Two endpoints, no key
One returns the rubric. One returns a published derivation. Between them you can reproduce any number this site has ever printed, without running its JavaScript.
Both are plain GET requests with no authentication, no rate-limit token and no registration. That is not generosity — a scoring method nobody can fetch and re-run is not published, it is merely displayed, and the whole argument of this site is the difference between those two things.
Static, on this domain
GET /rubric.json
The whole prior table: 13 rules, every distinct outcome with its assumed
probability and bits, every rationale, the correlation discount, the clamp, the band edges,
and the reference reading for a browser at every rule's modal outcome. It carries
measured: false and the provenance statement in the body, so a consumer cannot
strip the caveat by accident.
On the derivation store
GET /api/share/<id>
A published stub, exactly as it was stored. The id is the one in the stub's own
address — the tail of /s/<id>/. The response is the record, not a rendering
of it, and every row carries the probability the rubric assumed at the time so the total can
be re-derived without this site.
{
"id": "k7m4qp9rst",
"createdAt": "2026-08-14T09:12:44.108Z",
"kind": "score",
"payload": {
"form": "FS-1",
"rubricVersion": "1.0.0",
"measured": false,
"caveat": "Bits are surprisal against this rubric's published prior assumptions, not against a measured population. No sample backs these numbers yet.",
"correlationDiscount": 0.35,
"serial": "FS1-20260814-7K3Q",
"total": 26.4118,
"naiveTotal": 34.9024,
"band": "distinctive",
"crossedEdge": 20,
"oneInN": 89478485,
"oneInNCapped": false,
"signals": [
{
"id": "gpu",
"label": "GPU renderer",
"group": "graphics",
"value": "ANGLE (NVIDIA, NVIDIA GeForce RTX 4070 Direct3D11 vs_5_0 ps_5_0, D3D11)",
"prior": 0.0025,
"rawBits": 8.6439,
"bits": 8.6439,
"discounted": false,
"discountedAgainst": null,
"rationale": "The single highest-entropy stable signal in a normal browser. …"
}
],
"notRead": [
{
"id": "memory",
"label": "Device memory",
"reason": "navigator.deviceMemory is absent — Safari and Firefox do not implement it."
}
]
}
} Shown with one signal for length. A real record carries every rule that read, plus every rule that did not, with its reason.
Writing
POST /api/share
Body { kind: "score", payload: { … } }, answering { id, createdAt }. The payload is capped at 16 KB, which is roughly four times
what a full derivation needs and far below anything that could carry a probe.
What is deliberately not here. There is no endpoint that returns a stored fingerprint probe. A probe is around 13 MiB, contains an exhaustive property enumeration including the page's own markup, and serving it behind a short id would be a disclosure hole with an enumerable key. Publishing a stub writes a summary into a different table with no path back to the probe.
Counting
GET /api/stats
How many full readings and distinct devices the eight sites have stored, and since when:
{ probes, devices, visits, since, generatedAt }. It is the population the
counted columns are drawn from, refreshed every five minutes. A published stub carries its own
counts in payload.population, frozen at the moment it was torn off.
Reproducing a total
For each signal, take the assumed prior p and compute −log₂ p — that
is rawBits. Group the signals by their correlation group. In each group the
highest rawBits counts in full and every other member is multiplied by 0.35. Sum
the results. Compare against the four band edges at 12, 20, 28 and 36. That is the entire
algorithm, and it is implemented in about forty lines.
If your reimplementation disagrees with a published stub, one of us is wrong and the stub carries enough information to find out which. That is the property this whole site is built to have.
Two details bite reimplementations. The correlation group winner is chosen from the pass, not
from a fixed hierarchy, so you have to sort each group by raw bits before applying the 0.35
factor. And a rule that declines to read is excluded from the sum entirely — it is not zero
bits, and treating it as zero will silently agree with us on most browsers and disagree on
every Firefox, which does not implement navigator.deviceMemory at all.
What the JSON is actually good for
Three things this site cannot do for you. Auditing your own fleet: score profiles from your own harness against a table you did not write, and diff the results against the values your tooling claims to be presenting. Arguing with the table: the priors are data, so a disagreement can be expressed as a patch rather than as a complaint. Archiving: a derivation fetched today stays checkable after the rubric moves, because each row carries the probability that was assumed for it at the time.
What it is not good for is scoring a browser you cannot run code in. Values like the canvas hash, the renderer string and the detected font count only exist because something rendered; no header and no server-side inspection produces them.
The prior table as a page, with every rationale · the discount rule a reimplementation has to get right · the four band edges a total is compared against
Questions about the derivation API
Both endpoints serve rubric 1.0.0 data. Neither one requires or accepts credentials.
- Do I need an API key?
- No. There is no key, no account and no registration. The prior table is a static file on this domain and a published derivation is fetched by the id printed in its own address, which is the only credential involved.
- Can I fetch a stored fingerprint probe through this API?
- No, and no endpoint for it exists. A stored probe is a large document containing an exhaustive property enumeration, and serving one behind a short enumerable id would be a disclosure hole. Publishing a stub writes a summary into a different table with no path back to the probe.
- Is the rubric JSON versioned?
- It carries the rubric version, the basis field, the measured flag and the provenance statement in the body, so a consumer cannot strip the caveat by accident. If the table changes, the version in the file changes with it.
- What is the payload size limit for publishing?
- 16 KB, which is roughly four times what a full derivation needs and far below anything that could carry a fingerprint probe. The cap exists so the derivation store cannot quietly become a second probe store.
- Can I score a browser server-side with this table?
- You can implement the scoring anywhere, because the whole algorithm is about forty lines and the table it needs is served as JSON. What you cannot do is read the signals server-side: values like the canvas hash and the renderer string only exist inside a browser that rendered something.
Where the pens move
A published table lets you check a browser. Changing what that browser reports is a different job.
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.