The first full run, and what it measured
For the first time the AI played every stage in one hidden, silent, 20 times speed pass. All 24 stages reached their goal and nothing crashed. Nine passes ran over the day, with a round of fixes between each, so every change could be measured against the log.
| Pass | Real seconds | Game seconds | Deaths | Errors |
|---|---|---|---|---|
| 1 | 296 | 5,894 | 21 | 1 |
| 2 | 248 | 4,915 | 15 | 1 |
| 3 | 242 | 4,813 | 16 | 1 |
| 4 | 202 | 4,005 | 12 | 0 |
| 5 | 178 | 3,550 | 12 | 0 |
| 7 | 178 | 3,553 | 11 | 0 |
| 8 | 190 | 3,759 | 8 | 1 |
| 9 | 193 | 3,845 | 9 | 0 |
Speed averaged 19.7 times in every pass. (Pass 6 is left out: it ran while the PC was busy and took 221 seconds and 21 deaths, which turned two stages into watchdog takes.) In pass 9 all eight boss stages ended with the boss beaten, the AI found 37 of the 40 secret rooms and took 55 warp boxes, and it now writes a coverage line per stage so we can see what it did not touch.
Reading the log beat guessing
Several things the AI got stuck on were real level facts it had never been told about:
- The Boiling Room's toffee ribbon only appears once a cat pulls the end, so the AI could not plan across it. That stage went from a long stuck spell to 76 seconds with one death.
- A soaked cat has to shake before it can run. The AI had the wrong idea about this.
- It stood on its own bank shot in Behind the Boards, and stepped into sticky sugar in the Shop Floor.
- The 500 second Last Door was never the boss; two of the "slow" stages were the AI waiting for a boss that was asleep. After the fixes Last Door took 120 seconds and the Top Shelf 136 (were 500 and 508).
It also learned how to shoot a crate that blocks a corridor, how to swim narrow pools without hovering until the air runs out, and how to fight the sleepy Barnaby and the sniffing Snookums. The Shop Floor went from 538 seconds with three boss assists to 222 with none.
Cats inside the scenery
The log found a family of bug: a cat's box put inside solid ground for a frame. Three sources were fixed on the 13th: the AI's own helper teleport, the spot where Snookums goes to groom when she gives up, and the spot where Roo Roo watches when he gives up. A fourth, the place where Barnaby carries the cat, could be inside the counter. The game-side ones were real, because they would have happened to a person too. The lesson in the groom spot was that a "free space" probe narrower than the thing it protects misses exactly the edge cases that matter.
Timing breaks on a busy computer
A timing-based script is only as good as the machine running it. When a pass ran on a busy PC, two stages fell over. Scripted holds and gaps now run on a fixed game-time grid, and the log says when the machine was too busy to trust the run. A full pass is cheap, about three minutes, so it is worth running after every AI change on a quiet machine.
Art and difficulty
Roo Roo now reads as a terrier: a new head with a folded ear and a short muzzle, legs about a quarter shorter, and leftover pixels from the old head cleaned up. An independent reviewer scored it 7 for anatomy and 7 for style, a pass with minors. Small pose fixes for Samus, Pip and Bruiser pass with minors, and 78 real recordings with a loudness rebalance went into the build.
On difficulty, enemies in 3-1, 4-1, 5-1, 6-1, 6-2 and 6-3 were leashed to their posts, and two hearts were added where the logs showed repeated damage. All of this is measured from the AI's point of view: nobody has played the retuned stages or heard the new sounds yet. The one fight the AI still struggles with is Snookums in 2-3. She hides behind the mower and the AI keeps trying to stomp through it.