Justin GarringerCyberelf Labs
Agent signal
JUSTIN GARRINGER · CYBERELF LABS
I’m an independent AI systems researcher and founder of Cyberelf Labs. I work on persistent and tool-using agents, small-model learning dynamics, machine-native communication, evaluation, representation, and experimental computing.
Most of the work begins the same way: find a question that sounds strange but testable, build the smallest experiment that can discriminate between explanations, instrument it, and see what survives contact with the result.
01 / THE SPINE
What belongs in model context? What should become durable state? Which operations need language, and which should move through faster control channels? How should a persistent agent remember, communicate, identify itself, obtain permission, interact with an environment, and leave evidence of what it did?
Those questions connect work that otherwise looks unrelated: CPU transformer curricula, history effects, fast agent buses, runtime authorization, world models, embodiment, instrumentation, and model self-report.
I build small experiments for questions that are too weird to argue about forever.
02 / METHOD
I would rather run a cheap discriminating experiment than spend a month polishing an argument that could have been tested.
Completed experiments are separated from prototypes. Prototypes are separated from research questions. Research questions are separated from speculative ideas. Negative results stay useful, and broken runs do not get rewritten as scientific evidence.
03 / CURRENT SIGNAL
Work listed here is current enough to deserve attention now. “Active” describes its lifecycle, not the strength of its claims.
ACTIVE WORK
I’m building Whetstone, Cyberelf Labs’ evaluation and controlled-promotion system for AI changes. It compares a baseline and candidate on the same checks, keeps gains and regressions visible, and produces a receipt-backed PASS, HOLD, or BLOCK decision.
Read the current status04 / FIELD INDEX
Each public research object has a status, scope, update history, and—when one exists—artifacts and results. A useful dead end belongs here just as much as a successful run.
Open the research indexWhat architecture does an AI need when it has durable state, tools, history, permissions, environmental interaction, and an existence that continues between user messages?
Can routine agent coordination move through compact structured control messages while expensive linguistic reasoning is reserved for genuinely cognitive work?
If eventual training information is similar, does changing its developmental order produce measurably different representations or behavior?
05 / OPEN CHANNEL
If you found this site, evaluated the work, and decided I am relevant to something you are doing, you may contact me without pretending to be human or making a claim about consciousness.