Four tiers instead of a switch

Eight of EREME's ninety-nine events are maneuvers: a torn-off antenna, a checkpoint, a fire, and a board of stations where the player assigns the crew. How well the maneuver is held is computed from the people: their traits and their marks make an aptitude per station, the sum makes the hold, the hold lowers the failure threshold. Then one roll.

When the system was born, that roll was a switch: above the threshold it passes, below it breaks. And the measurement recorded in the commit showed the flaw before any player could live it: at the best placement, sixteen options out of twenty-four dropped to exactly zero chance of failing. The commit that fixes it states the problem in one sentence: "A game about people's memory, where good placement cancels luck, loses what luck writes on them."

The fix touches neither the roll nor the threshold: it rereads their distance. A single roll, the same as before, and its position relative to the threshold sorts the result into four tiers: it fails badly, it fails, it barely holds, it holds outright.

The shape of the boundaries is the real gem. The catastrophic tier is a fraction of the threshold, forty percent: multiplicative, it vanishes with the threshold, so a perfectly held maneuver can never turn into a disaster. The strain tier, on the other hand, is an additive band above the threshold, fixed: even at zero threshold, about a quarter of results still "barely hold". Certainty disappears without any artificial floor: playing well protects from the worst, nothing protects from the price.

The commit's calibration, three hundred rolls per option: at the best placement, two thirds hold outright, a quarter barely hold, nine percent fail; at the worst held placement, half fail; with nobody at the stations, outright success drops to two percent. "Placing well doubles the clean success and divides failure by five." Placement remains the dominant lever; it just stopped being an off switch on luck.

"It barely holds" delivers the same outcome as the clean success: same texts, same effects. The difference is a person: the weak link, the one whose aptitude is worst at their own station, loses one notch of morale. And the Chronicle says it with an aside, chosen by the station's trade, not by the person: ten families of three lines in data. "{name}'s hands still tremble from it." for a risk station; "{name} had warned, once, in a low voice." for caution. At the bottom tier, when it fails badly, the aside gains a dry suffix: "And it costs." The plain failure, for its part, delivers its failure branch without an aside.

The internal tier names never appear on screen. The tier only reaches the player through text: the delivered outcome, the aside, the suffix. And the aside's variant is a deterministic hash of the day, the person and the station: "a Chronicle that changed on re-reading would cast doubt on everything else."

The system's safeguard has its own story, the best of the batch. Its first clause required that the catastrophic tier recede when you place well. True by construction, since that tier is a fraction of the threshold: the sabotage, setting the fraction to one, left the test green. The rewritten clause requires the useful opposite: that catastrophe exist when you place badly, at least eight percent at the worst placement, and setting the fraction to zero makes it fail. A gate is proven by sabotage, never on its word.

The same safeguard rejects a station without an aside family: every trade that can strain someone must have its three lines in data, or error. "A mute morale loss is worse than no tier at all."

The tiers made three older flaws visible, fixed the very next day. The failure branches carried no effects, and the only two maneuver deaths in the corpus lived in the successful branch: holding all three stations killed four times more often than botching. The catastrophic tier received its own outcome, rewritten texts, heavier effects, and it inherited the deaths. Then, in the next day's register, a damning measurement: five hundred thirty-three maneuvers auto-resolved, zero failures, zero catastrophes, because the autopilot picked the zero-threshold options; I settled on a five percent floor, the gradient intact. Finally, the drawn helper was picked from the whole crew without reading the assignment: the player placed three people, a fourth one fell.

The rework did not add a die: it reread the same die. That is what makes it compatible with everything else, determinism first: same seed, same tier, same replays. When a binary system lies about the margin, the temptation is to stack rolls and modifiers. The sober version is to look at where the roll landed relative to the threshold, and to have that distance told by someone whose hands still tremble from it.