The paper was printed before you got here.
This form carries thirteen signal columns, a prior envelope for each, and four band edges, all pressed onto the stock before you arrived. The run decides where inside those printed ranges your line falls; then the form counts your values against the devices these sites have read.
Scores this page's reading of your browser against the printed priors, then counts each value against the devices these sites have read. What is stored.
The pens write the printed rubric. The black bars stamped after them are counts: the bits of each of your values among the devices these sites have read.
The derivation, line by line
Select any column on the form — or any row below — to print that signal's assumed probability, its surprisal, whether the correlation discount applied and against which sibling, and the rubric's written justification word for word.
| Signal | Observed | p | −log₂p | Bits | Sharing it | Bits, counted |
|---|---|---|---|---|---|---|
| Operating system | — | — | — | — | — | — |
| Screen resolution | — | — | — | — | — | — |
| Device pixel ratio | — | — | — | — | — | — |
| Color depth | — | — | — | — | — | — |
| CPU cores | — | — | — | — | — | — |
| Device memory | — | — | — | — | — | — |
| Time zone | — | — | — | — | — | — |
| Language preferences | — | — | — | — | — | — |
| GPU renderer | — | — | — | — | — | — |
| Canvas rendering | — | — | — | — | — | — |
| Audio sample rate | — | — | — | — | — | — |
| Installed fonts | — | — | — | — | — | — |
| Touch points | — | — | — | — | — | — |
| Total after discount | naive sum — | — | — | — | all 29, ×0.5 | — |
The two right-hand columns fill after a pass: how many of the devices these sites have read share each value, and what that is in bits.
The stub is what you leave with
Not a figure — an addition. Every row carries the value that was read, the probability this rubric assumed for it, the surprisal that follows, and whether the correlation discount applied. Tearing it off publishes the derivation under a permanent address so it can be argued with line by line, months later, by someone who was not here.
What publishing sends: the derivation summary only: the thirteen rows you can see, their counts against the stored devices, the rubric version and the caveat. It includes the observed values, so your GPU string and screen size become readable by anyone with the link. The full reading is never published.
- printed assumption vs. the reading
- a stroke, then one at 35% of it
- a rule that did not read
Tear-off stub · form FS-1
NOTHING TO TEAR OFF YET
value — · p — · −log₂p — · bits —
value — · p — · −log₂p — · bits —
value — · p — · −log₂p — · bits —
value — · p — · −log₂p — · bits —
value — · p — · −log₂p — · bits —
value — · p — · −log₂p — · bits —
value — · p — · −log₂p — · bits —
value — · p — · −log₂p — · bits —
value — · p — · −log₂p — · bits —
value — · p — · −log₂p — · bits —
value — · p — · −log₂p — · bits —
value — · p — · −log₂p — · bits —
value — · p — · −log₂p — · bits —
That figure is a property of the prior table, not a count of anything. The rubric caps the exponent at 53 before exponentiating.
The instrument, with its working shown
The chart form above, as the machine that would print it. Until you run a pass the pens write the rubric's modal browser, the reference printed with the form. After a pass they write your thirteen readings, and the black bars are your counts.
Set the inputs by hand
Every value below feeds the same pure function the live pass uses. Blank a field and the rule declines to read — its pen lifts and it leaves the sum entirely, rather than counting as zero bits. Currently seeded from the modal browser under this rubric.
Recomputed
| Signal | p | Bits |
|---|---|---|
| Operating system | 0.620 | 0.69 |
| Screen resolution | 0.220 | 2.18 |
| Device pixel ratio | 0.500 | 0.35 |
| Color depth | 0.940 | 0.03 |
| CPU cores | 0.280 | 1.84 |
| Device memory | 0.550 | 0.30 |
| Time zone | 0.050 | 4.32 |
| Language preferences | 0.340 | 0.54 |
| GPU renderer | 0.020 | 1.98 |
| Canvas rendering | 0.0020 | 8.97 |
| Audio sample rate | 0.720 | 0.47 |
| Installed fonts | 0.300 | 1.74 |
| Touch points | 0.680 | 0.19 |
Bits are surprisal against this rubric's published prior assumptions, not against a measured population. No sample backs these numbers yet.
What this instrument refuses to do
It will not tell you that you are rarer than 99.4% of users. The devices it counts are visitors of fingerprinting tools, not the web, so it prints the count and the population's size instead. It will not print a single glowing figure and hide the arithmetic behind a methodology link. And it will not report a percentile: neither the priors nor the counts support one.
What it does instead is print the assumptions first, then write your reading on top of them, then hand you the addition. Every line of it can be attacked separately, which is the only form of a uniqueness score that is worth arguing about.
- The whole prior table — thirteen rules, every assumed probability, every written justification, and the same table as JSON.
- The correlation discount — what a naive sum would have claimed, and why 0.35.
- The five bands — where the edges are and what each one does not tell you.
- Which signals actually move — and which ignore JS-level patching entirely.
The thirteen columns
In the order the rubric evaluates them. Each column's printed envelope spans every outcome rule 1.0.0 can produce for that signal — the assumption's full range, before it has seen anyone.
| Signal | Pen | Envelope | Modal |
|---|---|---|---|
| Operating system | platform | 0.69 – 5.64 | 0.69 |
| Screen resolution | display | 2.18 – 7.97 | 2.18 |
| Device pixel ratio | display | 1.00 – 5.64 | 1.00 |
| Color depth | display | 0.09 – 6.64 | 0.09 |
| CPU cores | capability | 1.84 – 6.64 | 1.84 |
| Device memory | capability | 0.86 – 5.64 | 0.86 |
| Time zone | locale | 4.32 – 7.38 | 4.32 |
| Language preferences | locale | 1.56 – 8.38 | 1.56 |
| GPU renderer | graphics | 5.64 – 8.64 | 5.64 |
| Canvas rendering | graphics | 8.97 only | 8.97 |
| Audio sample rate | audio | 0.47 – 6.64 | 0.47 |
| Installed fonts | typography | 1.74 – 5.64 | 1.74 |
| Touch points | capability | 0.56 – 6.64 | 0.56 |
What a bit of uniqueness actually is
One bit is one halving. If this rubric assumes a value is present in one browser out of two, reading it tells you one bit; one in four is two bits; one in a thousand is just under ten. That is the entire unit, and it is why a bit count is a better answer to how unique is my browser than a percentage: bits from independent readings add, and percentages do not.
The word doing the work in that sentence is assumes. Surprisal is only as good as the probability you feed it. The pens are fed a published table, so the printed total means under these stated assumptions, your thirteen readings are worth this many halvings. The stamped bars are fed counts of real devices, so they mean this many of the browsers these sites have read share your value. Each is checkable on its own terms.
Why thirteen readings and not three hundred
Because a rule that cannot state a defensible prior should not be in the sum. Every column on this form has a written justification you can attack; a table of three hundred signals would have three hundred numbers nobody could source, and the extra columns would mostly be restatements of the ones already there — which is the same double-counting the correlation discount exists to correct.
How to use the number without over-reading it
Treat it as an instrument reading, not a verdict. It answers one question well — which of your signals are surprising under a stated set of assumptions — and it is silent on the two questions people usually bring to a page like this.
It cannot tell you whether a specific site will recognise you, because that depends on what that site collects, how it correlates across sessions, and whether it has seen you before. And it cannot tell you whether your setup looks automated, because rarity and coherence are different properties: a rare-but-consistent browser is often less interesting to a vendor than a common one whose signals contradict each other.
What it does well is the operator's actual job. Take a pass, change one thing, run it again, and read which pens moved. The comparison view keeps both traces on one form, and the columns that refuse to move under a given tool are a direct statement about the level that tool works at.
Questions this form gets asked
Answered against rubric 1.0.0 as shipped. Every probability quoted anywhere on this site is a constant in the scoring code, printed before your run.
- Is this a measurement of how rare I am among real users?
- Half of it is. The printed bits are surprisal against a published table of assumed probabilities. The counted columns are real: how many of the devices the eight sites have read share each of your values. Those devices are visitors of fingerprinting tools, not the web at large, so every count is printed with the population's size.
- Why is there no percentile?
- A percentile ranks you against everyone, and the counted devices are not everyone. The form prints counts (3 of 214) instead, and the rubric reports a band with edges at 12, 20, 28 and 36 bits, printed on the form before any run.
- Why are some signals worth less than the surprisal they carry?
- Because screen size, pixel ratio and platform are not independent facts. Within a correlated group the strongest signal counts in full and every other member at 35%. The chart draws that discount as a physically shorter stroke and prints the arithmetic beside it.
- What does this site record about my browser?
- Every page view is recorded, and once per browser the page takes a full reading of about 120,000 browser properties in the background and stores it, again after a browser update or 30 days. RUN scores that reading. A stub publishes only the derivation summary. The privacy page lists everything and has a delete button.
- How unique is my browser, in plain terms?
- The run gives you a bit count and a band. A bit is one halving: if the rubric assumes a value is present in one browser in four, that value contributes two bits, because two halvings get you from everything to a quarter. Thirteen readings are summed, correlated ones are discounted, and the result is compared against four printed edges.
- Why is my score different in another browser on the same machine?
- Because several of the readings are properties of the browser rather than of the hardware. Firefox and Safari do not implement device memory at all, engines differ in what they report for pixel ratio under zoom, and the canvas hash depends on the text rasteriser the browser chose. Two browsers on one machine are two different rows on this form, which is the correct answer rather than an inconsistency.
- Can I see the calculation for a single signal?
- Click any column on the chart and the margin prints that signal, your value, the assumed probability, minus log base two of it, whether the group discount applied and against which sibling, and the rubric's own written justification for the prior, reproduced without edit.
Where the pens move
The pens that refused to move are the ones patched at JS level.
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.