Fire control Cut 6d map: where a hit lands is a dart throw, not a coin flip
Precision is currently a coin flip between landing exactly on the aimed item and landing uniformly anywhere on the hull, so aiming at a stern thruster scatters its misses across the bow. Operator direction: misses should cluster near what you aimed at, like throwing a dart at the grid -- and aiming should therefore be a real tradeoff, where a thin limb risks missing the ship entirely and centre mass is the safe play. One Gaussian kernel over the hull schematic answers both halves: its on-hull mass prices the stray into HitProbability, so the risk is legible on the HUD before the trigger, and its per-cell weights place the hit. One function, so the probability and the placement cannot drift apart -- Cut 3's named risk was exactly two functions answering where a shot goes. Records the rejected draft that clamped a stray dart to the nearest hull cell, which would have quietly removed the tradeoff, and the two consequences worth catching early: pOnHull may double-charge with the existing pSpread term, and every AI gates on this number so they will hold fire more once it lands. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
parent
35b8b6a550
commit
3bf1e40795
1 changed files with 118 additions and 0 deletions
|
|
@ -1297,3 +1297,121 @@ Cut 2 ship code without data.
|
|||
- The report carries the full table — cut, mutation, died/survived, and for any
|
||||
survivor the fixture defect behind it.
|
||||
- `dotnet test` green, with the count.
|
||||
|
||||
---
|
||||
|
||||
## Cut 6d: where a hit lands is a dart throw, not a coin flip
|
||||
|
||||
Date: 2026-09-19. Operator direction, 2026-09-19.
|
||||
|
||||
### The defect
|
||||
|
||||
`FireControl.Commit` places a hit like this:
|
||||
|
||||
```csharp
|
||||
if (aimedCells != null && aimedCells.Length > 0 && random.NextFloat() < shot.Precision)
|
||||
cell = aimedCells[random.NextInt(aimedCells.Length)];
|
||||
else
|
||||
cell = coords[random.NextInt(coords.Length)];
|
||||
```
|
||||
|
||||
`Precision` is a coin flip between *exactly on the aimed item* and *uniformly
|
||||
anywhere on the hull*. Aim at a stern thruster with Precision `.35` and the 65%
|
||||
of hits that miss it are spread evenly over the bow, the cockpit and everything
|
||||
else — no nearer the thruster than if the player had never aimed at all. There is
|
||||
no spatial term anywhere in it.
|
||||
|
||||
This is the same shape as 5.1's Tracking cliff: a quantity that should be a
|
||||
gradient implemented as a binary, so the failure case is maximally wrong rather
|
||||
than slightly wrong.
|
||||
|
||||
Operator, 2026-09-19: "I would presume that aiming for an item in the stern would
|
||||
still tend hit in that general area even if it misses the item. Like throwing a
|
||||
dart at the grid."
|
||||
|
||||
### The rule
|
||||
|
||||
One Gaussian kernel over the target's hull schematic, in cell units:
|
||||
|
||||
- **Aim point.** The centroid of the aimed item's cells, or the hull's
|
||||
`CenterOfMass` when nothing is aimed at. One path — the uniform-random branch
|
||||
is deleted, not demoted.
|
||||
- **Sigma.** Derived from the frozen `shot.Precision`, higher Precision giving a
|
||||
tighter group. Authored so that the unaided floor (`UnaidedPrecision`, joining
|
||||
`UnaidedAccuracy` and `UnaidedTracking` in `GameplaySettings`) is a sigma wide
|
||||
enough to read as spraying at the silhouette.
|
||||
- **Weights.** `w(cell) = exp(-d² / 2σ²)` for each occupied hull cell, `d` the
|
||||
distance from the aim point.
|
||||
|
||||
### The tradeoff, and why it cannot be a second miss stage
|
||||
|
||||
Operator, 2026-09-19: "It would also mean that aiming can be a tradeoff; if
|
||||
you're not precise enough, you're better off aiming center mass than risking
|
||||
misses on a thin limb."
|
||||
|
||||
That tradeoff only exists if a stray dart can miss the ship. An earlier draft of
|
||||
this section clamped a stray to the nearest occupied cell, which destroys it:
|
||||
aiming at a wingtip would cost nothing and merely redistribute damage inward.
|
||||
|
||||
But resolving the stray as a *second* miss stage would break R3 — one roll
|
||||
decides damage — and make `HitProbability` a lie, since a shot could pass the
|
||||
roll and then fizzle on placement. So the off-hull mass is priced into the single
|
||||
roll instead:
|
||||
|
||||
```
|
||||
pOnHull = Σ w(cell) / (2πσ²)
|
||||
```
|
||||
|
||||
the share of the kernel's total mass that lands on metal.
|
||||
|
||||
- `HitProbability` gains `pOnHull` as a factor. Aiming at a thin limb with a
|
||||
loose group drops the number **on the HUD, before the trigger** — the tradeoff
|
||||
is legible at decision time rather than discovered afterwards.
|
||||
- `Commit` draws the cell as a weighted pick over the same `w(cell)`. A hit
|
||||
always lands on metal, because the chance of not landing on metal was already
|
||||
charged at the roll.
|
||||
|
||||
One kernel, computed by one function, used for both the probability and the
|
||||
placement. Cut 3's named risk was two functions answering "where will this shot
|
||||
go" and eventually disagreeing; this does not reintroduce it.
|
||||
|
||||
`pOnHull` depends only on the aim point, the hull shape and the frozen
|
||||
`Precision`, so it folds into `PBase` at fire time and the freeze discipline
|
||||
(R10, Q6) is unchanged.
|
||||
|
||||
### Consequences, named rather than discovered
|
||||
|
||||
- **Every hit in the game moves**, not only aimed ones: unaimed fire becomes
|
||||
"aim at the centre of mass with the worst sigma", so big central sections soak
|
||||
more fire than extremities. That is the intended change, and it is a balance
|
||||
change.
|
||||
- **A fat hull is inherently easier to hit than a needle**, for free.
|
||||
- **The two targeting products' authored `Precision` values (`.05-.2` and
|
||||
`.35-.55`) become meaningless as written**, because they are probabilities and
|
||||
this makes Precision a grouping tightness. They are re-authored in the same
|
||||
pass as Cut 6c's reciprocal `Resolution`.
|
||||
- **Double-charging risk.** `HitProbability` already carries `pSpread`, which
|
||||
penalises a small target against a wide weapon spread. `pOnHull` also
|
||||
penalises small targets. Check whether the two charge the same thing twice
|
||||
before authoring numbers; if they do, one of them owns it and the other goes.
|
||||
- **AI willingness shifts.** `Combat.cs` and `TurretController.cs` gate on
|
||||
`AgentMinHitProbability` against this number, so NPCs will hold fire more than
|
||||
they used to the moment `pOnHull` lands. That is a tuning pass, not a bug, but
|
||||
it will show up in the first smoke and should not be mistaken for broken AI.
|
||||
|
||||
### Verification
|
||||
|
||||
- `AimingAtTheSternHitsTheStern`: with a mid-range Precision, hits aimed at a
|
||||
stern item land in the stern half far more often than the bow half. Mutation:
|
||||
restore the uniform-random branch. Must die. This is the assertion the shipped
|
||||
coin flip would fail.
|
||||
- `ThinLimbCostsHitChance`: `HitProbability` against an aim point on a
|
||||
single-cell extremity is strictly lower than against the centre of mass, at the
|
||||
same Precision — and the gap widens as Precision falls. Mutation: drop
|
||||
`pOnHull` from the probability. Must die.
|
||||
- `PlacementAndProbabilityShareOneKernel`: the empirical on-hull rate over many
|
||||
seeded draws matches the `pOnHull` the probability reported, within tolerance.
|
||||
Mutation: perturb sigma in one of the two call sites only. Must die — this is
|
||||
the test that would catch the two functions drifting apart.
|
||||
- `EveryHitLandsOnMetal`: no committed hit ever resolves to an unoccupied cell,
|
||||
across seeds and aim points including extremities.
|
||||
|
|
|
|||
Loading…
Reference in a new issue