Being turn-based what decides the order of events? This is determined by speed and action value. Speed is a base stat that all characters and enemies have. For some characters it is higher, while for others it is lower. Speed can be modified by different abilities, traces, relics, enemy effects, etc. The trailblazer's speed is 100 (which is constant even as you level them up), and other characters' speeds are within the neighborhood of 100, roughly \(\pm 20 \). Speed is calculated as $$\text{SPD} = \text{Base SPD} (1 + \text{SPD%}) + \text{flat SPD}$$
The base action value is defined as $$v_b = 10000 / s = c / s$$ where s is the speed and we defined the constant \(c = 10000\). So for the trailblazer the action value is 100. The entity with the lowest action value goes first, and their action value is subtracted from everybody else's action values. After the entity attacks, their action value is reset to the base value \(v_b \).
Speed modifiers affect the action value as follows $$v_1 = v_0 (s_1 / s_0)$$ so it just scales the current action value. Here \(s_1\) is the new speed and \(s_0\) is the original speed. To understand this, imagine \(v_0 = c/s_0 = v_b \) is the base action value (so no subtraction has happened yet). Then we have \(v_1 = c/s_1 \) which is just what the base value would have been if we use the new speed.
Action advancement / delay affects the action value as follows $$v_1 = v_0 - (c / s_0)(\text{Action advancement %} - \text{Action delay %})$$ To understand this, imagine \(v_0 = c/s_0 = v_b \) is the base action value with no subtraction. Then the formula becomes $$v_1 = (c / s_0) (1 - \text{Action advancement %} + \text{Action delay %}) = v_b (1 - \text{Action advancement %} + \text{Action delay %})$$ So this is just another way to scale the action value.
For example, if there is 100% action advancement (with no delay / retardation), the action value becomes 0. Otherwise, if subtraction occurs and we have \(v_0 < v_b\), a 100% advancement can even lead to a negative action value. Negative action values do not really break any mathematical laws. It is just that when subtraction occurs, we are subtracting the lowest action value, so this process will typically leave other action values non-negative. Thus, a negative action value can put you first in line and make you act quicker (assuming other entities are not given the same effect).
Note that above, a decrease in speed or action delay can cause the action value to increase. If we combine both the effects of speed modification and action modification, we have $$v_1 = v_0 (s_1 / s_0) - (c / s_0)(\text{Action advancement %} - \text{Action delay %})$$ Again if we imagine \(v_0 = c/s_0 = v_b \), the above simplifies to $$v_1 = v_b (s_1 / s_0 - \text{Action advancement %} + \text{Action delay %})$$ In other words, the speed modification can act as a kind of action advancement / delay where the percentage is related to the ratio of the new speed with the old speed.
So Star Rail combat follows discrete dynamics. Let \(v_t(i) \) denote the action value at time t of entity i. Initially we have \( v_0(i) = c / s_0(i) \). At each time step, we update the action value as follows: $$v_{t+1}(i) = v_t(i) - \min_j v_t(j)$$ and if \(v_t(i) = 0\) (meaning i had the minimum action value) we reset the action value \(v_{t+1}(i) = c/s_{t+1}(i) \). If there are speed or action modifiers, we will apply those as well using the formulae above. And thus the vector \(v_t\) describes the time evolution of action values for all entities in battle.
What range would an action value be in after the first subtraction? Obviously we can use the formulae above to compute everything exactly, but what if we just want an approximation? We can derive an estimate using the taylor series (first order approximation): $$v(i) = c/s(i) = c/(s(j) + \epsilon) \approx c/s(j) - \epsilon c/s(j)^2 = v(j) - \epsilon c/s(j)^2$$ where we assume \(s(i), s(j)\) the speeds of two entites differ by a small amount \(\epsilon \). The power series gives us a linear approximation relating the two actions values when the speeds differ. Suppose \(s(j) = 100\) like the trailblazer. Then recalling \(c = 10000 \) we have $$v(i) \approx v(j) - \epsilon = 100 - \epsilon$$ In other words, if the speed differs by a small amount \(\epsilon \) from 100, then the action value will also deviate by a small amount \(\epsilon \) from 100. The reason the equations simplify nicely is because the game chose a convenient constant \(c = 10000 = 100^2 \) related to the trailblazer's speed of 100.
So suppose there are just two entities i, j and that \(\epsilon > 0\) so that i is faster than j. Then after the first subtraction the action value for j will become \(v_1(j) \approx \epsilon \). Now suppose \(\epsilon < 10 \) is small compared to 100, then the action value \(v_1(j) < 10\) will be in a range of less than 10 after the first subtraction.
The exact formula for the difference in action values will be $$v(j) - v(i) = c(s(i) - s(j))/(s(i) s(j)) = c \epsilon / (s(i) s(j)) = 100 \epsilon / s(i) < \epsilon$$ Now this holds so long as i is faster than j, so \(s(i) > s(j) = 100\). And this hold true in general, not just for small \(\epsilon \), but even large deviations. That is the deviation in speed will give an upper bound in the deviation in action value. When the deviation in speed is small, then the taylor expansion above tells us that the deviation in action value will be similar in value (in the neighborhood) to the deviation in speed.