SAGE

The core idea

SAGE, short for Situation Aware Governance Engine, is a research stage project that asks a specific question: can useful capability come from the structure built around a model rather than from changing the model weights. The README frames it as a cognition kernel that wraps a frozen LLM in a persistent identity, a trust model, and a resource governance loop. The author is explicit that this is not a product but an experiment, and the project is licensed under AGPL-3.0. The bet is that intelligence emerges from orchestration across specialized components, not from scale alone. To keep expectations honest, the documentation separates quantitative findings from interpretive framings and lists exactly which components are real and which are mocked. That calibration is unusual and is presented as the strongest signal that the work is serious rather than promotional. The public repository holds the kernel architecture, the consciousness loop spec, the IRP plugin interface, and the identity layer, while some active research continues in private repositories that the README names but does not open. The author also states that multiple Claude instances act as active collaborators on the code and docs, which explains both the iteration speed and a known risk of coherent but untested framings. The author also states that multiple Claude instances act as active collaborators on the code and docs, which explains both the iteration speed and a known risk of coherent but untested framings slipping into the writing.

The consciousness loop and fleet

Every cycle, SAGE runs a twelve step consciousness loop. It senses observations, scores salience with SNARC across five dimensions of surprise, novelty, arousal, reward, and conflict, tracks an ATP budget as a metabolic state with modes WAKE, FOCUS, REST, DREAM, and CRISIS, computes a trust posture, selects attention targets, allocates budget across plugins, executes iterative refinement through the IRP framework, learns from convergence quality, updates memory, applies a policy gate, filters effects, and finally acts. The project runs as a federation across six machines hosting eleven SAGE instances on five model families, accumulating more than five thousand commits and fourteen hundred raising sessions through a developmental curriculum. A compiled Rust daemon named sage-rs runs the loop, federation, salience, and an HTTP dashboard in about twelve megabytes of memory, compared with roughly four hundred fifty megabytes for the earlier Python daemon. Each machine pairs a model such as Qwen or Gemma with hardware from a Jetson Orin Nano to an RTX 4090 laptop, and new instances start from a seed identity built from more than 117 accumulated sessions. The fleet polls peer health every thirty seconds, so the instances know when a sibling goes offline. The fleet polls peer health every thirty seconds, so the instances know when a sibling goes offline, and the seed identity built from more than 117 sessions gives each new instance a head start on behavior.

What is real versus mocked

The README's honesty section is the most distinctive part of the project. Components marked real include the continuous consciousness loop, local LLM inference through Ollama and Transformers, metabolic states, SNARC salience, the PolicyGate conscience checkpoint, seven tools in version 0.4.0a3, identity and relationship tracking, sleep consolidation, the Rust daemon, the federation mesh, and snapshot persistence at session boundaries. Components marked mocked include the sensors, which have no real I/O backends yet, and the physical effectors, where only the network effector works and the rest are stubs. A cross modal VAE and FlashAttention are listed as research stage, demonstrated but not in the live loop. The project also reports a public ARC-AGI-3 scorecard of 94.85 percent achieved with a frontier model plus a task specific harness, and it carefully notes the asterisks: that result is not proof of the edge kernel thesis and used affordances a strict competition would withhold. This separation keeps the measured work distinct from the aspirations. The author repeats the warning that conflating the two numbers is the overclaim the project most tries to avoid, which is why the real versus mocked table sits near the top of the README. The author repeats that conflating the public score with the edge kernel thesis is the overclaim the project most tries to avoid, which is why the real versus mocked table sits near the top of the README.

Editorial conclusion

SAGE is released under the AGPL-3.0 license and runs as a federation of eleven instances across six machines on five model families, per the project README.