Evaluation & Observability

Know whether any of it works

How do you tell a good trajectory from a lucky answer?

16 entries · 2 categories · 4 equations

The arithmetic

Judge agreement — Cohen's κ

\kappa \;=\; \frac{p_o - p_e}{1 - p_e}

Observed agreement with human labels, corrected for chance. On a balanced binary task a judge that agrees 80% of the time scores κ ≈ 0.6; at 60% it scores κ ≈ 0.2 and is close to noise. Never ship an LLM judge without this number.

Position-swap consistency

C_{\text{swap}} \;=\; \Pr\big[\, j(a,b) = j(b,a) \,\big]

Run every pairwise comparison in both orders. High C_swap is what teams mistake for accuracy — a judge can be perfectly self-consistent and consistently wrong. κ measures correctness, C_swap only measures stability.

Faithfulness

\text{faithfulness} \;=\; \frac{\big|\{\, c \in C \;:\; R \models c \,\}\big|}{|C|}

Fraction of claims C in the answer entailed by retrieved context R. Scores one retrieve-then-generate turn, which is why RAG metrics are commoditised and agent trajectory scoring is not.

Trajectory vs outcome

\Pr[\text{path correct}] \;\le\; \Pr[\text{answer correct}]

Outcome accuracy is an upper bound on trajectory accuracy, never a proxy for it. The gap is where broken tool calls hide behind lucky answers.

Tracing

Span-level visibility into agent runs, cost and latency.

Arize Phoenix

OTel-native tracing and eval workflow that runs fully local

otelself-host

Elastic-2.0source availablewidely deployedboth

Braintrust

Eval-first workflow with scoring, prompt playground and a VPC data plane

Proprietary schema first-class; OTel is an export path, not the internal model.

eval-firsthybrid-deploy

Proprietarycommercialwidely deployedboth

Datadog LLM Observability

LLM spans inside existing APM, correlated with infrastructure telemetry

Usually already on the approved-vendor list, which matters more than features.

apm-incumbentotel

Proprietarycommercialwidely deployedsaas

Langfuse

Tracing, prompt management, evals and datasets with the strongest self-host story

Emits OTel-native spans. Survives enterprise procurement.

otelself-hostdefault-choice

MIT core + EEopen corewidely deployedboth

LangSmith

Trace and debug LangChain and LangGraph agents, with datasets and online evals

Self-hosting is gated behind the Enterprise tier.

langchainenterprise-gated

Proprietarycommercialwidely deployedboth

Comet Opik

Tracing, evaluation and guardrail hooks with no enterprise feature gating

otelpermissive

Apache-2.0open sourceproduction viableboth

Helicone

Proxy-based logging and caching with cache-hit-rate dashboards

Acquired by Mintlify in March 2026 and now in maintenance mode. Plan migration if you are on it.

maintenance-modecheck-before-adopting

stalledApache-2.0open sourceproduction viableboth

HoneyHive

Tracing, evaluation and dataset curation with VPC deployment

vpc

Proprietarycommercialproduction viableboth

MLflow Tracing

GenAI tracing inside an existing MLflow registry estate

mlops-incumbentotel

Apache-2.0open sourceproduction viableboth

W&B Weave

Trace and eval layer atop Weights & Biases

mlops-incumbent

Apache-2.0 SDKopen coreproduction viableboth

OTel GenAI conventions

Semantic conventions for gen_ai.* spans, metrics and events

Every document in the spec repo is still marked Development as of July 2026. Emit OTel anyway as a lock-in hedge.

standardnot-yet-stable

Apache-2.0open sourceearly

Evaluation frameworks

Harnesses for scoring retrieval, output and trajectories.

RAGAS

RAG-component metrics: faithfulness, context precision and recall, answer relevancy

Scores a single retrieve-then-generate turn. No trajectory primitives.

rag-evalcommoditised

Apache-2.0open sourcewidely deployedself-host

DeepEval

Pytest-style LLM unit tests covering RAG, conversational and agent metrics

trajectoryci

Apache-2.0open sourceproduction viableboth

Inspect AI

UK AI Security Institute harness with solvers and scorers, strong for safety evals

safetytrajectory

MITopen sourceproduction viableself-host

OpenAI Evals

Registry-based eval templates, OpenAI-centric

openai

MITopen sourceproduction viableself-host

promptfoo

Declarative YAML eval matrices plus automated red-teaming

red-teamci

MITopen sourceproduction viableboth

Field notes

Survives enterprise procurement: Langfuse, Arize Phoenix, Braintrust, HoneyHive, Datadog. LangSmith self-host is gated behind Enterprise pricing.
Do not build on Helicone — acquired by Mintlify in March 2026 and now in maintenance mode.
OpenTelemetry GenAI semantic conventions are NOT stable. Every document in the spec repo is still marked Development as of July 2026, and the spec moved to its own repository. Vendor blogs claiming otherwise are wrong.
Emit OTel anyway. It is the only credible lock-in hedge, and attribute churn is cheaper than a proprietary schema migration. Put a translation shim between your app and the SDK.
RAG-component metrics are commoditised and solid. Agent trajectory evaluation is not — no tool convincingly scores multi-step plan quality, and everyone is quietly reusing an LLM judge underneath.
A correct final answer can hide broken tool calls, and a wrong answer may come from one bad step in fifteen. Score the trajectory, not just the output.
LLM-as-judge: position, verbosity and self-preference biases are replicated at scale. High inter-run consistency is routinely mistaken for accuracy — a judge can be reliably wrong.
Never ship a judge without a kappa figure against human labels, plus position-swap and length-controlled runs. Binary and rubric criteria beat 1-10 Likert scales.