This is where the work lives.
Scorched Earth Labs is built around a simple research principle:
Question the primitive before optimizing the implementation.
Our work begins with assumptions that have become so familiar they are rarely examined. We remove one, construct an alternative, instrument the result, and test what happens.
What survives becomes part of the next experiment.
The work below includes published research, technical papers, protocols, experimental architectures, and active research questions.
Cognitive Persistence & Verifiable Cognition
ASTP — AI State Tree Protocol
What does an AI system need to preserve if the goal is not merely remembering information, but maintaining a durable cognitive record?
ASTP defines a protocol for persistent cognitive state organized around Episodes, structural divergence, relationships, provenance, and append-only integrity.
Ariadne
Ariadne is the research program from which ASTP emerged and through which we continue exploring Cognitive Persistence and Verifiable Cognition.
The central question is not whether an AI can remember a prior conversation.
It is whether a persistent system can maintain an inspectable record of how its cognitive state evolved—and whether that record can support independent verification of what state was available to the system at a particular point in time.
Explore Ariadne →Cognitive Dynamics
Trajectories
What if coherence is movement instead of similarity?
Trajectories explores cognition as a dynamical system.
Rather than treating coherence as a static similarity measurement, the system models movement over time—including velocity, acceleration, deviation, re-anchoring, and the maintained relationship between human and AI trajectories.
The work emerged from a practical measurement failure: an AI following a human-driven topic shift should not be penalized for incoherence simply because the conversation moved.
That failure led us somewhere much more interesting.
Explore Trajectories →Multi-Agent Collaboration
The Social Contract
Most multi-agent systems are good at coordinating work. That does not necessarily mean the agents are collaborating.
The Social Contract explores the difference.
Drawing from Conversation Analysis, organizational behavior, and research into human group dynamics, the framework gives agents explicit mechanisms for evaluating whether to contribute, defer, or remain silent—and for responding to the contributions of other participants.
Initial production observations are encouraging. Controlled empirical validation remains ongoing.
Human–AI Interaction
Poseidon
What if the interface were a function of intent, not the application?
Poseidon is our research into intent-driven, surface-independent interface composition.
UI elements are decomposed into graph-resident semantic objects rather than bound permanently to pages or applications. The system composes those elements according to user intent, cognitive load, personalization, and the capabilities of the rendering surface.
The surface changes.
The underlying intent does not.
Poseidon also preserves temporal interface state, making it possible to ask not merely what happened in the system, but:
What was the user seeing when it happened?
Explore Poseidon →Cognitive Execution
Routing Complex
What happens when routing becomes a cognitive operation rather than a static model-selection rule?
The Routing Complex explores the selection and coordination of cognitive resources across a heterogeneous agent system.
Scheduler
A persistent cognitive system operates across time.
The Scheduler coordinates work, dependencies, deferred actions, and temporal state across a live agent network.
Model Profiler
LLMs are useful cognitive resources.
They are not trusted authorities.
The Model Profiler characterizes model capabilities and limitations so that Ignis OS can reason about which models are appropriate for a task and evaluate their outputs accordingly.
Papers
We publish early and revise in the open.
Each paper carries its version and date. Findings describe the system as it was when the paper was written. Later versions supersede earlier ones, and both stay available. Implementation calibration values are withheld; the observations are not.
ASTP: A Framework for Verifiable Cognitive Persistence in Multi-Agent AI Systems via Episodic Memory and Merkle-Anchored Provenance
Episodes as the unit of cognitive coherence, a three-layer coherence model, and Merkle-anchored provenance that makes the evolution of recorded cognitive state inspectable. Includes the first benchmark series on cognitive residue across Episodes.
Since this paper. The protocol was renamed ASTP and is now patent pending. The Adaptive Merkle Tree gained a DAG extension for capturing processes. The specification has moved through several major versions, and the belief and knowledge layers now record into it.
The Social Contract: Applying Human Conversational Norms to Multi-Agent AI Evaluation Architecture
Six evaluation dimensions drawn from Conversation Analysis that let agents decide whether to contribute, pass, or defer. Two case studies from sealed collaborative Episodes, including a five-day compliance analysis with its full quantitative record.
Since this paper. Turn evaluations are now recorded as signals in every collaborative Episode, and a second contract, who may consult whom, emerged from the running system. The next paper relates the two.
Where It’s Going
There’s a direction we’re looking that reframes how an agent relates to the world — and to the models it uses to understand it.
It is early work, and there is more to test before there is more to say.
But it connects several questions we have been pursuing about perception, belief, evidence, and how persistent AI systems build and revise a worldview.
Follow the work →Follow the Work
This index changes as the research changes.
If you’re working on adjacent problems, disagree with one of our premises, or simply want to see what survives the experiments: