An AI that knows the map is not testing the game
The Watch AI pilot passed every dive because it went straight to the repair stations in order. Each dive's layout is random, so that was not a skill. The pilot was reading every interactable in the station and planning routes through rooms it had never entered. An expert run was playing like a veteran with a printout.
It now tracks the rooms it has actually been in, only considers objects in those rooms, and when it has nothing known left to do it walks to the nearest unseen room next to one it has seen. Routes cannot cut through rooms it has not entered, unless no known route exists, which covers a hatch that will not give. It no longer knows where the octopus lairs are either: a grab teaches it, after which it plans around that room.
Exploring costs time, and the old exit margin was a flat 25 seconds. The first exploring dive died to the station explosion with 232 seconds still on the clock, because the longer shaft had made that margin too thin. The margin now grows with the ascent: the base margin plus 0.3 of the ascent time.
| before | after | |
|---|---|---|
| route knowledge | every room | rooms visited |
| exit margin | flat 25 s | base margin + 0.3 x ascent |
| test dive | stabilizes every dive | extracted with 44 s left, 1 of 5 repairs, 8 kg |
A planner still on the old ruler
The first exploring dives found something the old pilot never could.
The descent planner, which places events along the shaft, was called with the default depth scale, the original one. With the shaft at its new depth, a collapsed section could be placed on the old numbers, which put it inside the station hub, and drop a solid 20 m block across it. The pilot ended up trapped on one side for the rest of the dive and died to the explosion. Two diagnostic runs confirmed it.
That is not an AI-only bug. A human player could have been walled in the same way. The mission briefing was already using the scaled plan, so the intel said one thing and the world did another.
The fix is one argument: pass the real shaft depth to the planner, so events sit inside the shaft. The depth scale had rescaled everything that read depth through it, and this one call site had never been converted.
The wiggle that dropped it off the ladder
A report that the pilot sometimes could not climb a ladder and kept falling turned out to be two bugs.
- The ladder landing was not kept clear. The facility generator placed containers, repair points, beacons and records, and the view layer placed floor props, with no regard for where a hatch ladder lets you off. A server rack and an alarm panel could sit directly on the landing. Everything now slides out of a 2.6 m clear zone either side of each hatch, both rooms, floor props by their own half-width. A test checks it across 200 seeds.
- The unstick routine was the thing that dropped it. After 3.5 seconds without progress, the pilot "wiggled" sideways to free itself. On a ladder, sideways input above 0.75 lets go of the rungs. A pilot that stalled on a ladder would wiggle, fall, climb again, stall and fall again. On a ladder the wiggle now only goes up or down.
Making noise count for something
The target was that noise should account for at least 30% of what attracts sharks. In the balance report it was only 6 to 13%. Time at depth and the beacons were carrying nearly all of it.
Three constants moved: noise per unit of loudness 0.05 to 0.175, and the time-at-depth rates 0.07 to 0.035 for deep water and 0.02 to 0.01 for mid water.
| player style | noise share of attraction |
|---|---|
| quiet | 55% |
| balanced | 38% |
| greedy | 28% |
| reckless | 24% |
The two loudest styles land below the 30% target. The outcome tiers for quiet and balanced runs did not change.
Smaller ones
- A shark reading a destroyed target. In the frenzied state a shark read the position of an object that no longer existed. The feeding state already guarded against that; the frenzy state never got the same check.
- Tips for players who did not know. The carry limit is back to 12 kg, and the first-dive radio now chains two more lines: search lockers and crates, and weight counts. The playtest feedback was that it was not obvious you could take items down at all.
- Full screen at 1080p. The topside renders at 1280x720, which only fits once into 1920x1080, so it was drawing a 720p picture with black bars all round. Any 16:9 window now stretches it to fill, a 1.5x point-sampled scale with slightly uneven pixels. The dive scene is unchanged: 640x360 fits 1080p exactly three times.
- A ball that never gathered speed. In the new mini golf, a ball resting on a slope never accumulated speed, because the "still" branch reset its velocity every step.
- An asset with no script. A library asset was defined inside another class's file. Unity ties that kind of asset to its own file name, so it loaded with no script attached. It now has its own file.