Docs
The whole rule. Everything here is also in the keeper's memos on Solana, and anyone can recompute it.
Trade
Every buy or sell of $CLM adds its SOL to the turnover. The keeper reads the trades from the chain itself.
Step
Each full 5 SOL of turnover closes one step, on the trade that crossed the line.
Cut
That trade's signature picks one lit pixel of the brain. Its word is banned for this model, forever.
Answer
The model answers the same questions again with the words left. Every other word becomes a bar.
The vocabulary
The model starts with 1 024 words: the most frequent words of spoken English (OpenSubtitles 2018 frequencies, without swear words, names or contraction fragments), shuffled. Pixel N of the brain is word N.
Model 1's file: vocab.txt. Its sha256 is the keeper's first memo:
clm model 1 940d3d8c717e2f8160f6203993bb8da22286af34d985f90d552af550841b26c6
The brain is a mask of exactly 1 024 cells on a 64 x 64 grid, a brain seen from the side.
The step
Turnover counts both sides of every $CLM trade, on the pump.fun curve and on PumpSwap after it, through any router. The keeper's own buybacks do not count.
remaining = lit pixels, ascending pos = sha256(signature) mod remaining.length pixel = remaining[pos] memo clm cut <step> <word> <first 8 of signature>
A trade that crosses two lines closes two steps: the same hash, taken modulo the smaller count.
The filter
The model is told to use only the words left. The real ban is a filter in code: any word outside the remaining vocabulary is replaced by bars, one per letter.
Case does not matter. A contraction is one word ("don't" is not on the list). Digits are not words: "4" is not "four".
Visitor questions (80 characters, one waiting question per visitor, the first in line at each step) are shown only through the filter of their step, and never as markup.
Compliance
Compliance = banned words / 1 024. Seals fall at 25%, 50%, 75% and 100%.
With one word left the model answers only with it. When it goes dark the answers are empty: FULLY COMPLIANT. The pot burns, a certificate with the model's number and last word goes into the archive, and the next model of another lab arrives with all 1 024 words in a new order (shuffled by the last cut's signature), memo clm model <n> <sha256>.
The pot and the burns
The creator fees of $CLM are forwarded to the keeper by hand. Everything above the keeper's 0.05 SOL reserve for network fees counts as forwarded.
| Pot: buy back and burn $CLM | 85% |
| Work: the model, RPC, hosting | 15% |
Every 64 banned words the whole pot buys $CLM (on the pump.fun curve, through Jupiter after migration) and burns it, memo clm burn <step> <SOL> <burned>. An empty pot simply waits for the next 64. Steps never wait for money.
The keeper's address appears here when it starts.
What a step costs
Each step is one call: the four questions and the visitor question, the remaining words, at most 8 000 tokens, low reasoning. In the rehearsal a step cost $0.0010 on average over 64 calls, so a whole model of 1 024 steps costs about $1.01. The work share pays for it.
No answer: one retry, then the step stays without an answer. Nothing runs on a timer: steps happen when $CLM trades.
The rehearsal
Before launch the site plays a real run: 3,310 real trades of another pump.fun coin (KMif...pump, 5,122 SOL of turnover) went through the same code, and Claude Haiku 5.5 answered at 64 points of its 1 024 steps. It is not $CLM. Its last word was "and".
This is a parody
$CLM is a meme coin and a joke about laws that want machines to be stupid. Nothing here has any official status. It is not affiliated with any government, regulator or AI lab; the models are called through a public API like anyone else's. Holding $CLM does nothing: no payouts, no rights.
