inside that fee sit three things — hardware amortization, filters, and either scheduled technician visits (방문관리) or self-replacement kits shipped to you (자가관리). Self-managed plans are cheaper because you’re doing the labor.

Real numbers make the point better than description. Coway’s Icon Water Purifier 2 runs about ₩27,900/month on a six-year self-managed plan — roughly ₩334,800 a year. SK Magic’s Eco Mini Green 41 runs about ₩14,900/month on six years, roughly ₩178,800 a year. Both look like modest monthly commitments.

Now the part that catches people. Consumer reporting in Korea has documented that when contracts end and owners switch to buying filters themselves, the filter cost approaches the rental fee they were paying. Coway’s six-month self-replacement kit runs about ₩104,000; twice a year is roughly ₩210,000 annually against ₩334,800 in rental. SK Magic’s four-month filter runs about ₩40,000; three times a year is ₩120,000 against ₩178,800.

So the rental fee is mostly filters. Which means the widespread belief that the contract ending brings relief is wrong — the ongoing cost never had much to do with the hardware. It was consumables the whole time, and it continues.

That’s not an argument against renting. It’s an argument for pricing the filters first, then asking what the remainder buys. If the remainder is small, renting is close to a filter subscription with free repairs attached, which is a reasonable thing to buy. If it’s large, you’re financing hardware at an implied rate worth calculating.

And there’s a genuine case for visit-managed plans that has nothing to do with convenience. As established above, the replacement interval is the boundary of certified performance. A machine past its interval isn’t merely less effective — it’s outside the envelope where anyone verified anything, and it produces water that tastes fine regardless. If you know you’ll postpone filter changes, paying someone to show up is buying performance, not comfort. Same logic as the air purifier guide: a neglected filter is worse than no machine, because it convinces you the problem is handled.

The Specs

How to read this table: the “certified claims” column is the one that isn’t comparable across rows, and that’s the argument of this article. Every Korean row has cleared the same four-item mandatory floor (six for RO). Anything beyond that is optional testing the manufacturer elected to pay for, so two rows showing “mandatory floor” are not thereby equivalent — one may carry many optional claims and the other none. Verify per model in the certification listing rather than inferring from brand. Filter intervals assume roughly 10L/day for a four-person household and shorten with heavier use or poorer inlet water. Blank cells marked “Not published” mean the manufacturer does not publish a comparable figure, not that the value is zero. Rental pricing moves constantly with promotions, card-linked discounts, and contract length; treat these as structure, not quotes.

The Verdict

Municipal tap water, ordinary apartment, no specific contaminant concern: direct-flow, and it isn’t close. This is the overwhelming majority of Korean households. Chlorine, taste, odor, particulates, and downstream plumbing contributions are exactly what a certified carbon-and-membrane direct-flow unit handles, and you discard no water doing it. Confirm the model’s optional claims in the certification listing, then buy on filter cost and footprint.

A documented reason to remove dissolved contaminants: reverse osmosis, and accept the drain ratio. Well or private source water, confirmed lead plumbing, nitrate from agricultural runoff, infants in the household. This is what RO is for, and nothing else does it. Ask for the efficiency rating before buying — if the seller can’t produce it, that’s informative. Anything at or above 30 percent is doing meaningfully better than the 10–20 percent typical range.

Buying on “NSF certified” or “KC certified” alone: don’t. Pull the model number, open the certification listing, and read the claim list. Two minutes of database lookup outperforms any amount of marketing copy, and it’s the single highest-leverage action in this entire guide. A machine with four claims and a machine with 48 both say “certified.”

Small households and single-person apartments: consider unpowered ambient-only units. An SK Magic Eco Mini-class device at 2 kg and 94 mm wide, drawing no electricity, does the actual filtration job. Hot and cold water is a convenience feature that adds cost, bulk, standby power, and hygiene surface area. If you boil water for tea anyway, you may be paying a substantial premium for a function you’d replace with a kettle.

If you know you’ll skip filter changes: pay for visit management. Not for comfort — for the certification envelope. The interval is where verified performance ends, and the failure is silent. This is the same trade as any maintenance-dependent appliance, and self-knowledge is worth more than the monthly delta.

Renting versus buying: price the filters, then decide. Compute annual filter cost, subtract from annual rental cost, and look at the remainder. That remainder is what you’re paying for hardware financing, repairs, and labor. Sometimes it’s fair. Sometimes it’s a lot for a machine you’ll own anyway.

The thread here matches the smart bulb guide, where the headline spec was fine and an invisible attribute decided everything, and the smart scale guide, where a confident number on a screen was an estimate in a measurement’s clothing.

Water purifiers add a twist. The certification isn’t an estimate and isn’t wrong — it’s narrow, and the narrowness is invisible because the phrase “certified” is identical whether it covers four items or forty-eight. The specification you’re buying was never the filter material or the flow rate.

It was the list. Go read the list.

This article is for informational purposes only and is not investment advice.

Photo: Swanky Fella / Unsplash