1. Add the observer
The SDK sanitizes events locally and sends only structural labels, counters, and hashed state.
Live Evidence API
Every AI work session gets structural receipts automatically.
Connect CAUM to your agent workflow. Every observed step returns a zero-semantic evidence receipt and appears in the Live dashboard, without sending private prompts, files, customer messages, or business payloads.
Automatic receipts
Dashboard sessions
Observe-only
Zero private content
Latest API receipt
Observe-only
Session
T3
Boundary
clean
Receipt
step
Control
none
{
"mode": "zero_semantic_live_evidence_receipt",
"structural_health": { "tier": "T3", "public_class": "review_only" },
"evidence": { "algorithm": "sha256", "chain_head": "..." },
"review": { "agent_control_authorized": false },
"dashboard_path": "/live/sessions/"
}
The product flow
CAUM should feel like infrastructure: connect once, stream structural events, get receipts, and watch the session ledger update while the agent works.
2. Receive receipts
Each step returns an API-issued evidence receipt with tier, public class, hash chain, and claim boundaries.
3. Open the dashboard
Your CAUM Live API key opens customer-bound sessions, latest receipts, event counts, and structural health.
Three-line integration
The agent keeps its private work. CAUM receives structure only and returns the latest receipt immediately.
The API does not decide correctness, block the agent, or claim realized savings. It creates structural evidence for review.
Python SDK
Use the same API that feeds the customer dashboard.
from caum_sdk import observe
with observe("support-agent", api_key="caum_live_...", task_family="support_resolution") as caum:
caum.step("knowledge_search", private_result, tool_family="search")
print(caum.evidence_receipt["receipt_id"])
print(caum.dashboard_url)
What the dashboard shows
The dashboard is the customer-facing control room for CAUM Live: session ledger, workflow, event count, tier, live alert state, and the selected receipt. It uses the same `caum_live_...` key as the API.
1
Session ledgerRecent customer-bound runs, event counts, tier, public class, and chain head.
2
Selected receiptLatest structural receipt with hash-chain evidence and product boundary.
3
Sample runRun a controlled loop from the browser to verify the API key and dashboard path.
Dashboard session
Live
GET /v2/live/customer/sessions
Authorization: Bearer caum_live_...
{
"sessions": [
{
"workflow": "support-agent",
"event_count": 42,
"current_tier": "T3",
"receipt_id": "caum_live_receipt_..."
}
]
}
Current validation boundary
CAUM has controlled validation evidence on real model/tool-agent execution. The repeated-work class was intentionally prompted to create structural repetition, so this is engineering evidence for detection and privacy boundaries, not proof of customer prevalence.
150 controlled workflow runs
OpenAI Responses agents across support, software engineering, and research/RAG workflow families.
4,061 structural events
Zero-semantic event streams used to verify receipt generation, tiering, and repeated-structure detection.
Next proof target
Unprompted coding-agent traces from real recurring workflows, reviewed against CAUM's structural signals.
What teams do with a receipt
CAUM stays observe-only. When a workflow shows repeated structural pressure, the team decides whether to add a retry ceiling, checkpoint, fallback route, tool budget, or human review gate.
Retry ceiling
Stop repeating the same structural cycle after a team-owned limit.
Checkpoint
Pause the workflow when progress signatures stop changing cleanly.
Fallback route
Move the task to a safer model, tool path, or human review queue.
API first. Dashboard included.
CAUM Receipt is the one-trace entry point. CAUM Live Evidence API is the recurring layer that streams structural telemetry, receipts, task-family scope, and workflow memory into the dashboard.