The cull that reached too far: a gate that wasn’t supposed to kill anyone
The Long Watch promises that a distant animal is never erased outright: wander far enough from a herd and it folds into a population count instead of disappearing. In early July we added a small rule meant only to keep that promise honest at the edges. Tuning the same system for the largest worlds this August, we found real animals being deleted outright, with no record kept of them at all.
A promise, and the rule meant to guard it
Past the last watcher, a wild herd in The Long Watch stops being tracked animal by animal and folds into a single population count instead: a number that grows, holds, and drifts on its own until you come back to look. Nothing is lost in the folding. That is the whole point of the system: a place too far to watch is still alive, just counted rather than rendered.
Early in July we closed a hole at the edge of that idea: a location’s estimated population for a species should never be created in the first place if the location could never support even a single whole animal of that kind — ground too thin, too cold, too small to hold anyone. We called it the seeding gate. The design note that shipped with it was careful about what the rule was and wasn’t:
A seeding gate, not a new death path.
It promised that a never-viable population would never be created, and that it would refine the game’s persistence guarantee rather than reverse it. It shipped switched off, waiting for the day the deeper population layer was ready to lean on it.
What a performance pass turned up
That day arrived in August, when we started tuning how expensive the off-screen population layer is to run on the game’s largest world size. The layer had already burned us once, piling up duplicate herds until we taught it to rest; this time the surprise wasn’t speed. Watching the tuning run, we found animals disappearing from the world with nothing left behind: no population record, no count, nothing to say they had ever existed.
On one measured pass at the largest world size, just over half of all placed creatures (1,477 of 2,812) were vanishing this way, spread across 64 different species. We didn’t trust the number until we had a test that reproduced the loss on the unfixed game, then went looking for the fix.
Most of it wasn’t the rule at all
The obvious suspect was the seeding gate itself, so we split the losses apart. Of the 1,477 deletions, only 165 (about one in nine) actually belonged to the rule doing its job. Those 165 concentrated in seven species whose typical body mass sits at or above roughly 40.32 kilograms — reindeer, marsh deer, jaguar, Komodo dragon, green anaconda, capybara, and African rock python. On that large world, in the scenario we measured, nowhere within reach worked out to enough resource to support even one whole animal of any of them, so no population record for those species was ever created there. Every one of the seven lost every member that wandered far enough from camera, the gate working exactly as it was built to.
The remaining 1,312 losses had nothing to do with body mass or any location genuinely being too thin. They traced back to something else: a mismatch between two different maps of the same world. The check that decides whether a species can live somewhere at all runs on a coarse grid, cheap to evaluate. The system that actually places and moves individual animals off-screen runs on a much finer one. Most of these animals had wandered into fine-grid cells the coarse viability check had never evaluated at all, so when the moment came to fold them into a count, there was no record waiting to receive them. Finding nothing to add an animal to, the removal step removed the animal instead.
The fix, and what stayed the same
The fix leaves the viability rule alone. It only changes what happens the instant an off-screen animal would be removed: the removal step now checks first whether a population record for that species, at that location, actually exists. If one does, the animal folds into it as it always did. If none does, the animal stays — no longer deleted for want of somewhere to put it.
On the same measured scenario, the distant population count after a simulation update rose from 384 to 1,861, nearly five times higher, almost all of it animals that used to vanish without a trace. The tuning pass that surfaced the bug still delivered what it set out to do: the off-screen layer now runs about 3.73 times cheaper on the largest world size, with the population-viability threshold held fixed the whole time at roughly 40.32 kilograms of body mass, so none of the recovered animals came from loosening that rule either. The gate still keeps a never-viable population from ever existing. It just no longer takes anyone else down with it.
