Two numbers on one form, and they are never added together
The pens write surprisal against a printed table of assumptions. The stamped bars beside them are counts: how many of the devices the eight sites have read share each of your values. They answer different questions, so this site prints both and folds neither into the other.
Every one of your values, counted
The recorder counts the thirteen values the rubric scores. The full reading has 29 that the population is compared on; here are all of them.
Stored so far: counting…
What the printed number is
For each signal the rubric states a probability, its guess at how common that value is, and the bits are −log₂ of it. A value the table assumes is present in one browser in four contributes two bits; one in a thousand contributes just under ten. Add the signals up, discount the correlated ones, and you have the total.
There is no second step, no weighting learned from data, and no adjustment afterwards. You can reimplement it from the published table and get the same answer.
“Bits are surprisal against this rubric's published prior assumptions, not against a measured population. No sample backs these numbers yet.”
RUBRIC_PROVENANCE, rubric 1.0.0, basis “prior-assumption”.
What the counted number is
- The population: every distinct device the eight sites read in the last 365 days, its latest reading.
- Left out: this browser's own cookie identity, and headless or automated readings.
- The share: (devices with your value + 1) ÷ (population + 1). The bits: −log₂ of that, at most 33.
- The combined figure: half the sum over all 29 values, because they share information.
- Under 200 devices the form says the population is small, beside every figure.
Large surfaces (fonts, canvas, audio, WebGL) are compared by the collector's hash of the whole measurement, so two browsers share a value only when every font or pixel matched. Where the counted value is not the rubric's quantity, the form says so: the audio column counts the processing hash, not the sample rate; the fonts column counts the exact list, not the count.
Why the counts do not replace the priors
Every device in the population came from somebody who typed a phrase like how unique is my browser into a search box, or followed a link from a thread about proxies. That population is not the web. It over-represents privacy tooling, unusual operating systems, virtual machines, headless runners and people mid-way through configuring an anti-detect stack, which is to say exactly the values a uniqueness score is trying to price.
A frequency from that sample carries a decimal point and is wrong in a direction nobody can estimate. Publishing it as “rarer than 99.4% of users” would borrow the authority of counting without the part that makes counting work. So the form prints the count and the population's size, “3 of 214”, and leaves the inference to you.
How to argue with either number
With a prior: pick a line. Every prior has its own page, its outcome table and a paragraph on where it is weakest: the time zone rule pricing twelve zones identically, the screen rule with a cliff and no gradient, the GPU rule with exactly two outcomes. The reply is a source or a version bump, filed in public.
With a count: read where it disagrees with the prior. A value the rubric calls common and nobody here shares is the population's bias or the prior's error, and the two columns side by side are how you find out which.
Three things the printed side does without counting anyone
It prints the range first
Each of the 13 columns carries an envelope spanning every outcome that rule can produce, for anyone. A reading is a position inside a published range rather than a claim out of nowhere.
It prints a reference reading
A browser at every rule's modal outcome totals 23.61 bits after discount, band “distinctive”: a fixed property of the rubric, computed from the table.
It admits what it did not read
A rule that cannot read its signal lifts the pen. It is listed as not read, with the reason,
and excluded from the sum, never counted as zero bits. Firefox and Safari do not implement
navigator.deviceMemory, so a rubric that scored a missing value as zero would hand
every Firefox user a quieter total for a reason that has nothing to do with them.
Questions about priors and counts
Rubric 1.0.0, basis prior-assumption. The counts come from the stored devices at the moment you press, and change as the population grows.
- So the printed number is made up?
- The arithmetic is exact and the assumptions are published. What the printed side lacks is a sample, and it says so. The counted side is the sample: real devices, with the population's size printed beside every count.
- Why not replace the priors with the counts?
- Because the counted devices are visitors of fingerprinting tools, not the web. They over-represent privacy tooling, virtual machines and half-configured anti-detect stacks: exactly the values a uniqueness score is trying to price. A prior is wrong in a direction you can argue about; a biased count is wrong in a direction nobody can estimate. So both are printed, and neither is folded into the other.
- Is my own browser in the counts?
- No. Every reading made under this browser's own cookie identity is left out, so a browser is never compared with itself. The same machine read under another identity, in a private window for example, is in the population.
- Why does a value nobody else has score only a few bits?
- The share is smoothed: one is added to the count and to the population. A value seen nowhere else is at most one in the population plus one, not infinitely rare. With two hundred devices that is under eight bits, and the ceiling rises only as the population grows.
- How can a score be useful while the population is small?
- The comparison that needs no sample is against yourself: take a pass, change one thing, take another, and read the per-column delta. The counts add a second check: a value nobody else here shares is worth a look, whatever the printed prior says.
- Why does the stub still print a one in N figure?
- Because it is the vernacular form of a bit count and people ask for it. On the printed side it is two raised to the total, a property of the prior table, capped at an exponent of 53. On the counted side it is the API's estimate from the combined bits, capped at 33 bits, about the people on Earth.
Where the pens move
A rare value cannot be fixed in the page. Patching it makes a patched value.
Engine level, not JS level. Donut Browser spoofs through CDP injection in its Wayfern Chromium fork, so the value a page reads is the value the engine reports — there is no patched accessor to catch.
It refuses to contradict itself. A profile whose proxy exit geography, time zone or language disagrees with its fingerprint does not launch. That is a pre-launch gate in the app, not advice in a manual.
Open, and quiet. AGPL-3.0, desktop app and sync server both in the repository, and no analytics SDK anywhere in the source. macOS, Windows, Linux.