A deterministic framework for classifying autonomous agent trajectories through structural dynamics — without accessing semantic content.
CAUM operates under a simple premise: the structure of execution is independent from the semantics of execution.
An autonomous agent generates a discrete sequence of state transitions over time. While the content of each state may contain semantic information, the structural pattern of transitions reveals behavioral characteristics that are invisible to semantic analysis.
CAUM observes this structure.
It does not require access to what the agent is reasoning about, what APIs it calls, what data it processes, or what outputs it generates. The shape of execution — its continuity, recurrence, and progression — carries sufficient signal for structural classification.
Each trajectory is transformed into a canonical structural representation. The input format consists exclusively of:
This design ensures that CAUM can operate across any agent framework, any LLM provider, and any execution environment. The canonical representation is framework-agnostic and vendor-neutral.
CAUM performs deterministic structural evaluation over trajectory data. Identical input always produces identical output — there is no stochastic component, no model inference, and no probabilistic estimation.
Each execution produces a Structural Observation Report containing regime classification, severity tier assignment, canonical cryptographic commitments, and a complete Artifact Vault for offline reproduction.
CAUM classifies every trajectory segment into one of four canonical structural regimes. These regimes describe the shape of agent execution — not its intent, quality, or correctness.
Each regime classification is accompanied by a severity tier ranging from T1 (lowest structural risk) to T5 (highest structural risk). Severity tiers reflect structural risk intensity and are derived from regime persistence and structural continuity characteristics.
Every structural evaluation is committed through a layered cryptographic attestation framework. This ensures that results are independently verifiable, tamper-evident, and reproducible.
Each structural evaluation input and output is committed through SHA-256 canonical hashing. The hash covers the normalized trajectory, regime classification, severity assignment, and all intermediate structural coordinates.
Individual evaluation hashes are aggregated into a Merkle tree. The root provides a single commitment that covers the entire analysis, enabling efficient partial verification.
The Merkle root is signed with an Ed25519 key. This signature proves that the analysis was produced by an authorized CAUM instance and has not been modified since generation.
CAUM was designed with a zero-semantic-access architecture. The system never processes, stores, or transmits any content that could reveal the agent's reasoning, data, or objectives.
Structural coordinates are sufficient for classification. This architecture enables deployment in regulated environments, air-gapped networks, and compliance-sensitive contexts where semantic data exposure is unacceptable.
CAUM occupies a deliberate position in the agent infrastructure stack. It is an observation layer — not a control layer.
It does not steer agents. It does not enforce outcomes. It does not inject policy. It does not modify execution. It does not make decisions on behalf of operators.
It observes structure and reports deterministically.
This design philosophy ensures that CAUM remains compatible with any agent architecture, any governance framework, and any operational policy. The structural signal is provided; the action taken in response is always determined by the operator.
Upload a trajectory. Receive verifiable evidence of your agent's structural behavior.