Earned, but too rare
Since August, the one thing in combat that paid for effort was the committed hit. Swing fast enough and wide enough at the moment of contact, and the blow did 1.5× damage with its own spark and sound. When I last wrote a…
Since August, the one thing in combat that paid for effort was the committed hit. Swing fast enough and wide enough at the moment of contact, and the blow did 1.5× damage with its own spark and sound. When I last wrote about it, the bar had been following me upward, so I froze it for the session to make it hold still.
Frozen, it held still, and it still did not work. After pass 11 I said so in the headset: "I really wanted to heal there, but I was not getting committed hits when I was really trying."
The shape was the problem
In pass 11 I was trying hard, and the gate was frozen at 79.13 m/s and 229°. Of 44 landed hits, 16 committed. Eleven missed on speed only, six on sweep only, and eleven on both. My speed and sweep correlate at 0.33, so a committed hit meant clearing two thresholds on two axes that barely move together. A miss by 1 m/s paid the same as a miss by 40, which was nothing. Every retune I had made (0.75, then 0.65, then 0.55, then live to frozen) moved where the wall was. None of them changed the fact that it was a wall. And Mending, the affix that heals, was hung off it.
A slope instead
On 23 September Claude replaced it with a crit (Combat/Crit.cs). Each landed swing gets a quality q: the geometric mean of how its speed and its sweep rank against a frozen pool of my own 150 swings. A fast wiggle or a wide shove scores near zero, which keeps what the conjunction was for and drops the cliff.
chance = 5% + 70% × q² (zero under 90° of sweep; 75% at best)The 75% top is deliberate. A certain crit at the top would just be the gate again, set higher. A crit carries the old ×1.5, the old cues and Mending. The near-miss chime is gone, because there is no line to nearly miss. There is also bad-luck protection: each miss adds pressure to the next roll, and a crit spends it.
Pass 12: about right
I wore it that evening: 41 hits and 12 crits, 29% against a predicted 30%, with never more than 4 swings between crits. On the in-headset panel my answers were "mostly earned", "steadily", "about right", and that Mending on crits felt good.
The "mostly" was in the log. Two crits landed on limp swings, one at 10 m/s with quality 0.11, each cashing in pressure that a strong miss had just built. Claude added Crit.Payout the same night. A swing now cashes in pressure only in proportion to its own quality over the average quality of the misses that built it, capped at 1. In simulation the weakest quarter of swings fell from 15% to 4.5%, against a stated 6.4%.
Pass 13: earned, but too rare
The next evening I got 8 crits in 60 counted swings, 13.3%, and the longest drought was 11. The button I pressed read "Earned, but too rare now."
One of the two reasons was measured. The pool is frozen for the session, which was last time's fix. That day my median swing was 61.4 m/s over 219°, against the pool's 72.4 m/s and 234°, so an ordinary swing ranked about 0.31 instead of 0.5. Either the pool was built from harder swinging or I swung lighter that day. The log cannot say which, and neither can I. Freezing the pool stopped it running away from me. It did nothing about me drifting away from it.
The other reason was mechanical. Each miss added its own step of pressure, and a low-chance miss adds a tiny one, so a drought of ordinary swings barely wound the protection up. The swing that finally ended the eleven was a good one: q 0.67, with a stated chance of 36%. The roll came up 70, and only the drought's pressure got it over.
The comment that said nine
The protection's docstring said that at 30% a crit is "never more than nine" swings away. That only holds at a constant chance. Mine never is, and I had gone eleven. A comment that asserts a fact is a test that never runs.
Of the three fixes Claude offered (raise the curve, re-rank quality, or put a floor under the pressure) I chose the floor. Every landed miss now builds at least the pressure of a 30% miss. That top-up counts at a neutral quality of 0.5, so Payout still stops a limp swing from cashing it in. Replayed over pass 13's own swings (crit_sim.py), the rate goes from 14.6% to 21.6% and the median longest drought from 12 to 7, while limp swings stay just under their stated chance (5.2% against 5.7%).
Then Claude went back to the comment and worked out what the floor really guarantees. After eight landed misses, the next swing that cashes in fully is a certain crit: 8 × 0.119 + 0.057 = 1.009. After seven it is 0.890. The margin is under 1%, and a floor of 0.29 would make it nine. It is not a cap in swings, though. A swing weaker than the drought cashes in only its share, so if you keep swinging limply you can stay dry indefinitely. Both halves are tests now, and nudging the floor to 0.29 turns the first one red at 0.950.
What is still open
Nobody has worn the floor yet. Pass 14 is set up and waiting, so every number after pass 13 is predicted.
The floor does not bring pass 13's swings back up to 29%. If I say "too rare" again, the lever left is the quality baseline. Before anyone touches it, I want to know which story about 61 against 72 m/s is true. The chance the log prints also now undersells the real rate for anyone swinging below their pool. That is deliberate, but it means the number I read is not the number I feel.
What I am keeping from this is that no amount of tuning makes two weakly correlated thresholds feel fair. You replace the wall with a slope, and then you check what the slope does on a day you swing like somebody else.