Fingerprintscore Chart form FS-1 Rubric 1.0.0 Priors: assumed · Counts: real Donut Browser ↗

Every prior, in full

A browser uniqueness score is only as good as the assumptions under it, so here are all of ours: 13 rules, 60 distinct priors, seven correlation groups, one discount and four band edges. There is no second, private table.

Each rule turns one browser signal into a probability it assumes for the value it read, and the bits are −log₂ of that probability. Nothing else happens. If you disagree with a number on this page you disagree with a specific line, which is the only kind of disagreement a scoring page can usefully have.

Bits are surprisal against this rubric's published prior assumptions, not against a measured population. No sample backs these numbers yet.

Chart form FS-1, unwritten

024681012.0012 COMMON20 DISTINCTIVE28 RARE36 SINGULAROSSCREENDPRDEPTHCORESMEMORYZONELANGGPUCANVASAUDIOFONTSTOUCHONE OUTCOMEBITS / SIGNALTOTAL 0–40FORM FS-1 · PRIOR TABLE, RUBRIC 1.0.0 · PRINTED BEFORE THE RUN · PRIORS ASSUMED 024681012.0012 COMMON20 DISTINCTIVE28 RARE36 SINGULAROSSCREENDPRDEPTHCORESMEMORYZONELANGGPUCANVASAUDIOFONTSTOUCHONE OUTCOMEBITS / SIGNALTOTAL 0–40FORM FS-1 · PRIOR TABLE, RUBRIC 1.0.0 · PRINTED BEFORE THE RUN · PRIORS ASSUMED

This is the form as it ships: no reading on it, every envelope from the table below. A different rubric version prints a different form — that is what a version means here, and it is why an old stub stays readable years after it was torn off.


The prior table

Each block below is one rule. The outcomes are the distinct priors the rule can return, walked by calling the shipped scoring function itself rather than transcribing anything by hand — which is also why an open-ended input like a GPU string shows two outcomes and not two thousand. The rationale is the rubric's own text, reproduced without edit.

Operating system

Rule id platform · correlation group platform · envelope 0.69 – 5.64 bits

Desktop OS share is the best-documented signal here. Windows dominates, macOS is a strong second, Linux is a long tail. Assumed from broadly reported desktop share, rounded hard.

Outcomes for the Operating system rule
Outcome Assumed p −log₂p
Windows modal outcome — the printed line 0.620 0.69
macOS 0.180 2.47
macOS 0.180 2.47
Android 0.100 3.32
ChromeOS 0.020 5.64
Linux 0.020 5.64

Screen resolution

Rule id screen · correlation group display · envelope 2.18 – 7.97 bits

A handful of resolutions cover most of the web; everything else is a long tail. 1920x1080 is treated as the single most common, common laptop sizes next, anything unlisted as tail.

Outcomes for the Screen resolution rule
Outcome Assumed p −log₂p
1920x1080 modal outcome — the printed line 0.220 2.18
1536x864 0.070 3.84
1366x768 0.070 3.84
1440x900 0.050 4.32
2560x1440 0.050 4.32
1280x720 0.030 5.06
3840x2160 0.030 5.06
1512x982 0.020 5.64
1728x1117 0.020 5.64
3440x1440 any size not on the rule's list 0.0040 7.97

Device pixel ratio

Rule id dpr · correlation group display · envelope 1.00 – 5.64 bits

Effectively three buckets — 1, 2, and everything else. Fractional ratios come from OS display scaling and are individually uncommon.

Outcomes for the Device pixel ratio rule
Outcome Assumed p −log₂p
1 modal outcome — the printed line 0.500 1.00
2 0.280 1.84
3 any other whole ratio 0.040 4.64
1.25 any fractional ratio from OS display scaling 0.020 5.64

Color depth

Rule id colorDepth · correlation group display · envelope 0.09 – 6.64 bits

Almost universally 24. Anything else is notable on its own.

Outcomes for the Color depth rule
Outcome Assumed p −log₂p
24-bit modal outcome — the printed line 0.940 0.09
30-bit any depth other than 24 0.010 6.64

CPU cores

Rule id cores · correlation group capability · envelope 1.84 – 6.64 bits

hardwareConcurrency clusters on powers of two. Chromium also caps the reported value, which compresses the tail.

Outcomes for the CPU cores rule
Outcome Assumed p −log₂p
8 modal outcome — the printed line 0.280 1.84
4 0.220 2.18
12 0.130 2.94
16 0.120 3.06
2 0.070 3.84
6 0.070 3.84
10 0.040 4.64
20 0.030 5.06
24 0.020 5.64
5 any count not on the rule's list 0.010 6.64

Device memory

Rule id memory · correlation group capability · envelope 0.86 – 5.64 bits

navigator.deviceMemory is deliberately coarse and caps at 8, so it carries little information by design. Weighted accordingly.

Outcomes for the Device memory rule
Outcome Assumed p −log₂p
8 GB modal outcome — the printed line 0.550 0.86
4 GB 0.250 2.00
2 GB 0.100 3.32
1 GB 0.050 4.32
0.5 GB 0.030 5.06
0.25 GB any value not on the rule's list 0.020 5.64

Time zone

Rule id timezone · correlation group locale · envelope 4.32 – 7.38 bits

There are hundreds of IANA zones but traffic concentrates in a few dozen. Assumed uniform-ish across the busy zones with a long tail.

Outcomes for the Time zone rule
Outcome Assumed p −log₂p
America/New_York any of the twelve busy zones modal outcome — the printed line 0.050 4.32
Pacific/Chatham any zone outside them 0.0060 7.38

Language preferences

Rule id languages · correlation group locale · envelope 1.56 – 8.38 bits

A single `en-US` is by far the most common list. Every additional language multiplies the possibilities, so longer lists are treated as much rarer.

Outcomes for the Language preferences rule
Outcome Assumed p −log₂p
en-US modal outcome — the printed line 0.340 1.56
de-DE any single non-en-US list 0.030 5.06
de-DE, en any two-entry list 0.012 6.38
de-DE, en-GB, fr any list of three or more 0.0030 8.38

GPU renderer

Rule id gpu · correlation group graphics · envelope 5.64 – 8.64 bits

The single highest-entropy stable signal in a normal browser. The renderer string names the driver, the adapter, and often the exact backend, and there are thousands of distinct strings.

Outcomes for the GPU renderer rule
Outcome Assumed p −log₂p
Google SwiftShader any software renderer modal outcome — the printed line 0.020 5.64
ANGLE (NVIDIA, NVIDIA GeForce RTX 4070 Direct3D11 vs_5_0 ps_5_0, D3D11) any hardware renderer string 0.0025 8.64

Canvas rendering

Rule id canvas · correlation group graphics · envelope a single point at 8.97 bits

The rendered pixels depend on GPU, driver, OS text rasterizer and installed fonts together. Highly identifying, and heavily correlated with the GPU signal — which the group discount handles.

Outcomes for the Canvas rendering rule
Outcome Assumed p −log₂p
b41c9f2e77a0… every canvas hash, without exception modal outcome — the printed line 0.0020 8.97

Audio sample rate

Rule id audio · correlation group audio · envelope 0.47 – 6.64 bits

Overwhelmingly 48000, with 44100 second. Other values indicate unusual audio hardware or a spoofer.

Outcomes for the Audio sample rate rule
Outcome Assumed p −log₂p
48000 Hz modal outcome — the printed line 0.720 0.47
44100 Hz 0.240 2.06
96000 Hz any other rate 0.010 6.64

Installed fonts

Rule id fonts · correlation group typography · envelope 1.74 – 5.64 bits

Detected count from a fixed 78-name probe list. A stock OS lands in a narrow band; a design or development machine sits far outside it.

Outcomes for the Installed fonts rule
Outcome Assumed p −log₂p
28 of 78 21–35 detected modal outcome — the printed line 0.300 1.74
14 of 78 1–20 detected 0.200 2.32
44 of 78 36–50 detected 0.120 3.06
62 of 78 51–78 detected 0.030 5.06
none detected 0.020 5.64

Touch points

Rule id touch · correlation group capability · envelope 0.56 – 6.64 bits

Effectively binary on desktop (0) versus touch devices (5 or 10).

Outcomes for the Touch points rule
Outcome Assumed p −log₂p
0 modal outcome — the printed line 0.680 0.56
5 0.160 2.64
10 0.120 3.06
1 0.020 5.64
2 any other count 0.010 6.64

The discount

Within a correlated group the strongest signal contributes its full surprisal and every other member contributes 35% of theirs. The groups are display, platform, graphics, locale, audio, capability and typography.

A browser at every rule's modal outcome sums to 23.61 bits after discount, against 29.92 if the bits were simply added up. The full argument for 0.35 →

The bands

The rubric reports a band, never a percentile. The edges are constants in the shipped code and they are printed on the right-hand scale of the form.

Band From To
blends in 0 12
common 12 20
distinctive 20 28
rare 28 36
singular 36 —

What each band does and does not tell you →


How to read one of these blocks

Take any rule above. The outcome column is a value the rule can return — not every value a browser could hold, but every value that lands in a distinct price band. The assumed p column is the probability this rubric asserts for that outcome, and it is the only editorial judgement in the whole calculation. The −log₂p column is arithmetic on the column beside it and contains no opinion at all.

That separation is deliberate. Everything you can argue with is in one column, and everything you can check is in another. A rubric where the judgement and the arithmetic are mixed together — a weight applied here, a normalisation there — cannot be attacked at a line, and a score that cannot be attacked at a line is a matter of faith in whoever published it.

Why the outcomes are walked, not transcribed

The tables on this page are not typed out by hand. Each one is produced by calling the shipped scoring function against candidate inputs and reading back what it says, so a drift between this page and the code that scores you is not possible: the page would show the drift. The candidate list is the one authored thing, and it is why an open-ended input like a renderer string shows two rows rather than two thousand.

What publishing the table costs us

A published prior can be checked against your own knowledge in about a minute, and several of these will not survive that. The time zone rule prices twelve zones identically. The screen rule has a cliff and no gradient. The pixel ratio rule puts an extremely ordinary Windows scaling factor in the same bucket as a value almost nothing produces. Each of those is written on the column's own page under a heading that says so.

The alternative — a hidden formula — costs nothing to publish and cannot be improved, because nobody outside can see where it is wrong. That is the trade this site takes the other side of.

Reading a total against this table

A browser sitting at every rule's modal outcome sums to 23.61 bits after discount, against 29.92 if the bits were simply added up. That reference reading is a fixed property of the table, printed with the form. It is the closest thing to a middle this instrument has, and it is not a median of anybody.

Priors beside counts · which columns can actually move · what a version change would look like


Questions about the prior table

Every answer below refers to rubric 1.0.0 as shipped. If an answer and the table disagree, the table is right and this page is a bug.

What is a prior in this table?
A stated assumption about how likely a value is, written down before anyone is scored against it. The bits for a signal are minus the base-two logarithm of that number.
Why is every prior clamped at 1/4096?
Because an unclamped prior lets one rule dominate the entire sum on the strength of a guess. The clamp caps any single signal at 12.00 bits, which is why the left-hand scale on the chart form stops there.
Why not replace these priors with the counts the site collects?
Because the counted devices are visitors of fingerprinting tools, a sample that over-represents exactly the values this table prices. The counts are printed beside every prior instead. If a version ever adopts measured frequencies, the version increments, every prior gains a sample count, and the changelog says so.
Can I reimplement this scoring myself?
Yes, and that is the point of publishing it. Take each signal's assumed p, compute minus log base two of it, group the signals, keep the highest raw figure in each group at full weight and multiply the rest by 0.35, then sum. The same table is served as JSON so you never have to transcribe it from a web page.
Why do some rules have only two outcomes?
Because the rule tests for one condition and prices everything else identically. The GPU rule asks only whether the renderer string names a software rasteriser; the canvas rule asks nothing at all and gives every hash the same prior. Those collapses are drawn on the printed form as points rather than bands, because hiding them would misrepresent how coarse the table is.
Does running the check change the rubric?
No. The prior table is a static part of the shipped code and it is identical for every visitor. Probes submitted by a check are stored for a future version that could replace assumptions with frequencies; they do not feed the score you are shown today.

Where the pens move

A published rubric is checkable. A spoofing engine either moves a line on it or it does not.

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.

Download Donut Browser Source on GitHub — AGPL-3.0 ↗