What I'm actively building, what moved recently, and what is waiting for me to come back to it.
Ashes of the Unbound is the game I'm trying to make. SkyeEngine Code grew out of the way I coordinate development work. SkyeEngine and GLaDOS share the model-building pipeline, while HubN and KNET supply a steady stream of real infrastructure problems to solve.
These four projects share tooling, infrastructure and ideas, so I keep them together in the order they currently fit.
I'm building the survival factory RPG's foundations across gameplay plugins, dedicated persistence, gravity-relative movement, celestial lighting and repeatable evidence tooling.
Continue the planetary-systems work and close the unresolved first-connection SkyLight investigation without mistaking a useful diagnostic intervention for a fix.
I'm refining the desktop workflow I use to coordinate coding agents, preserve review state and recover interrupted work without letting chat become the source of truth.
Keep dogfooding it against real development, especially the boundaries between contract changes, holds, retries, recovered work and decisions which still need a human.
I'm moving reusable generation, evaluation, repair and training mechanics into SkyeEngine while GLaDOS keeps the identity, policy and model-specific decisions which should not be generalized away.
Freeze and arbitrate the measured three-judge disagreement set, finish the human and rework lanes, and certify the corpus before committing to another expensive 120B training run.
The peering workflow runs through a four-platform router lab while Registry, DNS, RPKI and Looking Glass state reconcile around it. KNET is the routed substrate underneath, including the queues, overlays and monitoring which reveal when the neat model meets an untidy link.
Keep the peering generator, router lab and deployed policy in step as the network changes, and keep making failures observable enough to diagnose without guesswork.
Finished investigations, decisions and write-ups from the last few weeks.
All three judgment censuses are complete; the first full comparison finds a 59.8% split-verdict rate and strongly directional disagreement.
Nearly sixty percent of complete three-way judgments split, with the direction and vocabulary of disagreement depending strongly on the judge.
Two simultaneous iBGP failures turned out to have two different underlay causes once I compared their directional packet counters.
A moderation category was measuring a repeated part of GLaDOS's figurative register rather than the content of the conversations.
These have not been disowned or quietly declared finished. They're simply not competing for this week's attention.