Invented Words

Order-n Markov chains over real vocabulary — noise at 2, plagiarism at 4.

words.pseudo · v1 · renderer text

deling
chand
crated
clamply
spired
overply
washbook
hexamp
elockin
preforch
foldfire
poliar

Controls · from schema()

How many

1 60

Chain order

2 4

2 is soup, 3 is pronounceable, 4 mostly regurgitates real words.

Shortest

3 12

Longest

4 16

Trained on

Entropy source

Picking a source changes the story on the receipt, never the quality: anything below class A is mixed with the OS CSPRNG before it reaches you. This result actually came from USGS Seismic Feed.

seed token QB6N27715VEWZVJ3QFXGT9XQH0

Re-render keeps this token. Generate replaces it. That is the whole idea.

entropy in

16 bytes

stream used

273 bytes

entropy out

207.6 bits

rejections

duration

12,295 µs

order

corpus

large

corpus words

7,776

rejected total

rejected wrong length

rejected as corpus words

rejected without vowel

exhausted

entropy note

Summed surprisal of the characters actually drawn — the chain is not uniform, so this is lower than log₂(alphabet) per letter.

Seed receipt

Class C · Observational cached

A magnitude 3.7 earthquake, 55 km SSE of Whites City, New Mexico, 6.1 km down — 46 minutes ago.

Carries only tens of bits of genuine surprise. Always mixed with the OS CSPRNG.

source: USGS Seismic Feed
observed: 2026-10-28T21:27:20+00:00
latency: 0ms
mixed with: csprng — always

https://earthquake.usgs.gov/earthquakes/eventpage/tx2026vgingz ↗

Permalink

The token is the seed, so this link recomputes the result rather than looking it up. No database, and it works for as long as v1 does.

This page and /api/v1/g/words.pseudo run the same code. There is no private endpoint.

Invented Words

Result

Measurements

Seed receipt

Permalink