From San Digital, who run live AI courses for engineers and business teams.Live AI training from San Digital. See the courses

declawd

A worked SynthID-Text fixture

Follow one public reference profile from fixed probability masses to an exact token-trace score. The values and contracts come from Declawd v0.2.0, and the interaction runs locally.

This is not Claude’s watermark, Anthropic’s detector or Google’s production implementation.

It cannot test arbitrary text or establish AI use. It uses a public educational key and fixed data.

Do not use it to judge a person.

Four fixed token candidates

The fixture starts with context token IDs 11 22 33 44. Its four candidate masses are integers out of 1000; they are not model logits or captured model output.

Each public key produces a binary g-value for a candidate. The first six keyed columns are shown because the registered tournament uses depths zero to five.

Scroll horizontally to see all columns

Fixed candidate masses and public keyed g-values
CandidateToken IDProbability massg0g1g2g3g4g5
cedar101400/1000 (40%)111110
fern202300/1000 (30%)000001
moss303200/1000 (20%)111001
reed404100/1000 (10%)001010

Six registered tournament rounds

Each row compares two fixed draws at one key depth. A higher g-value wins; a tie keeps the first draw. The fixture records the draws, so there is no random choice in this page.

Scroll horizontally to see all columns

Registered draws and winners
DepthFirst drawSecond drawWinner
0cedar (1)fern (0)cedar
1moss (1)reed (0)moss
2fern (0)reed (1)reed
3cedar (1)moss (0)cedar
4reed (1)fern (0)reed
5moss (1)cedar (0)moss

One prepared token trace

The scorer hashes five-token windows with 64-bit wrapping arithmetic. Repeated four-token contexts and positions at or after the first end token are excluded. This prepared trace has 8 valid contexts and no exclusions.

The raw score counts set g-bits. The weighted score applies the published integer weights. Both stay as unreduced fractions and use integer round-half-to-even for the 12-place decimal.

  1. 0101
  2. 1202
  3. 2303
  4. 3404
  5. 4505
  6. 5606
  7. 6707
  8. 7808
  9. 8909
  10. 91010
  11. 101111
  12. 111212
Valid contexts
8
Raw score
131/2400.545833333333
Weighted score
19846/382800.518443051202

Each mask below lists candidate position zero first. A 1 means that the position is included. All eight prepared positions survive the repetition and end-token checks; the digest pins each one-byte LSB0 mask.

Scroll horizontally to see all columns

Prepared repetition, end-token and valid masks
MaskPosition bitsPacked SHA-256
Repetition11111111a8100ae6aa1940d0b663bb31cd466142ebbdbd5187131b92d93818987832eb89
End token11111111a8100ae6aa1940d0b663bb31cd466142ebbdbd5187131b92d93818987832eb89
Valid11111111a8100ae6aa1940d0b663bb31cd466142ebbdbd5187131b92d93818987832eb89

The fixed values remain readable without JavaScript. Switching among the registered substitutions needs JavaScript because the exact score is recalculated in this browser.

Scroll horizontally to see all columns

Independent registered substitutions and their exact scores
EditToken indexChangeRaw scoreWeighted score
edit-a4505506134/240
0.558333333333
20770/38280
0.542580982236
edit-b7808818118/240
0.491666666667
17804/38280
0.465099268548
edit-c1011111112136/240
0.566666666667
21188/38280
0.553500522466

No detector threshold or text verdict

A raw reference score is not a detector result. This page has no private production key, compatible tokenizer, calibration set, decision rule or false-positive and false-negative measurements for Claude.

Anthropic says Claude uses a version of SynthID-Text. That family relationship does not make this public profile interoperable with Claude.

The browser accepts no prose and makes no network request. The same fixed trace can be scored with declawd lab synthid score from the immutable v0.2.0 release.

The four JSON Schemas are published under Contracts. Their vendored bytes are checked against the release source contract during every site build.