genai-interlingua · measured, not asserted

What translation costs

How a span is recognised

This is the only part that is an algorithm. Everything else on this page is a census — what these fixtures were observed to carry, counted. This is computed per span: each dialect counts the attributes only it would write, the highest total wins, and the number stamped on the span as interlingua.dialect.confidence is winner − runner-up. An integer margin, not a probability, because there is no denominator to make one out of.

What each dialect could not carry, and why

None of these is a spelling problem. Renaming prompt_tokens to gen_ai.usage.input_tokens is the part that looks like the work and is the part a lookup table can do. Every band above is the other kind: no field means the conventions have no concept for it, not merely no name for it; unstructured is a blob the schema cannot type; flattened is structure that existed in the source and has no canonical form; and ambiguous is one attribute that could have meant two different fields, recorded rather than guessed at. Matching names is not matching meaning, and this is where the difference is measured.

Lost to the vocabulary, not deleted. Under the default originals: keep every attribute above is still on the span, under the name its library gave it. What it lacks is a gen_ai.* name, so a query written in the conventions’ vocabulary will not find it and one written in the library’s will. originals: dedupe removes only exact copies a rename already made redundant. Only originals: prune takes them off the span, and that is a choice somebody has to make.

LiteLLM's 53 is not noise. It is an entire cost model — gen_ai.cost.input_cost, output_cost, margin_percent, discount_amount — plus every user_api_key_* field it records about who spent what. The conventions have no home for any of it, which is a finding about the conventions rather than about LiteLLM.

Field coverage, and the cost of pinning

A filled cell means that dialect's fixtures carried the field. A blank is silence, not a denial: the emitter may have had nothing to put there, or the fixture may not exercise it. Read a mark as evidence.

The cost of pinning a target

Both are defensible. Neither is correct. What is not defensible is making the choice silently — which is what a normalizer with a hardcoded attribute table does, and why the target here is a flag rather than a constant.

Every number on this page is generated from fixtures by TestConformanceTable, and CI fails when the table and the fixtures disagree. Rows are named by the key the field takes at genai-main; gen_ai.usage.cache_write.input_tokens is spelled cache_creation at v1.41.0. Provenance, per dialect: . A captured fixture is a claim about what that library really emits, recorded by running it; a hand-built one is a claim this repository makes about the library, checked only against itself.

The normalizer is genai-interlingua — an OpenTelemetry Collector processor and CLI, Apache-2.0. Every span it emits carries interlingua.lossy: the keys that span is not a faithful carrier of, on the span itself, so the loss above is queryable in production rather than only on this page.

There is already a first-party component for this. processor/genainormalizer landed in opentelemetry-collector-contrib in May 2026 — donated that February — and ships in the otelcol-contrib binary you are probably already running. If OpenInference and OpenLLMetry are the libraries you have, reach for it first: it needs no custom build and this does. What is measured here and not there is the cost. It has built-in tables for those two dialects, which is 31 of the 105 attributes above; it normalizes them and records nothing about what it could not carry. The remaining 74 are on dialects it does not claim — 53 of them LiteLLM’s cost model — so they are a gap in the conventions rather than in that component. The differences are being taken upstream rather than maintained in competition; the comparison is in the README, dated, because that component moves and a copy of it here would quietly go stale.

Every figure here is an aggregate over fixtures. The question underneath it — why one particular value means what it claims, which source attribute produced it, and what the translator declined to guess — is what the normalization inspector answers, one attribute at a time. Matching names is not the same as matching meaning, and ambiguous above is the band where that difference is measured rather than argued.