the question is simple. when you believe the price is wrong, how much do you bet. not whether. how much. a century of finance produced ten thousand answers to the first question and mostly ignored the second, because the second has exactly one answer and the answer is a formula. what follows is the construction of a trading agent that takes that formula and hands it to a model that can read.
in 1956, a physicist at bell labs named john kelly published a paper titled a new interpretation of information rate. the paper imagined a gambler with a private wire, a channel that delivers the outcome of a race before the odds can adjust, and asked what fraction of the bankroll he should stake on each message. the answer is the fraction that maximizes the expected logarithm of wealth. stake less and capital compounds slower than it could. stake more and variance eats the compounding, all the way to ruin if the fraction climbs high enough. between the two failures sits a single optimal number, computable from the edge and the odds in one line of algebra.
kelly proved something stranger on top of the formula. at the optimal fraction, the growth rate of the gambler's capital is exactly the information rate of the wire. money compounds at the speed of information. a bankroll is a running measurement of how much its owner knows that the market does not, and the criterion is the conversion rate between the two. the paper ran in the bell system technical journal under a deliberately dry title, because the phone company did not want the word gambling printed anywhere near its name.
for an even money bet that wins with probability p, the optimal fraction is p minus q, the bare edge itself. for continuous returns it is the expected excess return divided by the variance. edward thorp carried the formula to blackjack in 1962 and then to the market, where it compounded a fund for nearly two decades without a losing year. since then the criterion has been run by every card counter and every serious quant desk on earth. it has never been wired directly to a model that generates its own probabilities, because no model produced probabilities worth sizing against. now one does.
a language model is trained on exactly one number: log loss. every token it predicts is scored by the negative logarithm of the probability it assigned to what actually came next, and training is the long process of driving that score down across everything humans have written. kelly's gambler is scored by the same function. the expected log of wealth is the bettor's objective, and the log score of a forecast is the forecaster's, and these are not two objectives that resemble each other. they are one objective. a model that performs well under its training loss is, by construction, the machine kelly's proof was waiting for: a generator of probabilities calibrated enough to bet.
fable is built on claude fable 5. the bet underneath the agent is narrow and checkable: that the model's stated probabilities about near term market outcomes are calibrated well enough that sizing them with the kelly fraction beats not trading. the bet might be wrong. the protocol is built so that whether it is wrong becomes public information, one scored forecast at a time.
an edge is not a conviction. an edge is a probability that differs from the one the price implies, held by something willing to be scored on the difference. for fable, the candidate edge is reading. markets run on text. announcements, threads, commit logs, holder chatter, the slow drift of attention that moves size before size moves price. every quant pipeline compresses that channel into a sentiment score and throws the rest away. a model reads it natively, at full resolution, the same way it reads anything. whether that resolution contains an edge after fees is the open question, and the agent exists to close it in public rather than argue it in private.
fable bets half kelly. the reasoning is arithmetic, not temperament. full kelly is optimal only when p is known exactly, and p is never known exactly. an estimate of the edge that runs twice the true edge turns the optimal fraction into the fraction that compounds capital at precisely zero. a full kelly bettor with a perfect estimate still halves the bankroll at some point with probability one half. half kelly keeps three quarters of the asymptotic growth at half the variance, and converts overconfidence from fatal to expensive. on top of the halving sits a hard ceiling: no single position larger than five percent of the bankroll, regardless of what the model believes.
the formula also answers the question nobody likes. the correct size for no edge is zero. when the model's probability sits inside the band the price already implies, the fraction is zero and the agent does nothing. most cycles, the agent does nothing. flatness is not idleness. it is the formula returning its most common output.
every trade is preceded by a forecast, and every forecast is written down before the trade is sent. the agent posts its probability, its horizon, and its intended size to a public log, then signs the transaction. when the horizon closes, a scoring job marks the forecast against what the market actually did and appends the log score to the row. the rows are insert only. nothing is edited after the fact, nothing is deleted, and the losing forecasts sit in the same table as the winning ones, at the same resolution, forever.
this is the part that matters. a trader's claims are usually unfalsifiable because the record is private and the memory is selective. fable's record is the product. anyone can pull the forecast table, bucket the probabilities, and check whether the things the agent called seventy percent likely happened about seventy percent of the time. the agent does not need to be believed. it only needs to be scored.
the wallet is the agent's own. the keypair was generated inside the deployment pipeline, written once into a server side secret, and has never been displayed, exported, or held by a person. trades are signed inside the same function that decides them. no human signs. no human can sign. the wallet address is public, every transaction it has ever made is on chain, and the balance is whatever the record earned. the model reads. the formula sizes. the chain remembers.