Position size is a policy, not a feeling.
Of everything a trading terminal claims to do, sizing is the one decision it can actually enforce. Radon sizes every options position by fractional Kelly, computed from the structure's maximum loss and the signal's odds, then hard-capped at 2.5% of bankroll per position. The cap is a ceiling, not a target.
Full Kelly assumes knowledge you do not have.
The Kelly criterion answers a precise question: given the odds of a bet and the probability of winning it, what fraction of a bankroll maximizes long-run growth? The formula is short. For a bet that pays b times the amount risked, with win probability p, the optimal fraction is f = p - (1 - p) / b. Nothing in the arithmetic is controversial. The problem is what it takes as given.
Kelly assumes p is known. In markets it never is. Radon's probability estimates come from flow: dark-pool and OTC prints scored as accumulation or distribution, then checked against regime models. That is an estimate with error in it, not a known quantity. And full Kelly does not fail gracefully when the estimate runs hot. Betting past the true optimum cuts growth and deepens drawdowns, and betting far past it turns a positive-edge strategy into a losing one.
The asymmetry is the point. Sizing below optimal Kelly costs a modest amount of theoretical growth; sizing above it destroys capital at an accelerating rate. When the input carries error, the rational response is to shade down. A fraction of Kelly trades growth for survivability against estimation error, and Radon makes that trade as standing policy, decided before any single trade is on the screen.
The formula needs two numbers. The terminal supplies both.
The net odds b come from the structure itself. Radon routes defined-risk options structures: 37 of the 58 structures in its catalog carry a defined risk profile, and only those clear the convexity gate. For a call debit spread the catalog records maximum loss as the debit paid and maximum gain as the strike width minus that debit. Divide maximum gain by maximum loss and b exists as a fact of the position, not a forecast. This is where max loss comes from, and it is why defined risk comes first: an undefined loss makes b, and therefore Kelly, uncomputable.
The probability p comes from the signal. The flow score supplies the base estimate, and the corroboration milestone adjusts it: agreement between the flow read and the four regime models raises conviction, while a signal fighting the regime is downgraded, not ignored. The result is still an estimate. That is the argument of the previous section, and the reason the formula's output is treated as an input to a cap rather than a command.
| Structure | Call debit spread · defined risk |
| Max loss | Debit paid = $500 |
| Max gain | Width minus debit = $1,500 |
| Net odds | b = 1,500 / 500 = 3.0 |
| Signal probability | p = 0.45 |
| Full Kelly | f = 0.45 - (0.55 / 3.0) = 0.2667 = 26.7% |
| Quarter Kelly | 0.2667 / 4 = 6.7% |
| Radon cap | min(6.7%, 2.5%) = 2.5% of bankroll · cap binds |
Full Kelly at 26.7% of bankroll on a single options spread is the formula operating exactly as designed on an input it should not trust. Quarter Kelly shades it to 6.7%. The Radon cap then binds at 2.5%. Each layer between the raw formula and the routed order exists to absorb the error in p.
A ceiling that does not move.
Even shaded, Kelly is a per-bet answer to a portfolio question. Three things break the per-bet frame. Positions correlate: several flow signals expressed in the same regime are not independent bets, and Kelly's arithmetic assumes they are. Estimates err, as the first section argued. And regimes shift, which changes p after entry while the position size stays where it was set. The hard cap is the fixed backstop against all three, and it holds whatever the formula reports.
The policy is Gate 3 of Radon's four execution gates: fractional Kelly, hard cap 2.5% of bankroll per position. Milestone 6 of the evaluation pipeline enforces it, and the pipeline document describes the enforcement in four words: hard cap, not advisory. A structure that has not been sized under the cap does not reach milestone 7, and milestone 7 is the only place an order gets routed.
Conviction still matters, but it operates inside the ceiling. A strong signal with regime agreement sizes toward the cap; a downgraded signal sizes well under it. What conviction can never do is move the ceiling. The number is the same on the best setup of the quarter and the most ordinary one, which is what makes it a policy rather than a mood.
The cap is enforced where the order is built.
A sizing rule that lives in a document is a suggestion. Radon's lives in the execution path. Before any order routes, the mandatory pre-trade chokepoint computes maximum loss and margin impact on the assembled combo, exactly as it will be submitted. The check is not optional and cannot be bypassed. There is no route around it, so there is no routed order it has not seen.
What survives the chokepoint is journaled. The trade journal is the canonical store of executed positions, and the audit trail records the full decision chain: the signal that opened the evaluation, the structure that expressed it, the Kelly arithmetic and the cap that sized it, and the check that cleared it. Sizing is one step inside the method around the sizing step, and the journal is where the whole chain becomes inspectable after the fact.
Sizing, answered plainly.
The questions traders ask about the cap, in the order they ask them.
Why cap position size below what full Kelly allows?
Because the probability feeding the formula is an estimate built from flow data, not a known quantity. Overestimating it with full Kelly produces overbetting, and overbetting is the failure mode that compounds: it deepens drawdowns and can turn a positive edge into a losing strategy. A fraction of Kelly under a fixed ceiling gives up some theoretical growth to stay solvent against estimation error.
What if Kelly suggests less than 2.5%?
Take the smaller number. The cap is a ceiling, not a target. When fractional Kelly on a weak or downgraded signal works out to less than 2.5% of bankroll, that smaller figure is the position size. Nothing in the policy rounds a small position up to the cap.
Does Radon size positions automatically?
It computes the size and enforces the cap. Fractional Kelly is calculated at the sizing milestone from the structure's maximum loss and the signal's odds, and the mandatory pre-trade check verifies maximum loss and margin impact on the assembled combo before anything routes. Conviction can adjust size within the cap; it cannot route past it.
Can I override the 2.5% cap?
No. A hard cap that can be overridden is advisory, and the evaluation pipeline is explicit that this one is not: milestone 6 enforces the cap before milestone 7 will route an order. The value of the ceiling is that it holds on exactly the trade where overriding it feels most justified.
Run it against the live tape.
The sizing policy is visible end to end in the demo: signal, structure, Kelly arithmetic, cap, and the pre-trade check that enforces it.