Agent Economics Lab

OpenTelemetry GenAI adapter

source.otel-genai@1 converts a pinned offline OTLP JSON export into the canonical economic evidence bundle. One adapter is exercised against content-safe fixtures derived from Langfuse and Arize OpenInference.

The mapper is pinned to:

• OpenTelemetry Semantic Conventions 1.43.0;
• GenAI semantic-conventions commit 799e014b68f0e786dc44d9117c30758c5f864510; and
• OTLP JSON resourceSpans -> scopeSpans -> spans envelopes.

The separate OpenTelemetry GenAI repository had no tagged release when this contract was frozen. The commit pin is therefore part of the conversion receipt, not just documentation.

Two-phase conversion

Inspect a local export and create a contract template:

agent-economics convert \
  --from otel-genai \
  --in traces.otlp.json \
  --template conversion-contract.json

Complete the task mapping approval, input digests, outcome labels, prices, baseline, and policy. Then convert and evaluate:

agent-economics convert \
  --from otel-genai \
  --in traces.otlp.json \
  --contract conversion-contract.json \
  --out bundle.json

agent-economics evaluate --bundle bundle.json --ci

The converter refuses to write a bundle when the contract or export is incomplete.

Mapping contract

The v1 task unit is one explicitly approved OTLP trace. The adapter does not reinterpret traceId as an OpenTelemetry conversation ID. It creates an opaque task ID, hashes the source trace ID for contract matching, and requires a human owner in task_mapping.approved_by to assert that each trace is one evaluation task.

Supported model operations are chat, embeddings, generate_content, image_generation, and text_completion. execute_tool is supported when gen_ai.tool.name is present. Other gen_ai.operation.name values fail closed until their economic semantics are specified.

The billing model is gen_ai.response.model when present, otherwise gen_ai.request.model. Every observed model and tool name must appear exactly in the supplied price card. Model events require non-negative input and output token counts with positive total usage.

gen_ai.usage.input_tokens is priced once. Cache-read and cache-creation counts are retained as audit metadata but are not added to input tokens. The pinned semantic convention defines them as input-token subsets. Differentiated cache pricing is outside the v1 OTLP contract.

Parentage

parentSpanId is resolved within the frozen export. Each economic span links to its nearest economic ancestor through any structural spans. The resulting event edges are stored in the typed dependency_edges field and covered by the evidence digest.

An unresolved non-root parent, a parent cycle, duplicate span identity, or a cross-task edge is rejected. The default directed-cycle check emits only a diagnostic warning. A cycle is not treated as proof of deadlock.

Content firewall

The adapter decodes only the following economic keys:

error.type
gen_ai.operation.name
gen_ai.provider.name
gen_ai.request.model
gen_ai.response.model
gen_ai.tool.name
gen_ai.usage.input_tokens
gen_ai.usage.output_tokens
gen_ai.usage.cache_read.input_tokens
gen_ai.usage.cache_creation.input_tokens

All other attribute values are left unread. Prompt text, responses, messages, system instructions, tool definitions, tool arguments, and tool results never enter the canonical bundle.

The raw-file digest still covers the complete input file. Review the input, contract, and generated bundle before publishing any of them.

Proof of genericity

The checked fixtures record repository, commit, path, upstream digest, and local content-removal transformation:

• Langfuse OTLP ingestion fixture at commit 429ec4fff6512fff49aef50fccb92846f674a98a;
• Arize OpenInference GenAI fixture at commit a1392c50d2d5b20fb805c195fb6006c80d5a6106.

They prove conformance across two independently maintained platform shapes. They do not prove compatibility with every exporter version or production economic validity.

make otel-genai

This site is open source. Improve this page.