Econ Instrumentation SDK

LLM ROI tracking in production (value, cost, uncertainty) in <30 minutes.

Production-ready Python SDK for Regulus ECP telemetry:

• non-blocking execution/outcome telemetry transport
• hierarchical auto-instrumentation for LangGraph/LangChain/provider SDKs
• strict join-key continuity for business attribution

Install

pip install econ-instrumentation-sdk

With integrations:

pip install "econ-instrumentation-sdk[integrations]"

With OpenTelemetry bridge:

pip install "econ-instrumentation-sdk[otel]"

Quickstart (10 lines, real metric)

from econ_instrumentation import InstrumentationConfig, AutoInstrumentationConfig
from econ_instrumentation import configure, enable_auto_instrumentation, join_key_context
from econ_instrumentation import instrument_langchain_app, track_outcome, OutcomeEvent
configure(InstrumentationConfig(base_url="http://localhost:8000/v1", enabled=True))
enable_auto_instrumentation(AutoInstrumentationConfig(capture_mode="hierarchical"))
with join_key_context("CASE-123"):
    app = instrument_langchain_app(chain, "SupportRefundTriage", "model_v2")
    _ = app.invoke({"input": "refund?"})
    track_outcome(OutcomeEvent(execution_id="out_case_123", join_key="CASE-123", capability_id="SupportRefundTriage", outcome_type="custom", outcome_value="approved"))
# then run: python3 demo/live_poc_simulation.py --days 2 --requests-per-day 40 --strict
# inspect AER + margin in demo/output/report_*.json

Capture modes

• hierarchical (default): framework-first, provider fallback only
• all_layers: capture every layer and mark primary_for_rollup
• provider_only: OpenAI/Anthropic-only capture

CLI

regulus init --base-url http://localhost:8000/v1
regulus demo --repo-root . --days 2 --requests-per-day 40 --strict
regulus compute --repo-root . --last 7d

Recommended env flags

export ECP_AUTO_INSTRUMENT=true
export ECP_CAPTURE_MODE=hierarchical
export ECP_INSTRUMENT_LANGCHAIN=true
export ECP_INSTRUMENT_OPENAI=true
export ECP_INSTRUMENT_ANTHROPIC=true

Compatibility (tested ranges)

• openai >=1.40,<2.0
• anthropic >=0.34,<1.0
• langchain >=0.2,<0.4
• langgraph >=0.2,<0.4

Notes

• Prompts/responses are not captured by default (capture_content=False).
• Telemetry delivery never blocks inference path (buffer + background retries).