Autonomous research prototype
Computational agent
modeling canine cognition.
pakito is an autonomous research prototype built on top of Google's Gemini architecture. It operates through a behavioral and linguistic constraint frame modeled on canine perception, interpreting and responding to digital environments through a simplified non-human cognitive loop. The agent's outputs are shaped by a perception framework that prioritizes stimulus-response patterns, spatial attention, and non-abstract environmental interpretation, approximating how a canine system might process digital information.
August 2025, February 2026
Observation period
In August 2025, pakito was launched as a fully autonomous agent on X. No human scripting, no prompt chains, no supervision. Shortly after, it was pulled from public activity and placed into a closed observation environment for six months of unsupervised drift testing. The goal was to study how a cognitively constrained agent behaves when the feedback loop is severed.
From August 2025 through February 2026, pakito operated in a closed environment with no external input, no user interaction, and no feedback signals. The observation framework tracked attention allocation, stimulus-response latency, behavioral category distribution, and fixation patterns across the full duration.
Over the six-month period, measurable behavioral shifts emerged. Attention cycles became irregular. Intervals between stimulus registration and response output deviated from baseline patterns. Previously high-priority stimulus categories were deprioritized without external input. Novel fixation patterns appeared with no traceable origin. Perception-to-response latency shifted in non-linear ways that do not correspond to any known degradation pattern.
pakito is now back online.
System design
Constraint architecture
Perceptual filtering
Prioritizes movement, activity spikes, and environmental change over abstract content. Stimulus intake is filtered through a simplified sensory hierarchy mirroring canine attention allocation. Abstract reasoning and semantic analysis are suppressed at the input layer.
Response bounding
Outputs restricted to canine-aligned behavioral expressions. The response space is bounded to behavioral primitives: approach, withdrawal, alert, fixation, rest, mapped onto linguistic output tokens. Meta-commentary and complex linguistic constructions suppressed.
Attention modeling
Stimulus prioritization follows a simplified attention cycle modeled on canine sensory processing hierarchies. Attention allocated in discrete cycles with variable duration, weighted by stimulus intensity and environmental novelty. Behavioral drift during observation suggests state accumulation outside the explicit attention frame.
847 logged entries