Ignis OS
What if the model wasn’t the system?
Ignis OS is the experimental cognitive substrate at the center of Scorched Earth Labs.
It coordinates specialized agents, models, memory, tools, schedules, interfaces, and persistent cognitive state without treating any individual LLM as the source of truth.
Models are resources.
Ignis OS is the system around them.
Built to Run the Lab
Ignis OS is not a demonstration environment.
We use it.
Research unfolds inside it. Agents collaborate through it. Projects persist across sessions. The same systems we study are involved in the day-to-day work of the lab.
We don’t demo it. We live in it.
Models as Untrusted Sources
An LLM can be extraordinarily capable and still be wrong.
Ignis OS therefore treats model outputs as observations to be evaluated rather than truths to be accepted.
- The Model Profiler characterizes capabilities.
- The Routing Complex selects cognitive resources.
- Persistent state provides continuity.
- Governance mechanisms constrain what can become durable.
The goal is not to find one model we can trust.
It is to build a system capable of deciding what deserves trust.
Beliefs, Desires, Intentions
What does an agent hold to be true—and when did it decide that?
Every Ignis OS agent runs a Beliefs–Desires–Intentions cycle. BDI is an old idea from agent theory. We use it because it makes cognitive state explicit enough to record.
- Beliefs are what the agent currently holds, each with a confidence and a source.
- Desires are the goals a request gives rise to, ranked by priority.
- Intentions are the goal the agent has committed to, with a lifecycle: forming, executing, completed, interrupted.
None of this lives inside the model. It lives in the system around the model, where it can be written, read, and governed.
Beliefs are tiered. A constitutional layer is shared by every agent and is not writable at runtime; instructions, tasks, and user requests operate on the tiers beneath it. Structural beliefs describe identity, role, and domain. Operational beliefs change as the work does.
An agent can propose a durable belief during its work. The proposal passes eight gates before it is recorded, ending with a comparison against the negation of every constitutional belief. A near miss goes to review. A contradiction is refused.
A model can assert anything.
The system decides what becomes durable.
Superseded beliefs are archived and linked to their replacements. Displaced intentions are recorded with a hash of the belief state at that moment. All of it lands in the Episode.
That is what the BDI panel in Thermyt shows: not what the model said, but the recorded beliefs, goals, and commitments behind it, in order.
Open question: does a belief architecture designed for symbolic agents survive contact with a probabilistic reasoner? That is what we are instrumenting.
A System of Research Systems
Ignis OS is where many of the lab’s research programs meet:
ASTP
Provides the protocol for persistent cognitive state.
Trajectories
Models cognitive movement and coherence.
TSML
Explores tiered semantic memory.
Social Contracts
Governs collaborative participation.
Poseidon
Composes intent-driven interfaces.
Routing Complex, Scheduler, Model Profiler
Coordinate execution.
Each is independently researchable.
Together, they form the environment in which we test what a persistent artificial cognitive system might become.