A control layer for the thoughts LLMs are poised to say

What it does J-Space Cognition Suite is an inference-time control layer that treats a language model’s accessible internal representations—concepts it is “poised to say”—as a deliberately managed workspace. It provides a compact operating protocol for selective attention, state preservation across task seams, failure detection, and verified output, all without touching model weights or adding external services. The suite is distributed as a cross-platform Skill that modules load selectively rather than concatenating everything into every prompt.

The interesting bit The project borrows from interpretability research on global workspace theory, then operationalizes it as a set of text-level control grammars: a Dense Track notation (✓, ?, ✗) for epistemic states, a broadcast hub for shared constraints, and strict rules about when I versus we may appear to bind state descriptions to concrete next actions. It is essentially an attempt to engineer metacognition through prompt protocol alone.

Key highlights

• Model-agnostic and zero-dependency: runs as pure text protocol with an optional local Python controller that uses only the standard library.
• Targets four specific failure modes: working-set overload, representation drift, uncontrolled retry, and premature completion.
• Dense Track register provides a compressed internal notation for verified, asserted, and refuted states that must be losslessly expandable to plain language.
• Selective loading by design: modules and references are injected on demand rather than bloating every request.
• Packaged as a Skill for cross-platform agent use, with an optional controller that preserves state between task seams.

Caveats

• Generic integration requires manual file injection; plain API models cannot resolve local paths themselves, so the caller has to retrieve and feed the right modules.
• The protocol assumes the host environment can expose and respect Skill interfaces; without that, much of the routing logic falls back to manual wiring.

Verdict Worth a look if you run long-horizon agent workflows and suspect your model is suffering from attention drift or premature completion. Skip it if you want a drop-in weight patch or a fully automated wrapper that needs no caller intervention.