OBSERVING
CAUM Live

See when AI agents repeat, retry, or stall.

CAUM watches the shape of agent work and shows when compute stops converting cleanly into progress. Start with one receipt, then move recurring workflows into CAUM Live for zero-semantic monitoring, workflow memory, and review-only evidence.

$ caum live ingest events.jsonl
# Reading structural events, not private content...
OK Profile: agent_trace | decision: observe_only
OK Evidence: T5 critical structural cycle | no semantic judgment
OK Chain: sha256:e3b0c44298fc1c14...
Conservative Validation & Regression Evidence
1,350
Replay traces checked
54,100
Structural events in corpus
71
Conservative hard alerts
0
Production replay errors

Includes public replay and controlled regression traces designed to validate structural detection paths. CAUM does not publish these as a customer waste rate.

Commercial Entry Point

Start with a CAUM Receipt.

The entry product is simple: send one real agent workflow, get structural findings, reviewable exposure, and a fix receipt your engineering team can test.

CAUM Receipt

For small teams and builders. Upload JSON or JSONL and receive an after-the-run structural receipt for loops, retries, stagnation, token exposure, and cost exposure.

CAUM Live

For teams with recurring agent runs. Stream neutral events from production or internal workflows and review structural health while agents operate.

Workflow Memory

For repeated workflows. CAUM groups recurring structural fingerprints, drift, and review boundaries so teams can see whether the same problem keeps returning.

Core Capabilities

Observe. Classify. Attest.

CAUM does not steer, block, or judge your agent's answer. It classifies the structural shape of execution and produces reviewable evidence.

Structural Regime Detection

Classifies agent execution into discrete structural regimes - Explorer, Grind, Stagnation, or Loop - through proprietary analysis. No semantic interpretation required.

Hash-Chain Attestation

Every analysis can produce committed evidence artifacts: canonical hashes, chain heads, Merkle-style roots, and report metadata that can be reviewed later.

Zero Semantic Access

CAUM does not need prompts, source code, customer messages, or business payloads. Strict mode uses opaque state IDs, event shape, counters, timestamps, and hashes.

Observation Model

Structure reveals what semantics cannot.

Agent work can stall silently: repeated calls, circular plans, retries, and long runs where compute stops converting cleanly into new structural movement. CAUM observes these patterns from event shape, counters, timestamps, hashes, and cycle evidence. No access to reasoning, prompts, code, or private data is required.

  • Review looping agents before repeated runs expose more compute
  • Audit historical agent traces for structural loop and cost exposure evidence
  • Maintain strict data privacy with opaque IDs, counters, timestamps, and structural metadata
  • Classify behavior into structural health tiers (T1-T5) for review, audit, and customer-owned policy hooks
Structural Regime Analysis Sample Output
Explorer
T1
Grind
T2
Stagnation
T3
Loop
T5
Processing Pipeline

Five steps. Fully deterministic.

Identical structural input produces the same structural evidence. No content truth judgment and no autonomous blocking.

1
Upload
Submit a JSONL trajectory file with discrete state transitions.
2
Canonicalization
Data is parsed and normalized into the canonical structural format.
3
Structural Analysis
Proprietary classification of trajectory dynamics and structural regime assignment.
4
Classification
Each trajectory window is assigned a Structural Regime and Severity Tier.
5
Report & Vault
PDF observation report and hash-chain Artifact Vault are generated.
Data Flow

Minimal input. Reviewable structural evidence.

What You Provide

A minimal trace with opaque state identifiers, event type, tool or phase labels, status, timestamps, and optional cost/token counters. No private user data. No API keys. No proprietary code.

  • Opaque state or action identifier
  • Sequence index, status, phase, and tool family
  • Optional latency, token, and cost counters
// sample_trajectory.jsonl {"step": 1, "state": "a3f8c2", "ts": "2026-01-15T08:41:02Z"} {"step": 2, "state": "b7d1e0", "ts": "2026-01-15T08:41:03Z"} {"step": 3, "state": "a3f8c2", "ts": "2026-01-15T08:41:05Z"}

What You Receive

A deterministic, independently verifiable report detailing the structural regime of the trajectory, with a hash-chain Artifact Vault.

  • Structural Regime (Explorer, Grind, Stagnation, Loop)
  • Severity Tier (T1-T5) per observation window
  • Structural Observation Report (PDF)
  • Evidence receipt - canonical hashes, chain heads, and report metadata
Deployment

Integrate on your terms.

Structural

Offline Analysis

Analyze historical trajectory datasets. Generate institutional structural observation reports for post-incident investigation and compliance audits.

  • Batch JSONL processing
  • Offline deterministic execution
  • Air-gapped compatible
Best for structural review, audit trails, and compliance workflows.
Live

Real-Time Monitoring

Integrate CAUM into running systems. Receive structural health tiers, loop evidence, and cost exposure while agents execute.

  • Low-latency REST API
  • Webhook stagnation alerts
  • Cloud-managed infrastructure
Best for continuous monitoring of autonomous agents.
Open CAUM Live
Enterprise

Embedded SDK

Deploy CAUM directly within your infrastructure. Full local control. Zero external data transfer.

  • Native Python & Go bindings
  • Zero-telemetry architecture
  • Bare-metal execution
Best for sensitive or high-volume deployments.
Output

The Structural Observation Report

A strictly formatted PDF designed for institutional audits, compliance reviews, and structural trace review. Every report includes canonical hash commitments so later tampering is detectable.

No sensitive context is ever exposed. Reports contain only structural regime classifications, severity assignments, and hash-chain attestations.

Structural Observation Report
Ref: CAUM-2026-0742
Date: 2026-02-15
Dataset Hashe3b0c44298fc1c14...96fb924
Structural HealthT5 - critical exact-cycle evidence
VerificationObservation-only - claim-audited receipt attached
Steps 0-14: Explorer regime. Healthy structural expansion.
Step 15: Structural recurrence detected. Regime transition.
Steps 16-40: Exact structural cycle repeated. Review before repeating similar compute spend.
Trust Architecture

Independently verifiable.

Every report can ship with canonical SHA-256 commitments, chain-head evidence, and a Claim Audit that records what the report does and does not assert.

Live Demonstration

Watch CAUM observe.

CAUM performs structural observation on a simulated execution, detects anomalies, and generates verifiable attestations - all without accessing semantic data.

Observer Idle
START END OBSERVING
Phase I
Detection
Phase II
Attestation
Phase III
Resolution
Structural Observer
CAUM initiates read-only structural observation. The system classifies execution behavior through state transition analysis without accessing underlying payload data.Privacy boundary: semantic content is not required.
Evidence Stream
Resources

Research & Formal Documentation

Run structural analysis on your dataset.

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

Frequently Asked Questions