CAEL provenance bridge

Optional source-linked reasoning tools for LangGraph and CrewAI. This adapter returns the actual local CAEL response, exact source quotes and IDs, input and engine hashes, CAEL's parsed interpretation, and a finite Horn-rule proof. A separate checker validates the derivation and replay against trusted original input.

The adapter glue is MIT-licensed. The external CAEL 0.4.0 engine and model retain their separate non-commercial personal/research licence. They are not included in this package, and its MIT licence does not change their terms.

Installation

python -m pip install 'cael-provenance-bridge[langgraph,crewai]==0.1.0'

Use Python 3.10–3.13 when installing both frameworks. Provide a trusted CAEL checkout containing the exact engine/model files pinned by the adapter:

git clone https://github.com/Vovala14/CAEL.git cael-runtime
git -C cael-runtime checkout 49359eae9aa197d5b92631c16efb74085180a27c
export CAEL040_ROOT="$PWD/cael-runtime"
export OTEL_SDK_DISABLED=true
export CREWAI_TELEMETRY_ENABLED=false
export CREWAI_DISABLE_VERSION_CHECK=true
export LANGSMITH_TRACING=false

No external model/provider, API key, embedding service or network fetch is needed to compute an answer. Installing the optional frameworks installs their own dependencies; this is separate from invoking a neural model.

Framework entry points

import os
from cael_provenance import CAELRuntime, verify
from cael_provenance.crewai import CAELProofTool, provenance_flow
from cael_provenance.langgraph import graph_builder, proof_tool

runtime = CAELRuntime(os.environ["CAEL040_ROOT"])

• LangGraph: a StateGraph builder, checkpoint/resume support and a standard tool retaining the complete certificate in ToolMessage.artifact.
• CrewAI: a typed BaseTool returning the complete Pydantic envelope and a deterministic Flow, without Agent/LLM planning.
• verify(certificate, original_request, runtime) binds the certificate to the original trusted source text and independently checks the finite proof.

See the complete usage examples and API limits.

Scope and measured results

The installed wheel passes 79 checks with zero skips on Windows; the same bridge source also passes all 79 on Linux with the real frameworks. The 17 distinct requests include 12 new ordinary cases, three unsupported boundaries and two retained historical parser misses. Intended answer agreement is 12/14 outside the boundaries, including unknown/conflicting answers. These are integration and integrity checks, not a benchmark or a claim of general reasoning ability. Full outputs and failures are retained in the validation report.

Verification is relative to CAEL's parsed premises. It does not establish external truth, authenticate source authorship, sign documents or prove correct interpretation of arbitrary English. Origins are caller-declared; hashes are not signatures. Keep the original request separately as the trust anchor. Unknown, false, conflicting, unsupported and unverified results are distinct. Downstream prose rewritten by another model is not automatically covered by the certificate. The two known parser misses remain unfixed.