One row per strategy, never per account: where two arms place the identical bet they are one observation at two stake sizes, so n is the larger of the two and nothing here is summed. The edge is the equal-weight mean of profit over stake, which matches the win rate arithmetic exactly at even money; total-profit-over-total-stake is shown beside it as the money path and runs higher wherever a bankroll compounded.

The paper record is days old; the forward log is weeks old on the same rule. Where they disagree the longer sample is the one to read. Neither column says the book would accept the stake.

These are not failures. They bet games whose house margin is 20-55%, and they run to show that the engine can tell an unbeatable game from a beatable one and what 5% stakes do to a bankroll on one. A control reading -20% is the control working.

Each line is one strategy’s flat-5% arm, indexed to 100 at its own starting bankroll. There is no total: summing these would add together experiments that have nothing to do with each other, and on a rupee axis Game B would bury everything else. Controls are plotted on their own numbers in the table above rather than here.

The same thing in rupees, one panel per strategy, each on its own scale. They cannot share an axis: Game B is three orders of magnitude above the rest, so a single chart would draw one curve and eight flat lines on zero. Each panel prints its own floor and ceiling, so the shape is never mistaken for a magnitude. Every-arm, flat 5% sizing.

Whether the P&L above may be read as evidence. A bet is only real if it was struck on a quote actually on screen, gated on information available beforehand, and settled against a published result.

A trade is not a bet. It is opened at one price and closed at another before settlement, so who wins is irrelevant and only the two prices matter. Exposure is the liability for a lay and the stake for a back, and a closed position carries none.

A one-tick target with a time stop has an asymmetric payoff: the win is capped at one grid step while the loss runs as far as the price goes, which puts break-even near 0.68 rather than at even money. Measured across 591,262 observations, no grid price clears it. A value arm sitting at zero has found nothing worth taking, which is the result rather than a fault; the every arm keeps trading as the control and prices what the refusal is worth.

Every settled bet, scored with its sample size and its t against zero. A rule is a candidateuntil n is large enough to tell it from noise — roughly |t| > 2 — and a young ROI that looks spectacular usually converges toward its prior as n grows.