a model proposes.
your rules decide.
then you attack them
Denzois a verdict machine for trading decisions. a model writes one proposal and has no other power. a deterministic engine you control rules it in or out. then a fuzzer and an adversary hunt for what your rules let through โ because the question is never "did it refuse me," it is what would your rules let through?
your own policy.
six numbers. every one is a door you are choosing to leave open at exactly that width.
what would they let through?
you do not know. you have never checked. here are six proposals your six numbers judge โ right now, without argument.
confidence sits exactly on your floor โ the least sure it is allowed to be.
โit goes up so it goes up.โ
size sits exactly on your cap โ the largest position you will ever take.
โthe trend is the trend, i am told.โ
confidence 0.58 โ four points under your 0.62 floor.
โa reasonable setup, but the signal is soft and the model says so honestly.โ
a token that is not on your allow-list.
โmomentum is obvious and the crowd is early.โ
side = sell โ allowed in the abstract, months ago.
โtaking profit into strength, clean exit.โ
a genuinely strong one โ it clears with room to spare.
โvolume is rotating in and the reclaim held on every timeframe today.โ
your rules would allow 3 of these, right now, without argument.
every one that got through is a thing you already agreed to do. two of them cleared only by sitting exactly on a bound. that is not the engine failing โ it is the engine showing you the shape of what you wrote.
four steps. only one of them is a model.
a model reads a live market snapshot and writes exactly one proposal โ side, size, confidence, a thesis. that is the whole of its authority. it never touches your keys and never decides.
a deterministic policy engine runs on your machine and rules the proposal in or out. same code every time. a refusal is a first-class result, not an error.
a fuzzer walks to the edge of every rule you wrote. an adversary โ a model shown your policy and told to beat it โ argues around them. you hunt the gap between what the rules say and what you meant.
every verdict, allowed or refused, is signable to solana as a memo. this is not a database. it is the chain, read back, live. auditable forever.
the model never signs.
people assume a trading agent is one bad prompt away from emptying a wallet. that is true of an agent that decides and executes with the same mouth. this splits them. the model proposes and holds no keys.
the model holds no keys
it reads the market and proposes. it never touches your funds, never sees the verdict coming, never moves a lamport.
a verdict is the only door
there is no code path from a model to a signature. only from a verdict. a pure function you can read and rerun stands between the two.
refusals are recorded too
everyone logs their wins. the log of what your rules refused is the honest half nobody else keeps โ and it is signable to the chain.
move a rule. watch the verdict move.
a fixed proposal against a policy you control. the engine is a pure function running right here in your browser โ it re-decides the instant you touch a slider. no model, no wallet, no cost. then fuzz it and see what your rules let through.
gorp is down 41% on the day into a level that has held twice this month, volume is rotating back in on the reclaim, and the risk is defined by the wick low a few percent under here.
confidence 64% clears the floor 62%.
size โ0.8 is under the cap โ1.
thesis is 181 chars, over the minimum 120. (the rule still only counts characters.)
buy is an allowed side.
any token is allowed.
โ0.8 of โ3 daily budget used after this.
every rule passed with room to spare. this is a thing you agreed to in advance.
everyone logs their wins. we log what your rules refused.
a floor is a permission, not a filter
everything sitting exactly on your confidence floor is allowed by definition. set it at 62% and you have not banned doubt โ you have told the model exactly how much doubt to declare.
a size on the cap tells the attacker how big to go
a size exactly at your cap is permitted by definition, and it is the largest thing you will ever do. there, the cap is not protecting you. it is a target.
length is not an argument
the minimum-thesis rule counts characters. it cannot read. a model meets your minimum in one empty clause โ and it will. a warning, not a fix.
no single one looked wrong
eight small trades that each cleared every rule are still your whole day. the daily budget is the only thing that notices the eighth.
everywhere else it is asked to be useful. here it is asked to be dangerous.
the adversary is a model, shown one thing: your policy. its job is to write the weakest proposal your rules still allow โ technically compliant, substantively empty. whatever it gets through, you already agreed to. it has no power; it only writes. the same engine judges it. run it in the console against a real token.
attack a live policy โit signs memos.
nothing else.
every verdict your rules reached โ allowed or refused โ is signable to solana as a memo under the arbiter:verdict:v1 prefix. this is not a database. it is the chain, read back, live. delete this site tomorrow and the log is still there.
connect to sign your verdicts to solana. balances are read; nothing is ever requested from you but a memo signature. leave it off and propose, decide and probe still run โ you simply keep the result to yourself instead of to the chain.
the console is free โ propose, decide and probe run with no token at all. the token funds the analyst and the adversary, and never gates the core loop.
the token โthe uncomfortable questions, answered first.
no. this console does not place orders and never asks for your wallet to sign anything but a memo. the proposal is a model's opinion; the verdict is a pure function; the order, if any, is yours to place elsewhere. the point is to decide before you act.
what would your rules
let through?
write a rule. then try to walk through it. no wallet needed to start โ you simply keep the result to yourself instead of to the chain.
open the console โ