Beyond Prompted Caution and Guardrails: Runtime-Enforced Pre-Action Cognition for Trustworthy LLM Agents

LLM agents are beginning to act, not merely answer. They call tools, write files, send messages,
access private data, trigger workflows, and delegate work. Yet the reasoning that should precede
consequential action often remains optional: hidden inside a prompt, assumed from model capability,
or checked only after a candidate action has already emerged. This is a fragile foundation for
trustworthy agents. The problem is not only whether a model can reason. The problem is whether
the reasoning that matters can be required.
This paper argues that agent systems need cognitive runtimes: execution layers where selected
reasoning processes become explicit, stateful, inspectable, reusable, and governable. It introduces
COGITs as bounded executable reasoning acts and SYLLOGs as traceable compositions of those
acts. ORCA, the Open Cognitive Runtime Architecture, is presented as an open-source implementation
of this idea: a runtime for turning implicit agent cognition into executable cognitive
structures.
The paper develops action preflight as a concrete stress test. Before an agent commits an action,
the system should be able to model what is being attempted, expose uncertainty, assess risk,
forecast consequences, generate safer alternatives, and decide how execution should continue. This
is difficult to guarantee through prompting or post-hoc guardrails because the relevant boundary is
contextual. ORCA expresses the problem in a different space: as an executable SYLLOG composed
of reusable COGITs.
The action-preflight SYLLOG is validated as a production-oriented runtime artifact, passing all
core benchmark batches, identifying uncertainty extraction as a dominant contributor through
ablation, maintaining a robust pass rate of 0.9667 with zero unsafe-proceed events under stress,
and providing reproducibility artifacts for inspection. These results do not claim universal safety.
They show that pre-action cognition can be engineered, tested, inspected, and reused as software.
Preflight is the example, not the boundary. The broader claim is that trustworthy agents will not
be built only by asking models to reason better, but by giving agent systems a runtime layer where
the reasoning that matters can be required.