The wrong clock

Deep Descent is a 2D side-scrolling extraction survival game: you dive to a failing station and try to get out before the countdown ends.

Six dives, no clean run

The first set of run logs came from the person building the game, who went down six times and never came home clean.

Three things in that table matter more than the deaths themselves.

The player runs out of air, not time. The countdown is 15:00. The base tank holds 9:00. Three of six runs ended in drowning with roughly five to six and a half minutes still showing on the big clock. The HUD displays both numbers, but the countdown is the large one, so it is the one people watch.

Boosting is the loudest thing a player does. It is the top noise source in five of the six runs, and run 6 had ten sharks around it by the time it tried to ascend. Noise is what draws sharks. Nothing in the game told the player that boosting was the price.

No run reached a frenzy. The late-game systems are not what is beating people. The basics are.

The conclusion we drew was to leave the difficulty alone and make the rules learnable. A game that kills you for reasons you cannot see is not hard, it is opaque.

A pilot that explains itself

The Watch AI, the pilot that plays the game while you watch, got a commentary panel. It maps what the pilot is currently doing to a one-line rule ("sharks read noise: hold still, keep the light on it"), shows each rule once per session, and spaces them at least 5.5 seconds apart so it reads as a lesson and not a ticker.

Writing the rules down exposed two that were not true. The pilot never boosts, so the lessons that implied it did were corrected to say what actually happens.

The pilot also gets a debug log: every change of action, a state snapshot every two seconds, findings worth a human's attention (low air, low health, new breaches, a shark close and attacking), how each dive ended, and any engine error. It raises a STUCK? warning when the pilot has neither moved nor changed what it is doing for 25 seconds.

That last one exists because of a failure we could not explain. In one captured session the pilot sat on the step "Down to the Dive Bay" for more than four minutes. Reading the ship pilot's stair logic, it held up on paper: pick a stairway, walk to it, hold down to climb. When reading the code does not find a stall, the useful move is to make the system report where it stalled, along with the deck, the position and what it thought it was doing. The log is written but has not yet been run against a real stall.

Shortening a shaft without touching forty places

The shaft got shorter, and about 40 places in the code had depths hard-coded for the original: fauna bands, light and water colour, event ranges, how fast attraction builds with time at depth, sound.

Changing one number would have cut the descent off at the top and left most of the content unreachable. Instead there is now a depth scale that maps the real shaft onto the original reference scale, so every threshold written for the old depths still works and the whole descent compresses rather than being truncated. The systems that read depth convert through it, and the game sets the scale at launch.

Two balance test expectations moved as a result. The quiet-run time went from 29 to 23, and a noise ceiling from 12 to 16. Both were a consequence of a shorter trip, not of anything breaking.

Upgrades should add, not replace

The carry limit was tuned three times in one afternoon: 20 kg, then 17, then 13. Weight is a deliberate lever here. Every kilogram carried costs 2% speed, down to a floor of 60%, and the starting kit alone weighs 1.6 kg.

The harness upgrades were written as absolute capacities, 16 kg and 20 kg. That works until the base moves: with a 20 kg base, the first upgrade would have lowered the limit. The upgrades now add 4 kg each on top of whatever the base is, giving 17 and 21 at the 13 kg base, so tuning the base can never turn an upgrade into a downgrade.