What "Inspectable" Actually Means in a Health App

Half the health AI field now says it shows you the numbers. A big figure on a card, a sparkline, a "based on 30 days". That sounds like proof. It is usually a picture of a number. Here is the short test that tells a real provenance card from the same claim built out of decoration.

By MetricBridge · Published 16 November 2026 · ~6 min read

AI-assisted, human reviewed. How we use AI

Quick answer

"Inspectable" means one answer can be checked end to end: the raw samples behind it, the date window, the calculation that turned them into a number, and the same number when you ask again. If an app shows a figure but cannot name those four things, it is displaying data, not letting you inspect it. That gap is the whole argument.

What "we show the numbers" usually means

Showing a number is the easy part. Any app can compute a 7-day average and print it at 30px. The claim in the marketing is bigger than that, though. It implies you, the reader, can follow the number back to where it came from and see the work.

Most products stop at the display. The card shows a result and a label. It does not show which readings went in, which days were included, what happened to the missing ones, or what "average" was applied. You are asked to trust the pipeline that produced the label exactly as much as you trust the number. Nothing about that is inspection.

Inspectable is a claim about four things

Strip the phrase down and an inspectable answer has to be four things at once:

  1. The samples. The individual readings the number was built from, not a count of a hidden set.
  2. The window. The exact date range, so a "this week" is a set of days you could point at, not a vibe.
  3. The calculation. What was done to those readings: an average, a delta, a baseline comparison, and how gaps were handled.
  4. A reproducible re-run. Ask the same question again and get the same figure, because the same code ran over the same data.

Three of those are description. The fourth is a test you can actually run. That is where most products quietly fail.

A real provenance card, so the shape is concrete

Here is an answer from MetricBridge's own on-device Ask, computed from a seeded 30-day series. This is the whole card, not a screenshot of a prettier version:

51.2 ms
HRV, 7-day average
up 4.8% versus the prior 14 days
samples: 7 · window: 9 Nov to 15 Nov
identical on a re-run · sha256 7c1d9ab4

Read it as the four parts. The samples: seven of them. The window: 9 to 15 November, stated so you can check it against Health on the same phone. The calculation: a 7-day mean compared against the two weeks before it. The re-run: the same ask returns 51.2 ms again, because the phone reruns the same arithmetic over the same readings.

That last line is the one worth buying. A number you can reproduce is a number you can argue with. A number you cannot reproduce is a number you can only accept.

Why reproducibility is the hard part

Reproducing an answer is easy when the calculation happens where the data lives. It is hard the moment a model sits between you and the figure. Ask a hosted chat "what was my HRV last week" and the reply is generated text: it may pull a real figure, it may round it, it may blend it with an earlier question, and asking again can give you a slightly different sentence. The wording moves because a language model is writing it, not deriving it.

On-device changes that. MetricBridge computes the number in a function you can run again, then renders a card from the result. The number is data before it is ever prose. Ask twice and the figure is stable, because nothing sampled it the second time.

The one-line test. Pick a number the app gave you and ask the same question again. If it moves, something you cannot see took part. If it holds, you are looking at a calculation over your own data, not a reply about it.

The claim you cannot check

The failure mode is not a lie, it is a shortcut. "Based on your last 90 days" with no window. A "readiness" gauge with no baseline. A trend arrow with no sample count. Each of those can be true and still leave you unable to inspect any part of it, because the inputs are on a server you do not hold and the method is not shown.

There is a fair version of that design: a product that states its method, dates it, and links to it. If a number's path is described but not runnable by you, call it documented. It is a step up from decoration. It is still not inspection, because the samples sat somewhere you cannot reach and you cannot re-run the calculation to confirm it.

What has to be true before it means anything

For "inspectable" to hold, the app needs the data on the device, the method fixed and published, the output rendered from a deterministic result, and no server in the path that could change the answer between asks. MetricBridge is built that way: the app reads Apple Health read-only, computes every Ask answer on the iPhone, and shows a provenance card on the result. Your health data never reaches us or a cloud AI, and the MCP server makes zero network calls. The methods behind each number are published at /methods/.

That is the whole claim, and it is checkable in a minute on your own phone. Which is the point.

Four questions to ask any health app

  1. Where were the samples? Can it name the readings behind a number, or only a count of days?
  2. What window, exactly? Two dates you could point at, or "recently"?
  3. What was the calculation? A named operation over those readings, or a described outcome?
  4. Does it hold on a re-run? Ask the same question twice. If the figure moves, part of it came from something you did not supply.

One-time, not metered inspection

Inspection is cheap to run and awkward to sell, so it tends to arrive metered: access with a prompt count, where each check costs a request and a request costs the vendor. MetricBridge has no server of ours in the path, so there is nothing to meter. A 7-day free trial, then a single $24.99 one-time unlock covering the export, the MCP tools, on-device Ask with a provenance card on every answer, the Weekly Health Brief, and the Mac and Apple Watch apps. Ask as often as you like, because checking an answer should not carry a price per question.

For the design behind the card, see how Ask runs on the phone. For the other half of the surface, the MCP server hands the same data to your own agents with no server in between.

FAQ

What does it mean for a health app to be inspectable?

It means one answer can be checked end to end: the app shows the raw samples it used, the date window, the calculation that turned them into a number, and the same figure when you ask again. A rendered number you cannot trace back to those four things is a claim, not an inspection.

Is a number a health app shows the same as an inspectable number?

No. Displaying a figure is easy; anyone can print a large number on a card. The inspectable part is the path back from that number to the inputs. If the app cannot name the samples and the window behind it, the display is decoration, not provenance.

Why does reproducibility matter so much?

If the same question returns the same figure every time, the answer is a calculation over data you hold, and you can re-run it to confirm. If it drifts between asks, some part of it came from something you cannot see: a cached reply, a server model, or state you did not supply.

Can a cloud app be inspectable?

It can be honest about method, but inspection stops at the boundary you cannot cross. You can read the description of how a server computed a number; you cannot re-run it yourself, and the samples it used sat somewhere you do not control. On-device calculation puts the whole path in your hands. Wellness information, not medical advice.

See your own numbers, and the working behind them

Every Ask answer computed on your iPhone from your own Apple Health data, every figure shown with its samples and window, every re-run identical. No account, no server of ours in the path.

Where it goes: the app reads Apple Health read-only and computes answers on the device. Your health data never reaches us or a cloud AI. What you get: all 194 metrics, on-device Ask with a provenance card on every answer, the Weekly Brief, the export and the MCP tools (21 tools, 19 read-only, and 23 prompts), and the Mac and Apple Watch apps. Price: $24.99 one time after a 7-day free trial, not a subscription. Wellness, not medical advice.