CAUM Agent Evidence Kit v0.1

Add evidence receipts to agent workflows.

Install one lightweight observer beside LangChain, LangGraph, browser agents, coding agents, or custom workflows. CAUM returns a structural receipt with health, review readiness, chain evidence, token exposure, and cost exposure without reading private content.

Python SDK LangChain callback Zero-content Observe-only No savings guarantee
Evidence Receipt
Generated
Version
v0.1
Health
T1-T5
Content
not sent
Chain
sha256
receipt: caum.agent_evidence_kit.v0.1 | mode: zero_semantic_agent_evidence_receipt | allowed_to_block: false

CAUM is not your agent framework. It is the evidence layer beside it.

Teams can keep LangChain, LangGraph, OpenAI Agents, Claude Code, browser-use, CrewAI, n8n, or their own stack. CAUM gives the buyer a receipt that says how the run behaved structurally, not whether the output was semantically true.

Install where work happens

Wrap one workflow, tool loop, agent invocation, or callback stream.

Keep content private

Prompts, completions, source code, files, and customer payloads stay outside CAUM.

Produce a receipt

Return structural health, chain hash, review status, cost exposure, and claim boundaries.

Minimal integration

The default observer works with any Python agent. It hashes private values locally and sends only structural telemetry to CAUM Live.

Claim lock: CAUM observes and records. It does not decide, interrupt, certify legal compliance, detect hallucinations, or guarantee financial savings.

Python

Wrap one agent run and print the receipt.

from caum_sdk import observe

with observe("support-agent", agent_id="agent-7") as caum:
    result = agent.invoke({"input": "private request"})
    caum.step("agent.invoke", result, cost_usd=0.012)

receipt = caum.receipt.to_dict()
print(receipt["structural_health"])

LangChain / LangGraph

Use the callback when the workflow already emits agent events.

from caum_sdk.integrations.langchain import (
    CaumLangChainCallbackHandler,
)

caum = CaumLangChainCallbackHandler(
    workflow="refund-agent",
    agent_id="refund-agent-v1",
)

agent.invoke(
    {"input": "private request"},
    config={"callbacks": [caum]},
)

receipt = caum.receipt().to_dict()

Receipt output

The output is designed for audits, pilots, and customer evidence packs.

{
  "version": "caum.agent_evidence_kit.v0.1",
  "mode": "zero_semantic_agent_evidence_receipt",
  "structural_health": {
    "tier": "T5",
    "public_class": "hard_alert",
    "publishable_as_waste_rate": false
  },
  "cost_and_tokens": {
    "reviewable_exposure_opportunity_usd": 0.01139514,
    "opportunity_is_realized_savings_claim": false
  },
  "product_boundary": {
    "raw_content_returned": false,
    "allowed_to_block": false
  }
}

First integration, then pilot revenue.

The kit gives CAUM a direct path into real builders: integrate in one workflow, generate receipts, then move recurring workflows into CAUM Live when structural exposure repeats.