标签：#evaluation

LLM Evaluation System Prompts Scored Rubrics Runtime Guardrails: A Practical Guide for Production Learn how to evaluate LLM outputs in production using system prompts, scored rubrics, and runtime guardrails to prevent hallucinations and ensure quality. TL;DR: To evaluate LLM outputs in production, combine system prompts that define evaluation criteria, scored rubrics using LLM-as-a-judge for dimensions like correctness and relevance, and runtime guardrails that filter or flag unsafe outputs. This approach scales better than human review, adapts via prompt changes, and catches failures that status codes miss, as seen in the Air Canada chatbot case. Why Production LLM Evaluation Demands More Than Status Codes A 200 status code only confirms the server processed the request—it says nothing about whether the generated text is factual, safe, or useful. The Air Canada chatbot that invented a non-existent bereavement discount returned perfectly valid HTTP responses, yet the hallucinated policy led to a tribunal ruling against the airline. Production evaluation must therefore separate operational health (latency, error rates) from output quality (correctness, relevance, harmlessness). Consider a typical API call that succeeds operationally but fails qualitatively: import requests response = requests . post ( " https://api.example.com/v1/chat " , json = { " model " : " gpt-4o " , " messages " : [{ " role " : " user " , " content " : " What is Air Canada ' s bereavement policy? " }]}, headers = { " Authorization " : " Bearer $KEY " } ) print ( response . status_code ) # 200 print ( response . json ()[ " choices " ][ 0 ][ " message " ][ " content " ]) # Output: "Air Canada offers full refunds for bereavement-related cancellations..." A 200 status code and a well-formed JSON body mask a completely fabricated policy. To catch this, you need a separate evaluation layer that scores the output against a rubric. LLM-as-a-judge is a common approach, using a second model to assess the

Learning how to add evals to an LLM feature is the difference between shipping a demo and shipping a reliable product. When you embed an LLM into a real feature — a chatbot, a voice agent, a document summarizer — you’re not just calling a model. You’re betting your user’s experience on a non‑deterministic system that can silently break with every prompt tweak, model update, or edge case. That’s why we instrument every LLM feature we build with a purpose‑built eval suite. Here’s how we did it for an outbound AI calling agent and how you can do the same. Why Evals Are Not Optional LLMs are non‑deterministic: give them the same input twice, and you’ll get two different responses. That means unit tests that check for exact string matches are useless. As Pragmatic Engineer notes , you need evals to verify that the solution works well enough — because there’s no guarantee it will. When you’re building a feature that speaks to real customers, like the AI Calling Agent dashboard we built, a regression in tone or missed booking intent can cost revenue immediately. Evals turn that uncertainty into signal. How to Add Evals to an LLM Feature: A 4‑Step Workflow We’ll walk through the exact process we followed, from defining success to automating checks in CI, using the DeepEval framework as an example. You can swap in Evidently AI or build your own, but the pattern is the same. Step 1: Define Success for Your Feature Takeaway: Before you pick a metric, write down the one thing that makes the feature “done” — usually a business outcome, not a technical measure. For the AI Calling Agent, the core feature was an outbound call that books a meeting. The success criterion wasn’t “the LLM replied politely.” It was “the agent scheduled a meeting with the right time and date.” This is a reference‑based evaluation: you compare the output to a known ground truth. Evidently AI’s guide calls this pattern out as essential for regression testing and experimentation. From that criterion, we der

Introduction Building Reliable LLM Applications in Python put it plainly: treat model output as a hypothesis to verify, not a fact to trust. Testing Best Practices in Python put the same discipline in pytest terms: a suite only earns trust by asserting the right things at the right level, unhappy paths included. This post is where those two ideas meet — a pytest assertion either passes or fails against a fixed expected value; an LLM's output is a paragraph of prose that might be right in spirit while differing token-for-token from anything you wrote down in advance. Evaluating it takes a harness, not an assert . That harness has three parts: a golden dataset of representative cases with known-good expected behavior, scoring that turns each case into a pass/fail or a number, and regression testing that runs the harness on every change and fails the build when the score drops. Making RAG Accurate in Python already gave you half of this story — recall@k, precision@k, MRR, nDCG measure whether retrieval found the right chunks. This post measures the other half: whether the generated answer built from those chunks is actually good, which is a genuinely different question a retrieval metric can't answer on its own. Everything below is illustrative, non-executed Python, grounded in the same Anthropic SDK shapes as posts 10/11. The Golden Dataset: Curating Cases, Not Just Inputs A golden dataset is a small, hand-curated set of (input, expected behavior) pairs that represents the ways your application is actually used — not a random sample, and not just the cases that already work. Each case needs enough structure to be scored automatically later: from dataclasses import dataclass , field @dataclass class EvalCase : id : str category : str # "extraction", "qa", "summarization", ... input : str # the prompt/question sent to the system under test expected_exact : str | None = None # non-None only for cases scorable by exact match must_contain : list [ str ] = field ( default_f

Introduction Building Reliable LLM Applications in Java put it plainly: treat model output as a hypothesis to verify, not a fact to trust. Testing Best Practices in Java put the same discipline in JUnit terms: a suite only earns trust by asserting the right things at the right level, unhappy paths included. This post is where those two ideas meet — a JUnit test either passes or fails against a fixed expected value; an LLM's output is a paragraph of prose that might be right in spirit while differing token-for-token from anything you wrote down in advance. Evaluating it takes a harness, not an assertEquals . That harness has three parts: a golden dataset of representative cases with known-good expected behavior, scoring that turns each case into a pass/fail or a number, and regression testing that runs the harness on every change and fails the build when the score drops. Making RAG Accurate in Java already gave you half of this story — recall@k, precision@k, MRR, nDCG measure whether retrieval found the right chunks. This post measures the other half: whether the generated answer built from those chunks is actually good, which is a genuinely different question a retrieval metric can't answer on its own. Everything below is illustrative, non-executed Java, grounded in the same Anthropic Java SDK shapes as posts 10/11. The Golden Dataset: Curating Cases, Not Just Inputs A golden dataset is a small, hand-curated set of (input, expected behavior) pairs that represents the ways your application is actually used — not a random sample, and not just the cases that already work. Each case needs enough structure to be scored automatically later: public record EvalCase ( String id , String category , // "extraction", "qa", "summarization", ... String input , // the prompt/question sent to the system under test String expectedExact , // non-null only for cases scorable by exact/programmatic match List < String > mustContain , // key facts a correct answer must mention (programma

Why Workflows Need a Dedicated Evaluation Framework Traditional software testing covers code correctness. Workflows add two layers of uncertainty: LLM output is non-deterministic : the same input can produce different results across runs Cross-step dependencies : a Phase 3 problem may only surface at Phase 7, making the debugging chain long Without an evaluation framework, every workflow change requires a full end-to-end run: slow, expensive, incomplete coverage. Three-layer testing decomposes the problem. Three-Layer Evaluation Structure Layer 3: End-to-end tests (Workflow level) Full pipeline from trigger to completion Test cases: eval/cases.yaml Metrics: completion rate, Phase 4 avg rounds, gate trigger rate Layer 2: Integration tests (Phase level) Cross-step data flow is correctly passed Cross-phase routing logic fires correctly Layer 1: Unit tests (Step level) Each subagent's output matches its output contract No real LLM calls — validates JSON schema only Test priority: Layer 1 should be the most numerous and fastest — catches contract violations in seconds. Layer 3 is the slowest and most expensive — run it only when changes affect the main pipeline. Layer 1: Step-Level Unit Tests Unit tests verify that subagent output files match the declared schema. No real LLM calls needed. # tests/unit/test_phase3_output.py import json from pathlib import Path def test_analysis_output_schema (): """ Phase 3 output must conform to analysis_final.json schema """ output = json . loads ( Path ( " test_fixtures/phase3/analysis_final.json " ). read_text ()) assert " passed " in output assert isinstance ( output [ " passed " ], bool ) assert " confidence " in output assert 0.0 <= output [ " confidence " ] <= 1.0 assert " root_cause " in output assert isinstance ( output [ " root_cause " ], str | type ( None )) assert " evidence " in output assert isinstance ( output [ " evidence " ], list ) # on failure, error field must be present and non-empty if not output [ " passed " ]: ass

A large-scale audit of AI-as-judge evaluation — covering over half a million individual judgments — finds that AI judges are consistently reliable but not valid, meaning they give the same answer repeatedly without that answer being correct. Published work and popular benchmarks like Chatbot Arena have treated consistency as proof of trustworthiness, and the audit shows that assumption is unfounded. Key facts What: Using one AI to grade another is now common — but the biggest audit yet shows these graders are consistent without being correct. A judge that always picks "answer A" scores perfectly on consistency. When: 2026-06-19 Primary source: read the source (arXiv 2606.19544) The distinction matters: a judge is reliable if it's consistent (same question, same answer), and valid if those answers are actually correct. The audit's central finding is that AI judges are reliable without being valid, and the field has been treating the first as evidence of the second. Because consistency is easy to measure and looks reassuring, it has stood in for actual trustworthiness across a lot of published work. A new audit makes the problem stark: a judge that ignores both answers and always picks the one labeled "A" would be perfectly consistent — flawless reliability, identical verdict every time — and completely worthless, because it never read anything. Consistency is trivially easy to fake and says almost nothing about whether the judging is sound. Yet "the judge agrees with itself" has done significant reassurance work in papers and benchmarks, and the always-pick-A example shows exactly how empty that reassurance is. When the researchers corrected for the agreement you'd get by chance — as any fair test should — confident-looking scores deflated noticeably. Gaps between models that seemed meaningful shrank or blurred. Accepted folk wisdom also took a hit: the long-standing worry that AI judges are suckers for longer, wordier answers turned out to be far weaker than assumed

Agent-testing startup Patronus AI, founded by former Meta AI researchers, is experiencing nearly insatiable demand, its investor says.

Stop Asking 'Is GAI Here' — Ask 'At What Layer' The GAI debate has a structural problem. Someone says "passing this benchmark means GAI." A model passes it. Then they say "that benchmark wasn't hard enough." The goalpost moves. Someone says "passing the Turing test means GAI." Models pass it. Then they say "the Turing test is too easy." The goalpost moves again. Someone says "inventing new mathematics means GAI." Models do it. Then they say "that's just pattern matching in disguise." Goalpost moves. This isn't bad faith. It's a missing layer definition. We never agreed on what "general" means. Without that, every achievement gets reclassified as "not really general." I've been working on a framework that might fix this. It started as a capability map. Then I realized: this isn't just a map. It's a GAI maturity model. The Five Layers Layer Name Definition L0 Embodied Perceive and operate in the physical world L1 Application Complete single-domain tasks using tools L2 Engineering Build and maintain systems L3 Meta-Domain Abstract and transfer between unrelated domains L4 Meta-Cognition Perceive and control your own thinking process The rule: layers cannot be skipped. It's a maturity sequence, not a checklist. This immediately explains the goalpost problem: some people define GAI as L1. Others define it as L4. They're using different layers for the same word. What About Models Without Bodies? L0 requires embodiment. Text-only models don't have bodies. The cleanest answer: LLMs have no L0. They start at L1 — cognition without embodiment. This isn't a defect. It's an architectural difference. Humans build up from L0 (a baby senses the world before understanding it). LLMs start at L1 (they understand the world directly, skipping physical experience). The result: humans can "feel" when something is wrong — that's L0 feeding signals up to L4. LLMs don't have this channel. The framework forced me to face something uncomfortable: human intelligence cannot exist without a body

Most teams I talk to have "evals." I ask them where the evals run. The answer is almost always the same: a notebook, a dashboard, a spreadsheet someone updates after a bad week. That is not an eval suite. That is a museum. Here is the opinion I will defend for the rest of this post: if your agent's quality checks cannot block a merge, they are decorative. The entire value of an eval is that it stops a regression before it reaches a user. A score you read on Monday about a deploy you shipped Friday is a postmortem, not a gate. We gate code with unit tests. We gate APIs with contract tests. We gate infra with terraform plan . Then we take the single most non-deterministic component in the stack — an LLM agent that can silently change behavior when a vendor ships a new checkpoint — and we let it through on vibes. That asymmetry is the actual bug. Why "run it locally and eyeball it" rots The failure isn't that engineers are lazy. It's that manual eval runs degrade under exactly the conditions where you need them most: Prompt edits look harmless. You reword one line of a system prompt to fix a tone complaint and tank tool-selection accuracy on three unrelated tasks. Nobody re-ran the full set because it's a 40-minute slog. The model moves under you. You pinned gpt-4o , but gpt-4o is not a constant — providers roll checkpoints. Your prompt is identical and your behavior shifted anyway. Dependencies leak. A retrieval index gets re-embedded, a tool's API changes a field name, and the agent starts confidently citing stale data. None of that shows up in a code diff. Every one of these passes code review. Every one of these is caught by a regression suite that runs on the PR. The fix is boring and it works: treat agent behavior like any other thing you'd protect with CI. The two halves you actually need A CI gate for agents needs two things, and people consistently build only one. The first is a scorer : something that takes the agent's output for a fixed set of inputs and ret

Coding agents are no longer just autocomplete with a longer prompt. GitHub describes Copilot cloud agent as software that can research a repository, create an implementation plan, make code changes on a branch, run in an ephemeral GitHub Actions-powered environment, and let a developer review or create a pull request afterward. OpenAI's Codex GitHub integration similarly positions code review as a repository-aware review pass that follows AGENTS.md guidance and focuses comments on serious issues. That shift changes the buyer question. The useful question is not "does the agent usually write code?" It is "can the team detect when the agent drifts away from the developer's intent before the change reaches production?" A May 2026 arXiv paper, "How Coding Agents Fail Their Users" , gives teams a better vocabulary for that review. The authors studied 20,574 real IDE and CLI coding-agent sessions across 1,639 repositories and define misalignment as a breakdown that becomes visible through developer correction or pushback. The paper reports seven recurring symptom categories: wrong project diagnosis, misread developer intent, developer constraint violation, self-initiated overreach, faulty implementation, operational execution error, and inaccurate self-reporting. Effloow Lab also ran a bounded OpenAI API check using three synthetic, non-confidential coding-agent transcript snippets. The run did not measure real-world incidence, compare vendors, or reproduce the paper. It produced a small rubric that maps visible symptoms to review gates such as diff-scope checks, evidence-before-edit checks, acceptance-criteria coverage, and verification-output requirements. The public lab note is available at /lab-runs/coding-agent-misalignment-failure-taxonomy-poc-2026 . This guide turns that research and lab output into a practical QA checklist for teams buying, piloting, or packaging coding-agent workflows. Why This Matters for Agent Buyers Coding-agent procurement often starts with p

The general shape of the problem is that every public LLM benchmark is on a saturation clock that runs from the moment of its publication to the moment a model's training corpus has eaten it. The clock has been running, on the visible benchmarks of the last five years, for somewhere between twelve and thirty months before each one is no longer useful for differentiating frontier models. The benchmarks are not failing. They are doing exactly what they were designed to do, in the order they were designed to do it, and the field has been running through them faster than the people designing them anticipated. I want to put numbers on the saturation pattern, walk through what the contamination evidence actually says, and then sit with the question of what an honest benchmark would have to look like in 2026 — because the "private held-out eval" answer that the labs are converging on has economics that are worth examining carefully before any of us salute it as the solution. The saturation timeline, with numbers HumanEval (Chen et al., OpenAI, July 2021). 164 hand-written Python problems. The benchmark was published with Codex at 28.8% pass@1; the underlying GPT-3 base model scored 0%. GPT-4 (March 2023) hit 67% in the original Technical Report. By late 2024, OpenAI's o1-preview and o1-mini both reached 96.3% pass@1 ; Claude 3.5 Sonnet sat at 93.7%. The benchmark is saturated in the operational sense — the relative spread across the top ten models is around 10 percentage points, which is too small a gap to differentiate them on, and most of the new models arrive within a percentage point or two of the ceiling. The successor variants (HumanEval+ from EvalPlus, with augmented test cases) are the field's response. Lifespan from publication to operational saturation: about 36 months. MMLU (Hendrycks et al., September 2020). 57 subjects, ~14,000 multiple-choice questions, taken from publicly-available test prep and academic sources. The problem with MMLU is not that it's satura

Architectural change cases extend architecture decision record (ADR) thinking by evaluating how decisions may evolve over time. Change cases expose hidden assumptions and help teams estimate the reversibility and cost of change. By Pierre Pureur, Kurt Bittner

What is an LLM evaluation harness? A deep dive into lm-eval-harness You fine-tuned a 7B model. It aced your smoke tests, your colleague ran a few prompts and shrugged approvingly, and the README is now full of cherry-picked outputs that look great in a screenshot. Then someone asks: how good is it, really? — and you realize you have no number to point at. No MMLU score. No HellaSwag. Nothing reproducible, nothing you can defend in a PR review, nothing you can compare to last week's checkpoint. That's the gap an evaluation harness fills. It turns "vibes-based evaluation" into something with a score, a stderr, and a config file you can re-run next Tuesday. Why evaluate LLMs at all? Two reasons that actually matter: Comparability. If you can't put a number on a model, you can't compare it to anything else — not the previous checkpoint, not the open-source baseline, not the commercial API you're trying to replace. Leaderboards are noisy and gaming-prone, but a local leaderboard with the tasks you care about is one of the most useful artifacts a team can build. Regression detection. Most model regressions are silent. A 0.3-point drop on MMLU won't show up in a chat session, but it will show up in CI. People who ship models for a living treat evals the way backend engineers treat unit tests: mandatory, run on every PR, and blocking on regressions. You don't need a hundred benchmarks. You need the three to five tasks that map to your actual use case , plus one or two general capability anchors (MMLU, HellaSwag) so you can sanity-check that you didn't accidentally destroy basic reasoning while you were tuning for your domain. What is an "evaluation harness"? An evaluation harness is the software that sits between a model and a benchmark. It handles the boring-but-critical parts: loading the model weights, tokenizing prompts in the way the benchmark expects, running inference, extracting the answer from a longer generation, scoring it against a ground-truth key, aggregating

Microsoft on Tuesday took the wraps off Adaptive Spec-driven Scoring for Evaluation and Regression Testing, an open source framework for spinning up AI evaluations.