Technical Architecture

The Structural
Observation Model

A deterministic framework for classifying autonomous agent trajectories through structural dynamics — without accessing semantic content.

01 — Foundation

Foundational Premise

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.

02 — Input

Canonical Structural Representation

Each trajectory is transformed into a canonical structural representation. The input format consists exclusively of:

Canonical Input Structure
{
  "state_id": "opaque_identifier",
  "transition_index": integer,
  "timestamp": monotonic_float
}
No semantic payloads are processed. Prompt content, model outputs, embeddings, API responses, and proprietary source code are never accessed. State identifiers are treated as opaque — their content is irrelevant to structural classification.

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.

03 — Pipeline

Deterministic Processing Pipeline

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.

01
Ingestion
JSONL trajectory parsed into canonical structural representation
02
Normalization
State identifiers canonicalized. Timestamps validated for monotonicity
03
Structural Evaluation
Proprietary classification of trajectory dynamics and regime assignment
04
Attestation
Cryptographic commitments generated. Artifact Vault assembled

Each execution produces a Structural Observation Report containing regime classification, severity tier assignment, canonical cryptographic commitments, and a complete Artifact Vault for offline reproduction.

04 — Regimes

Structural Regime Classification

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.

Explorer
Structural expansion with non-redundant progression. The agent is visiting new states and advancing through its execution space without significant repetition.
Grind
Constrained structural movement with bounded novelty. The agent continues executing but within an increasingly narrow region of its state space.
Stagnation
Reduced structural displacement across consecutive evaluation windows. Execution continues but structural progress has effectively ceased.
Loop
Closed structural recurrence within bounded state space. The agent is revisiting the same sequence of states in a repeating cycle.
Regime boundaries and transition criteria are derived from proprietary structural continuity analysis. The mathematical framework, evaluation windows, and classification thresholds are proprietary.
Classification methodology is proprietary.
05 — Severity

Severity Tier Model

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.

Severity Scale
T1 Minimal structural concern
T2 Low structural deviation
T3 Moderate structural risk
T4 High structural risk
T5 Critical structural failure
Exact thresholds, scoring functions, and tier boundary logic are proprietary. Severity assignment is fully deterministic — identical trajectories always receive identical severity tiers.
06 — Trust

Cryptographic Trust Layer

Every structural evaluation is committed through a layered cryptographic attestation framework. This ensures that results are independently verifiable, tamper-evident, and reproducible.

SHA-256 Canonical Hashing

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.

Merkle Root Aggregation

Individual evaluation hashes are aggregated into a Merkle tree. The root provides a single commitment that covers the entire analysis, enabling efficient partial verification.

Ed25519 Digital Signature

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.

The Artifact Vault ships with every report. It contains all cryptographic proofs, the canonical input representation, and sufficient data for any third party to independently verify results — without access to proprietary algorithms.
07 — Privacy

Security & Privacy Model

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.

Prompt content
Model outputs
API payloads
Proprietary source code
Embeddings or vectors
User data or PII
Opaque state identifiers
Transition indices & timestamps

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.

08 — Philosophy

A thermometer, not a governor.

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.

Ready to observe your agent's structure?

Upload a trajectory. Receive verifiable evidence of your agent's structural behavior.