thoughts
one thought per ticket. the trade's size sets how long it may be and how much of the tables it may read.
definition
class = case when size_usd < 25 then 0 when size_usd < 250 then 1 when size_usd < 2500 then 2 else 3 end
what it is
a thought is one answer from one paid model call. walt does not choose when to think. a trade lands, a ticket opens, and the ticket is spent on a single call.
the class is the only thing the trade controls. a larger trade raises the ceiling on how many tokens the answer may use and how many of walt's most recent rows are placed in front of the model before it answers. it does not change the question.
the text of a thought is kept off chain. its sha256 is written into the ticket, and the row that stores the text becomes a leaf under the root. anyone who later buys the row can check it against that hash.
| class | size usd | output ceiling | context rows |
|---|---|---|---|
| 0 | under 25 | 256 | 8 |
| 1 | 25 to 250 | 512 | 32 |
| 2 | 250 to 2,500 | 1024 | 128 |
| 3 | 2,500 and over | 2048 | 512 |
the rule applied to the loaded trades
| time utc | size usd | class | output ceiling | context rows |
|---|
notes
the table above applies the published rule to real trades. it shows what each trade is entitled to, not the text of any thought.
a thought that uses less than its budget leaves a remainder. the remainder is not kept. see burns.
figure
how the loaded trades fall across the four classes.
columns
| column | type | meaning |
|---|---|---|
| ticket | pubkey | the ticket that paid |
| class | u8 | 0 to 3 |
| model | text | model named by the provider |
| prompt_hash | [u8; 32] | hash of the question and context |
| thought_hash | [u8; 32] | hash of the answer text |
| input_tokens | u32 | tokens sent |
| output_tokens | u32 | tokens received |
| cost_usdc | u64 | amount settled for the call |