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

Bits are not just added up

A 3024×1964 display at 2× essentially implies a MacBook Pro. Counting the resolution, the pixel ratio and the platform at full weight charges you three times for one fact.

So rubric 1.0.0 does not add its bits up. Signals are placed in seven correlation groups; in each group the loudest reading of that pass counts in full and every other member is multiplied by 0.35. On the reference reading alone that is the difference between 29.92 bits and 23.61.


The discount as a lever

Every discounted signal goes through the same machine: its raw bits lift one end, and what it adds to the total is read 0.35 of the way along. With a pass stored in this browser, the lever steps through your own discounted signals; otherwise through the modal browser's.

Fig 2Reduction lever, 0.35
024681012ADDED× 0.35GROUP DISCOUNT · 0.35024681012RAW−log₂ p
Drag the carriage · tap for the next signalStatic view
A grouped signal's raw bits lift the carriage on the right. The follower sits 0.35 of the way along the lever, so it rises 0.35 as far: that is what the signal adds once the strongest member of its group has counted in full.

What a naive sum would have claimed

Take the modal browser — every rule sitting at its own likeliest outcome, which is a fixed property of the table and not a person. Adding its per-signal surprisal without any correlation model gives 29.92 bits. Applying the group discount gives 23.61 bits: a difference of 6.31 bits, or a factor of 79 in the “1 in N” form that people quote.

That gap is the size of the exaggeration built into any score that adds its signals up. It is not a rounding difference. It is most of the headline.

How it is drawn, not described

On the chart form, a discounted signal is a second physical event. The pen deflects to the signal's full surprisal — inside the printed envelope, where it belongs — and then drops to exactly 35% of that stroke, with a printed bracket connecting the two heights and ×0.35 set beside it. You can see which signals were discounted without reading a single number.

Which member takes full weight is decided per pass. If your canvas readback is blocked, the GPU string takes full weight in the graphics group instead. The table does not fix a winner.

The groups, worked through

Below is the modal browser's own pass. In each group, one signal counts in full and the rest are discounted against it.

Group Full weight Discounted Contributed
graphics Canvas rendering 8.97 bits, in full GPU renderer — 5.64 × 0.35 = 1.98 10.94
locale Time zone 4.32 bits, in full Language preferences — 1.56 × 0.35 = 0.54 4.87
display Screen resolution 2.18 bits, in full Device pixel ratio — 1.00 × 0.35 = 0.35 Color depth — 0.09 × 0.35 = 0.03 2.57
capability CPU cores 1.84 bits, in full Device memory — 0.86 × 0.35 = 0.30 Touch points — 0.56 × 0.35 = 0.19 2.33
typography Installed fonts 1.74 bits, in full sole member, no discount 1.74
platform Operating system 0.69 bits, in full sole member, no discount 0.69
audio Audio sample rate 0.47 bits, in full sole member, no discount 0.47
Total naive sum 29.92 saved 6.31 23.61

Your own pass is worked the same way on the chart form, and every row of the tear-off stub records whether the discount applied and against which sibling.


Where the model is wrong

The groups are hand-drawn. Audio sample rate is heavily determined by the platform, and the table does not connect them at all — so a platform that implies 48000 Hz gets charged for both. Fonts and canvas overlap for the same reason: the text rasteriser is part of both readings, and they sit in different groups.

A single flat 0.35 is also a simplification. Real dependence between screen size and pixel ratio is not the same strength as the dependence between a GPU string and a canvas hash, and one number cannot express both.

The honest position is that these are the two weakest joints in the rubric, they are written here rather than discovered by a critic, and they are the first things a measured version would replace — with a covariance estimated from data rather than a constant chosen by argument.

If you have a source that bears on either of them, open a challenge. Every 13 rules and the discount itself are open to it.


Which groups actually discount anything

Seven groups exist, but only 4 of them have more than one member in a full pass — graphics, locale, display, capability. The rest are sole occupants, and a sole occupant always contributes its full surprisal because there is nothing louder beside it to discount against. That is worth knowing before you spend effort on a column: the audio sample rate and the installed font count are never protected by the discount, so every bit they carry reaches the total intact.

The reverse is the more common surprise. A visitor changes their screen resolution, watches the raw figure for that column drop by several bits, and sees the black summing pen barely move — because the pixel ratio was the loudest member of the display group in that pass and the screen reading was already being charged at 35%.

Why the dominant member is chosen per pass

A fixed hierarchy would be simpler to describe and wrong in a specific, common way. If the canvas rule were permanently the dominant member of the graphics group, then a browser that blocks canvas readback would have its GPU renderer — 8.64 bits, the loudest stable signal a normal browser hands over — discounted to 3.02 for no reason other than a table entry.

Choosing the dominant member from the pass avoids that. Whichever reading actually carried the most surprisal this time counts in full, so removing a signal can never make the remaining signals cheaper than they are. It also means the discount is a property of your run rather than of the paper, which is exactly why the printed envelopes on the chart form show undiscounted surprisal and the pen strokes show the discounted stroke.

What eight bits of movement is worth at a band edge · read the per-column delta from two passes · the full prior table with every group printed


Questions about the correlation discount

The discount is a published constant of 0.35, versioned with the rubric. Changing it would change the rubric version and reprint the form.

Why 0.35 and not some other number?
It is a published assumption, chosen to be large enough that a grouped signal still contributes and small enough that it cannot dominate. It is not fitted to data: the counted population is still too small and too skewed to fit it to. It is versioned, so changing it changes the rubric version.
Does the discount ever apply across groups?
No. Only within a group, and only to members that are not the strongest in that pass. Which member is strongest is decided per run, not fixed in the table.
What is a correlation discount in a uniqueness score?
It is the correction for signals that are restatements of the same underlying fact. One display produces a resolution, a pixel ratio and a colour depth; one GPU produces a renderer string and the canvas pixels. Charging full price for each of them counts one machine several times, and the total inflates without anything being learned.
Which signal in a group takes full weight?
Whichever one carries the most raw surprisal in that particular pass. It is decided per run rather than fixed in the table, so a blocked canvas readback promotes the GPU renderer to full weight in the graphics group automatically.
Do most fingerprint scores discount at all?
Almost none disclose that they do. A total that is the plain sum of its per-signal bits is either double-counting correlated signals or has never been checked for it, and there is no way to tell which from the outside when the method is not published.

Where the pens move

Correlation is why one changed setting rarely moves the total. Consistency is why the wrong change makes it worse.

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 ↗