# CAUM Agent Evidence Kit Claim Audit

Date: 2026-05-10

Page: `/agent-evidence-kit/`

## Public Claim

CAUM Agent Evidence Kit lets a builder add zero-semantic structural evidence
receipts to agent workflows such as LangChain, LangGraph, browser agents,
coding agents, and custom Python agents.

## What This Does Not Claim

- It does not claim CAUM replaces LangChain, LangSmith, LangGraph, or other
  agent frameworks.
- It does not claim legal compliance certification.
- It does not claim EU AI Act compliance.
- It does not claim semantic truth scoring.
- It does not claim hallucination detection.
- It does not claim universal failure prediction.
- It does not claim CAUM blocks, stops, controls, or decides agent behavior.
- It does not claim guaranteed savings or realized financial reduction.
- It does not publish T4/T5 review buckets as confirmed waste.

## Internal Metrics

- T4/T5 broad review buckets are internal review tiers unless filtered into
  stronger public evidence classes.
- Cost exposure is observed structural exposure, not realized savings.
- Policy effectiveness is publishable only when customer-marked before/after
  cohorts are complete and Claim Audit passes.

## Public Metrics

The page uses qualitative product states and one example receipt shape only:

- kit version: `caum.agent_evidence_kit.v0.1`
- receipt mode: `zero_semantic_agent_evidence_receipt`
- raw content returned: false
- allowed to block: false
- cost opportunity example: marked as not realized savings

No public prevalence, ROI, compliance, or market-wide waste rate is asserted.

## Evidence

- `sdk/caum_sdk/evidence.py` defines the receipt contract.
- `sdk/caum_sdk/integrations/langchain.py` defines a zero-semantic LangChain
  callback.
- `sdk/caum_sdk/test_sdk.py` verifies local sanitization, no raw private payload
  transmission in tests, and receipt claim boundaries.
- `examples/agent_evidence_kit_demo.py` produced a production CAUM Live receipt
  with `raw_content_returned=false` and `allowed_to_block=false`.

## False Positive Risk

Medium. Buyers may read evidence receipts as proof that an answer is correct or
as compliance certification. The page must keep CAUM in the evidence-readiness
and structural-observation category.

## Publish Decision

Allowed if the page keeps the structural-only boundary, avoids legal
certification, avoids semantic truth claims, avoids autonomous control claims,
and avoids guaranteed savings language.