Netsky Lab — a lab of one. Independent R&D on the substrate models actually run on: bounded context, run continuity, model surgery, inspectable agent memory. Public on GitHub — code first, claims after.
Training runs, fine-tunes, and local inference fail silently unless runs, configs, failures, and evaluation traces are first-class engineering objects. Most teams treat them as logs and wonder why nothing reproduces.
A capable agent should remember what failed, what was verified, what policy changed, and what the next run must read before editing code. Without that, every session is a stranger to the last.
Token budgets are not hoarded. Context gets ranked, admitted, broadcast, and evicted through capacity-limited workspaces that can be inspected — gating and broadcast are engineering objects, not folklore.
Claims should point to sources, quotes, traces, benchmarks, byte-stable manifests, or code. A beautiful answer without provenance is still a liability.
The public org is meant to be inspected. If the work is good, the code should carry more weight than the landing page.
The work lives at the seams — where model behavior, memory, and evaluation refuse to compose cleanly.
Specific substrate over stacks. The public repos are the work, not the pitch.
Questions and proposals — channels below.