← Benchmarks 🕐 10 min read
Benchmarks

Enterprise AI Evals Field Guide — 2026

> **Ingestion status:** The Awesome Evals ingestion materializes 318 cited rows, 318 golden-quality exemplar rows, 269 raw mirrors, 143 mirrored BenchFlow notes, and 83 Chandra-extracted PDF rows, inc

Source files: sources/21-benchmarks/awesome-evals-primary-sources-2026.md · sources/21-benchmarks/awesome-evals-primary-sources-2026.jsonl · sources/21-benchmarks/awesome-evals-golden-exemplars-2026.jsonl · sources/21-benchmarks/eval-framework-native-schema-primary-sources-2026.md · sources/21-benchmarks/frontier-lab-eval-sdk-primary-sources-2026.md · sources/21-benchmarks/eval-native-schema-boundaries-2026.json · sources/21-benchmarks/parlance-auto-evals-production-trace-error-analysis-raw.md · sources/21-benchmarks/model-routing-meaningfulness-hse-robustness-raw.md · statebench/schemas/eval_result_v1.schema.json · statebench/schemas/failure_taxonomy_v1.schema.json.

Production companion pages: Enterprise Evaluation for Knowledge Bases and Coding Agents · Eval Framework Schema Crosswalk · Frontier Lab Evaluation Research and SDK Primitive Crosswalk · Decentralized GitLab Eval Federation

Production machine-readable artifacts: Awesome Evals source ledger · Awesome Evals source summary · Awesome Evals golden exemplars · PDF/Chandra queue

Ingestion status: The Awesome Evals ingestion materializes 318 cited rows, 318 golden-quality exemplar rows, 269 raw mirrors, 143 mirrored BenchFlow notes, and 83 Chandra-extracted PDF rows, including 2 chunked extractions. The 46 media rows were attempted through the transcript path; public captions were unavailable, and the local ASR pass is queued because audio download/transcription requires the podcast/video ASR pipeline.


Answer

An enterprise eval is not a score. It is an evidence contract that says: for this task, this model, this prompt, this retrieval index, this tool catalog, this runtime, and this judge/check, here is what happened, where it failed, and what evidence is retained.

The durable architecture is:

  1. Run evals repo-locally where application code, fixtures, prompts, and tool schemas live.
  2. Normalize only the result row into a firm-wide schema.
  3. Keep native framework exports as evidence, because DeepEval, RAGAS, LangSmith, Phoenix, Hugging Face, provider SDKs, and custom pytest suites do not share one stable native schema.
  4. Promote only reviewed failures into knowledge-base updates, SFT rows, preference pairs, or policy changes.

The Awesome Evals corpus strengthens the prior finding: evals are becoming a product capability, not a one-off QA activity. The field is moving from static benchmark scores toward verifier design, RL environments, benchmark integrity checks, agent trajectory scoring, security/adversarial harnesses, and observability-backed eval loops.

Where Evals Live In The Agentic Chat Cycle

Cycle point Eval artifact Main failure domains Framework or lab primitives
User/task intake task card, persona, policy boundary, expected artifact ambiguous intent, wrong persona, missing authority, unsafe request OpenAI eval item schema, Anthropic task design guidance, Google ADK eval sets, custom golden work packets
Prompt/context assembly prompt template, memory state, retrieved context, tool catalog stale context, ACL leak, prompt drift, missing source, wrong tool exposure RAGAS contexts, DeepEval context fields, LangSmith/Phoenix traces, Bedrock KB traces
Model generation model call, sampling config, response hallucination, refusal error, format drift, cost/latency blowout OpenAI graders, model eval suites, Hugging Face LightEval/Evaluate, lm-eval-harness
Retrieval and grounding retrieved chunks, citations, source bundle wrong source, unsupported claim, stale source, citation span mismatch RAGAS faithfulness/context metrics, DeepEval RAG metrics, Phoenix retrieval spans, FACTS/LongFact-style grounding
Tool use and action tool schema, tool call, observation, approval wrong tool, invalid args, side effect, failed recovery, permission bypass provider traces, Claude Code/MCP hooks, AutoGen traces, Qwen-Agent function calls, custom deterministic checks
Model/tool routing routed actor, candidate pool, route rationale, perturbation set redundant model pool, unstable assignment, cost-only routing, accuracy/robustness split AI gateways, RouterBench/EmbedLLM-style matrices, HSE, perturbation robustness
Multi-turn trajectory state deltas, handoffs, retries, final answer goal drift, loop failure, bad recovery, compounded stochastic failure Anthropic agent eval concepts, Google trajectory/final response split, LangSmith/Phoenix spans, AutoGenBench
Judge and review rubric, judge model/version, human label, rationale judge bias, criteria drift, validator invalidity, label error LLM-as-judge frameworks, EvalGen-style validator validation, human annotation queues
Governance rollup normalized row, source bundle, failure code, review decision unverifiable score, missing trace, benchmark overclaim, unapproved training promotion eval_result_v1, failure taxonomy, GitLab component, warehouse rollups

Eval Domains And Hard Boundaries

Domain What it measures What it does not prove Enterprise storage boundary
Public model benchmark base capability on a known benchmark product reliability or private workflow fit model registry evidence only
Product task eval output quality for a workflow general model intelligence repo-local eval pack plus normalized rows
RAG/KB eval retrieval, grounding, citation, abstention source authority, freshness, ACL compliance unless separately checked source bundle plus retrieval trace
Coding-agent eval diff correctness, test pass, repo navigation, tool behavior maintainability or production safety unless reviewed repo snapshot, patch, test logs, trace
Agent trajectory eval tool choice, state updates, recovery, handoffs final answer truth by itself trace spans and action evidence
Judge/verifier eval rubric or reward reliability that the underlying model is safe or correct judge config, judge model version, calibration set
RL environment eval whether success can be verified cheaply and repeatedly that the reward cannot be hacked environment spec, verifier, exploit tests
Security/adversarial eval jailbreak, prompt injection, sabotage, exfiltration business-task success red-team suite plus policy failure code
Observability eval trace completeness and production drift task correctness without labels trace coverage, metrics, sampled reviews
Routing meaningfulness eval whether a router is choosing among differentiated actors and making stable assignments final-answer correctness, safety, or cost by itself route logs, behavioral matrix, HSE score, perturbation robustness
Training/SFT promotion reusable corrected behavior eval result validity redacted training example with lineage and approval

The hardest boundary is faithfulness vs truth. A RAG answer can be faithful to an unauthorized, stale, or wrong source. Grounding metrics must therefore sit beside source authority, freshness, ACL, and citation-span checks.

Framework Overlap

Concept DeepEval RAGAS LangSmith Phoenix Hugging Face Provider SDKs Corporate schema
Case input input user_input example/run input span input dataset sample user message/task task.input
Expected/reference expected_output, context reference, reference_contexts reference output annotation reference labels expected artifact/check expected_artifact
Actual output actual_output response run output span output prediction/generation final response attempt.actual
Retrieved evidence retrieval_context retrieved_contexts trace docs retrieval spans dataset docs KB/tool observations evidence.role=retrieved
Metric/judge metric object metric evaluator/feedback evaluator metric module/task grader/check metric_result
Trace optional/external external native run trace native trace logs/bundles native run spans trace_ref
Model/runtime metric/provider config evaluator/provider config run metadata span metadata model/backend revision model/runtime call model, harness, runtime_context
Failure class custom reason custom label feedback tag annotation postprocess error/event type failure_code

The overlap is enough for a shared warehouse row. It is not enough to force every team into one runner.

Lab And SDK Divergence

Lab/platform Eval posture SDK/runtime primitive Divergence to preserve
OpenAI eval definitions, graders, agent traces, third-party eval trust Evals API, graders, Agents SDK traces strong grader abstractions; trace and eval artifacts still need warehouse linkage
Anthropic agent-specific eval design, MCP/tooling, Claude Code-style permissioned agents Agent SDK, MCP, tool traces, Claude Code hooks emphasizes task design and trajectory/final outcome split more than one public schema
Google/DeepMind grounding/factuality research, ADK evals, A2A protocols ADK eval sets, Gen AI evaluation service, A2A task/artifact concepts protocol-level conformance matters alongside model scoring
Microsoft multi-agent and enterprise orchestration AutoGen, Foundry monitoring, Magentic-style systems team graph and handoff behavior are first-class eval surfaces
AWS Bedrock Agents, Knowledge Bases, AgentCore-style observability KB citations, action groups, Bedrock traces source authority and enterprise controls are platform-specific
Meta/Hugging Face/open model ecosystem model cards, datasets, LightEval/lm-eval style harnesses Hub revisions, LightEval, Evaluate, TRL/SFT, PEFT registry/training substrate is separate from product eval warehouse
NVIDIA safety and guardrail eval/tooling garak, NeMo Guardrails, runtime observability adversarial and policy checks sit beside task quality
Chinese and second-tier labs model reports, OpenAI-compatible APIs, function calling, local agent frameworks Qwen-Agent, DeepSeek APIs, Kimi/GLM/Tencent/Baidu SDK surfaces API compatibility does not mean behavior, tokenizer, tool policy, or eval compatibility

What Awesome Evals Changed

The 318-row ingestion does not overturn the existing architecture. It sharpens it.

Prior finding Updated after Awesome Evals ingestion
Normalize result rows, not frameworks Confirmed. The corpus spans runners, dashboards, papers, observability tools, talks, and RL environment work; one native schema would discard evidence.
RAG eval is a retrieval/faithfulness problem Expanded. RAG eval is also source authority, permissioning, freshness, abstention, and citation-span evidence.
Agent eval needs traces Strengthened. Several sources emphasize outcome-vs-trajectory, state deltas, pass@k/pass^k, recovery, stochasticity, and harness engineering.
LLM-as-judge is useful but risky Strengthened. Judge bias, criteria drift, validator validation, label errors, and benchmark ground-truth mistakes need their own eval lane.
Automated trace analysis can replace error analysis Rejected. Parlance Labs’ 100-trace production experiment shows strong recovery of trace-visible failures, but all tested systems missed product-context failures that looked acceptable in the trace.
Model routing is solved by cost/quality optimization Rejected. Google DeepMind’s routing-meaningfulness paper shows clean accuracy can diverge from behavioral diversity and perturbation robustness.
Benchmark scores are procurement weak evidence Strengthened. Benchmark integrity, label error, contamination, live/held-out tests, and leaderboard incentives are recurring themes.
GitLab runners can decentralize eval execution Confirmed. Repo-local execution is the only place private fixtures, code, tools, and workflow traces are naturally available.
SFT promotion should come from eval failures Expanded. The corpus points to verifier/RL environment design and trajectory training, not just prompt/completion SFT.

What The Parlance Auto-Evals Experiment Adds

Parlance Labs adds a missing empirical lane: automated error-analysis tools over real production traces. Their July 2026 experiment compared 100 human-annotated apartment-leasing agent traces against Braintrust Loop, Arize AX Alyx, LangSmith chat-agent analysis, Codex, Factory Droid, and Claude Code. The best system recovered 87.2% of the 39 human-labeled failures; Arize had the highest reported precision at 91.0%; coding agents were competitive with dedicated platforms on finding failures.

The important result is not the vendor order. It is the boundary:

Boundary What automated systems caught What they missed
Trace-visible failure unsupported claims, repeated asks, tool-output contradictions, unbacked promised follow-ups product objective failures that looked polite or complete inside the transcript
Product-context failure only when the rule was present in trace or prompt sales-objection handling, SMS Markdown, voice interruption behavior, handoff rules
Workflow value platforms kept findings linked to traces and supported annotation/rerun workflows every tested system treated the job as one-shot instead of interviewing the user or learning from annotations

This sharpens the operating model: automated trace analysis should be a review-acceleration loop, not an autonomous judge. The corporate schema should record error_analysis_mode, failure_visibility, human_annotation_ref, auto_issue_cluster, review_decision, and severity_calibration so teams can measure which failures are trace-visible versus product-context-dependent.

What The Routing Meaningfulness Paper Adds

Google DeepMind’s July 2026 routing paper adds a structural eval for AI gateways and multi-model systems. A router is only meaningful if the candidate actors are behaviorally differentiated and the assignment policy is stable under meaning-preserving perturbations.

This adds four failure codes to the enterprise eval taxonomy:

Failure code Meaning Evidence
routing.redundant_society routed models/agents behave too similarly for routing to matter actor list, behavioral matrix, HSE score
routing.assignment_instability paraphrases or minor rewrites route to different actors perturbation set, original route, perturbed routes, robustness score
routing.coreset_overbuild extra models add governance/cost burden without meaningful diversity HSE-vs-size curve, marginal diversity gain
routing.accuracy_robustness_split clean accuracy improves while perturbation robustness falls clean accuracy, perturbed accuracy, robustness by perturbation level

The practical deployment rule: gateway evals should report accuracy, cost, and latency, but also the diversity of the model pool and route stability. Otherwise a firm can buy a sophisticated router that mostly arbitrages redundant models or routes equivalent requests inconsistently.

Golden JSONL Exemplar Shape

Each Awesome Evals citation now has one exemplar row that can seed a corporate golden work packet. A row is intentionally not a full benchmark. It is a typed seed that says how this source should influence failure classification and evidence retention.

{
  "exemplar_id": "awesome_evals::ae-006-demystifying-evals-for-ai-agents",
  "source_id": "ae-006-demystifying-evals-for-ai-agents",
  "source_url": "https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents",
  "source_type": "web_article",
  "eval_domain": "agent_trajectory",
  "recommended_failure_code": "trajectory_error",
  "golden_case": {
    "persona": "enterprise_eval_owner",
    "task_id": "ae-006-demystifying-evals-for-ai-agents::classify_failure",
    "input": "Classify a failed AI-agent run using this source's eval concept and preserve the evidence boundary.",
    "expected_artifact_type": "source_document",
    "expected_fields": [
      "harness.name",
      "harness.version",
      "model.provider",
      "model.name",
      "runtime_context.prompt_template_id",
      "runtime_context.retrieval_index_id",
      "attached_tools",
      "trace_ref",
      "failure_code",
      "source_bundle_ref"
    ]
  }
}

The exemplar contract forces the team to record the harness, model version, tools, retrieval index, trace reference, failure code, and source bundle. That is the minimum shape needed for a failure to roll up from repo CI into a team dashboard, firm-wide scorecard, KB update, or training-data promotion queue.

Decentralized Rollup Model

The target operating model is:

  1. Product repo owns evals/<pack_id>/cases.jsonl, fixtures, source bundle manifest, and native runner configs.
  2. GitLab runner executes the native framework close to the code and secrets.
  3. Runner emits JUnit, JSONL, trace bundles, annotations, and raw framework exports.
  4. Team collector normalizes rows to eval_result_v1 and adds review status.
  5. Firm collector aggregates normalized rows, not private fixtures.
  6. Knowledge owners promote only reviewed, redacted, source-backed failures into wiki pages, KB patches, SFT rows, preference pairs, or policy gates.

This keeps private repo evidence decentralized while still giving leadership a firm-wide view of failure rates, model regressions, retrieval index drift, judge drift, and source-quality gaps.

Remaining Verification Backlog

The ingestion is enough to update the synthesis and wiki. It is not enough to declare every source fully processed through every modality-specific pipeline.

Backlog item Count Required next step
Media items 46 Captions unavailable; run yt-dlp plus Whisper/WhisperX ASR through the podcast/video path and keep needs_asr until transcript artifacts exist.
Paper/PDF-like items 83 All 83 paper/PDF rows have Chandra text evidence, including 2 chunked extractions.
Metadata-only rows 3 Revisit blocked/indirect sources and either mirror allowed content or keep as metadata-only citations.
Native schema gaps tracked separately Add exact export fixtures for framework versions whose hosted exports are not stable public schemas.

The practical conclusion is that the evaluation pillar now has a stronger map, a larger citation graph, and a machine-readable exemplar set. The remaining work is modality completion for the video/podcast/lecture rows.