How to measure "quality" in AI outputs (beyond accuracy)
Your accuracy score is green.
Your users are quietly churning. Both things can be true at the same time — and if you're only watching one metric to measure AI output quality, you probably won't notice until the retention data catches up.
The core problem is that AI output quality isn't a single thing. It's at least seven independent dimensions: usefulness, relevance, hallucination rate, task completion, safety, latency, and user satisfaction. Each one can move in a different direction after a model update.
A change that improves response speed can quietly degrade groundedness. A safety tuning pass can tank task completion rates. A model that scores well on benchmarks can still fail three out of four experienced users in production — which is exactly what a METR study of AI coding tools found. Single-score evaluation doesn't surface these trade-offs. It hides them.
This article is for engineers, PMs, and data teams who are shipping AI features and want a more honest picture of what's actually working. Here's what it covers:
- Why accuracy and other traditional metrics break down for generative AI
- The seven quality dimensions that actually matter — and why they move independently
- The three-layer evaluation stack (automated metrics, human review, and behavioral signals) and what each layer can and can't measure
- How to close the gap between model-level scores and real user outcomes
- How to build continuous monitoring so quality degradation is detectable before users report it
Each section builds on the last, moving from the structural problem with single-metric evaluation to the practical infrastructure for measuring AI output quality in production — including where tools like GrowthBook fit into the monitoring and rollback layer. By the end, you'll have a framework you can actually act on.
Why accuracy alone is a misleading proxy for AI output quality
If your team is currently evaluating AI outputs using accuracy, F1-score, precision, or recall, that's a reasonable place to have started. These metrics have decades of research behind them, they're well-understood, and they integrate cleanly into existing ML pipelines.
The problem isn't that accuracy is a bad metric — it's that it was built for a fundamentally different kind of system, and applying it to generative AI creates blind spots that can quietly undermine your product while your dashboards look healthy.
How accuracy metrics were designed — and what they were designed for
Accuracy measures the ratio of correct predictions to total predictions. It was designed for classification tasks: spam or not spam, fraud or not fraud, cat or dog. These are bounded-output systems where every prediction can be checked against a discrete ground-truth label. The math is clean, the interpretation is intuitive, and for the right problem, it works.
Except even in its native context, accuracy is fragile. Consider a fraud detection model that predicts "no fraud" for every single transaction. On a dataset where 99% of transactions are legitimate, that model achieves 99% accuracy — and completely fails at its actual purpose.
This isn't a generative AI problem; it's a structural flaw in accuracy as a metric when class distributions are imbalanced. Generative AI doesn't introduce a new problem so much as it amplifies an existing one into something much harder to ignore.
Why generative AI breaks the assumptions accuracy relies on
Classical accuracy requires a ground-truth label to compare against. Generative AI rarely has one. When you ask a language model to summarize a document, draft a response to a customer complaint, or explain a technical concept, there are thousands of valid outputs — and no single correct answer. The entire premise of the metric collapses.
This isn't just a theoretical concern. A Deloitte survey from mid-2025 found that one-third of generative AI users had already encountered incorrect or misleading answers from models that score well in controlled lab evaluations. The lab-to-production gap is real, and accuracy scores don't surface it.
A model can be "95% accurate" in benchmark conditions and still fail users through slowness, inconsistency, hallucination, or contextual inappropriateness — none of which the accuracy score will flag.
The METR study of experienced open-source developers using AI coding tools reinforces this from a practitioner angle: three out of four participants saw reduced performance when using AI assistance, despite the tool's productivity claims. Model capability, as measured internally, did not predict real-world user outcomes.
The gap between what a model scores and what it delivers is not a rounding error — it's a structural feature of how these systems work.
The hidden cost of relying on a single score
When a single metric becomes the primary signal for AI output quality, it doesn't just fail to capture what it misses — it actively creates a false sense of security. A composite accuracy score can remain stable while latency degrades, hallucination rates climb, and user trust quietly erodes.
The erosion part is particularly dangerous. Most users won't file a bug report when an AI output is subtly wrong, off-tone, or unhelpful. They'll just stop using the feature. By the time that shows up in retention data, the accuracy score has been green for months. The signal you were monitoring was never measuring the failure mode that mattered.
This is the core structural problem: accuracy collapses multiple independent failure modes — safety, relevance, groundedness, task completion, latency — into a single number. Optimizing that number doesn't mean you've addressed those dimensions. It means you've made them invisible.
The implication isn't that accuracy should be discarded. It's that it needs to be understood as one narrow signal within a broader measurement stack. What that stack looks like — and which dimensions it needs to cover — is where the real work of measuring AI output quality begins.
The seven dimensions that actually define AI output quality
The instinct to reduce quality to a single score is understandable. Single scores are easy to track, easy to report upward, and easy to compare across model versions. The problem is that AI output quality is not a single thing.
It is a composite of at least seven independent dimensions — usefulness, relevance, hallucination rate, task completion, safety, latency, and user satisfaction — each of which can move in a different direction at the same time. A model update that improves one dimension can silently degrade another, and if you're only watching one metric, you won't see it happening.
This is not a new insight. In 1996, researchers Richard Y. Wang and Diane M. Strong published "Beyond Accuracy: What Data Quality Means to Data Consumers," identifying fifteen distinct quality dimensions for data and arguing that accuracy alone was an insufficient proxy for the rest.
Thirty years later, teams building AI-powered products are rediscovering the same structural problem: the dimensions are independent, they trade off against each other, and collapsing them into one number destroys the signal you need to make good decisions.
Each quality dimension can degrade without moving the others
Each quality dimension captures something the others don't. Usefulness measures whether the output actually helps the user accomplish what they came to do. Whether the response addresses the specific question asked — rather than a plausible-sounding adjacent one — is what relevance captures. Hallucination rate tracks how often the model generates confident-sounding content that is factually wrong or unsupported.
Task completion is distinct from relevance: a response can be on-topic and still fail to complete the user's underlying goal. Safety captures whether the output avoids harmful, biased, or policy-violating content. Latency measures how quickly the response is delivered. And user satisfaction aggregates the user's subjective experience, which often diverges from any of the objective dimensions above.
The critical point is that these dimensions are not correlated by default. A model tuned aggressively for safety may refuse valid requests, driving task completion rates down. A model optimized for low latency by truncating responses may produce outputs that score well on speed but fail on usefulness. A model with a low hallucination rate on benchmark datasets may still generate contextually inappropriate responses that tank user satisfaction in production.
The trade-off problem in practice
When teams optimize for a single dimension without monitoring the others, they tend to create what looks like improvement but is actually a transfer of degradation to an unmeasured axis. This is the same dynamic that GrowthBook's Metric Correlations feature is designed to surface: the platform explicitly asks whether maximizing one key metric is coming at the cost of another.
The docs describe a scenario where an experiment drives purchasing behavior — revenue goes up — while retention goes down, flagging this as a potential dark pattern. The same logic applies directly to AI quality dimensions. An experiment that improves response speed while quietly reducing groundedness is not a quality improvement; it's a trade-off that hasn't been acknowledged yet.
The METR study of experienced open-source developers using AI coding tools illustrates this at the practitioner level. Three out of four participants saw reduced performance when using AI assistance, despite the expectation of productivity gains. The developers were, in effect, optimizing for one dimension — code generation speed — while degrading others, including actual task quality and accuracy.
The one developer who saw positive results had more than fifty hours of experience with the tool, suggesting that even "task completion" as a dimension is context-sensitive and cannot be read from throughput metrics alone.
Volume metrics as a masking problem
High-volume metrics — tokens generated, responses per second, code lines written — are particularly dangerous proxies because they measure activity, not quality. As Collibra notes in the context of data quality: "You cannot rely on just one metric."
When applied to AI outputs, this means that a team watching throughput while ignoring hallucination rate, task completion, and user satisfaction is not measuring quality at all. They are measuring production volume, which can increase while quality across every meaningful dimension simultaneously declines. The dimensions don't average out. They have to be tracked separately, with explicit attention to the trade-offs between them, or the picture you're building is incomplete by design.
Three evaluation methods, each measuring what the others miss
The instinct to reach for automated metrics when evaluating AI output is correct — but it's incomplete. No single evaluation method covers the full surface area of quality dimensions like groundedness, contextual appropriateness, task completion, and user satisfaction.
What teams actually need is a layered stack where each method is assigned to the quality signals it can genuinely measure, with defined handoffs between layers. The question isn't which method to use. It's understanding what each one captures, where each one breaks, and how they work together.
Automated evaluation frameworks: fast, scalable, and narrow
Automated metrics — precision, recall, F1, BLEU, ROUGE, perplexity — were designed for bounded-output tasks: classification, regression, structured extraction. For those tasks, they provide objective performance data and enable direct comparison between model versions. They run continuously, integrate cleanly into CI/CD pipelines, and scale without human intervention.
The problem is structural. When you ask a generative model to summarize a document or answer a question, there are dozens of valid responses — all slightly different, all potentially correct. That flexibility is what makes generative models useful. It's also what breaks automated scoring.
Most automated metrics work by comparing the model's output against a reference answer and measuring how closely they match. BLEU, for example, counts how many words or phrases the model's response shares with the reference. That tells you something about surface similarity — but it can't tell you whether the response actually makes sense in context, whether it's appropriate for the specific user asking, or whether it would help someone get their job done.
A model that scores well on these metrics might still produce outputs that confuse users, miss the point of the question, or confidently state something false. Automated scoring is fast and scalable, but it's measuring a narrow slice of quality — not the full picture.
Human-in-the-loop review: the calibration layer
Human review is the only method that can reliably evaluate dimensions like tone, nuanced usefulness, safety in context, and whether a response is appropriate for a specific user situation. Automated metrics cannot score these signals. For that reason, human review remains the gold standard for evaluating the quality dimensions that matter most.
The constraint is scalability. Human review cannot scale across large prompt datasets or high-output production systems — it's expensive, slow, and inconsistent without structure. This means human review should function as a calibration layer, not a primary monitoring mechanism. When automated scores shift unexpectedly, human review validates what's actually happening. When a new model variant ships, human review samples outputs across representative prompt categories before full rollout.
When human review is conducted, it should be systematic. The VALID-AI framework from Toronto Metropolitan University offers a practitioner-facing checklist: validate data quality and representativeness, analyze algorithm behavior including bias in reinforcement learning from human feedback, account for legal and ethical considerations, assess interpretability and transparency, and evaluate AI-specific behaviors like refusal patterns and prompt sensitivity.
Ad hoc human review produces inconsistent signals. Structured human review produces calibration data that improves the automated layer over time.
Behavioral and engagement signals: the closest proxy to real value
Behavioral signals — what users actually do after receiving AI output — are the most direct proxy for quality dimensions like usefulness, task completion, and user satisfaction. Session engagement, retry and regeneration rates, feature adoption, task completion rates, and downstream retention metrics all reflect whether AI output is actually working in the context users care about.
This is also the layer where the most dangerous assumptions live. The METR study finding holds here too: high-volume behavioral signals from early adoption periods can be systematically misleading — users unfamiliar with a tool depress performance independently of model quality. Flat engagement metrics also miss task-type segmentation: AI output quality looks very different on well-documented, common tasks versus complex, niche ones.
A feature flag rollout approach to AI evaluation is built around this behavioral layer — using gradual rollouts with engagement data monitoring, A/B testing model variants against user behavior outcomes, and statistical guardrails to trigger deactivation if quality signals degrade.
Landon Smith, Head of Post-Training at Character.AI, describes the value directly: "We can compare different modeling techniques from the perspective of our users — guiding our research in the direction that best serves our product." That framing — quality as experienced by users, not scored by automated metrics — is what the behavioral layer is designed to capture.
Building the stack: assigning each method to what it can actually measure
The practical synthesis is straightforward, even if the execution isn't. Automated metrics run continuously and provide scalable coverage of bounded quality signals — they're the monitoring layer. Human review runs on sampled outputs at defined intervals or when automated scores shift — it's the calibration layer. Behavioral signals captured through production A/B tests and feature flag rollouts provide ground truth on real-world value — they're the outcome layer.
Each layer has a defined role and a defined limitation. Automated scoring catches regressions fast but misses judgment-dependent quality signals. Human review evaluates those signals but can't scale to production volume. Behavioral signals reflect actual user value but can be confounded by adoption curves, task distribution, and user familiarity.
The stack works because each layer compensates for the others' blind spots — not because any one of them is sufficient on its own.
Closing the gap between model-level metrics and real user outcomes
The three-layer evaluation stack described above can tell you a great deal about how a model performs — but there's a measurement crisis hiding inside most AI deployments that even a well-instrumented stack can miss. The problem isn't that organizations lack data — it's that the data they're collecting answers the wrong question. Model-level metrics tell you how the AI system is performing. They don't tell you whether users are actually better off.
The benchmark-to-reality gap
The most clarifying data point on this problem comes from DX's analysis of over 400 companies. Despite a 65% increase in AI tool usage across those organizations, median pull request throughput increased by only 8%. Most companies landed somewhere in the 5–15% range.
That's a real gain — but it's a far cry from the 3x to 10x productivity improvements that some vendors imply in their marketing materials, or the claims that AI is now writing 20–30% of production code.
The gap isn't a fluke. It's structural. Benchmarks are run on controlled tasks with experienced users operating under favorable conditions. Production environments involve fragmented tool use, varied skill levels, and the kind of messy, context-dependent work that controlled evaluations are specifically designed to exclude.
A model that performs well on a benchmark has demonstrated that it can perform well on that benchmark — nothing more.
The METR finding reinforces this directionally: adoption metrics and outcome metrics are different phenomena, and treating one as a proxy for the other is a measurement error.
How operational metrics mislead
Google Cloud has flagged this directly: most organizations are "confusing operational efficiency gains with end state goals". Faster response times, lower token costs, reduced latency — these are real improvements to the AI system's performance. They are not evidence that users are accomplishing more meaningful work.
The DX defect rate finding makes this concrete. Some teams have been shipping up to 50% more defects since AI adoption. At the same time, daily AI users are merging 60% more pull requests than non-users. Both of those statistics are true simultaneously — which is precisely the problem.
Volume metrics and quality outcomes can move in opposite directions, and a measurement stack that only captures throughput will read the situation as a success while the product degrades.
Operational metrics measure how the AI system is running. Outcome metrics measure whether users and the business are better off. These are not the same measurement, and they don't reliably correlate.
Instrumenting for downstream outcomes
The practical question is what to measure instead. DX's framework argues that AI performance requires tracking utilization, impact, and cost together — that any single layer in isolation produces a misleading picture. Glean's approach to measuring AI's impact on decision-making goes further, identifying decision quality, time to decision, user confidence, and business outcomes as the metrics that actually matter, explicitly positioning these as a step beyond simple adoption rates.
Google Cloud's four-layer KPI framework offers a useful organizing structure: model quality, system performance, user engagement, and financial impact. Most teams instrument layers one and two. The argument here is that layers three and four are where the actual accountability lives.
The practical implication is that teams need to define what downstream success looks like before deploying AI features — what user behaviors should change, what business metrics should move, and over what time horizon.
Platforms that enable gradual rollouts tied to engagement data and surface metric correlations across experiments — throughput up, quality down, for instance — give teams the ability to observe the trade-offs that single-metric measurement would obscure. For teams whose outcome metrics live in a data warehouse, warehouse-native experiment analysis allows results to be computed directly against that data, keeping the outcome layer connected to the same platform managing rollouts and guardrails.
Landon Smith, Head of Post-Training at Character.AI, describes the orientation shift directly: "We can compare different modeling techniques from the perspective of our users — guiding our research in the direction that best serves our product." That's the move this section is arguing for: from model-level signals to user-level outcomes, measured in production, with the ability to act on what you find.
Quality is not a one-time eval: building continuous monitoring for AI outputs
Closing the gap between model-level metrics and user outcomes requires not just better measurement at launch, but measurement that persists. Running a thorough evaluation before you ship an AI feature is necessary. It is not sufficient.
The structural problem is that LLM outputs are non-deterministic, model providers push updates without announcement, and the distribution of real user inputs drifts away from your test set the moment you go live. A passing eval score is a snapshot of quality at a specific moment under specific conditions — not a guarantee of what users will experience tomorrow.
Why static evals fail in production
Researchers studying clinical AI deployment found that even after successful integration, ML and AI algorithms are "sensitive to changes in the environment and liable to performance decay". Their recommendation was unambiguous: continuous monitoring and updating, not one-time evaluation. The clinical context is different from a product LLM, but the performance decay logic is identical.
The Claude Code regression in January 2025 is the clearest concrete example of what this looks like in practice. A harness issue was introduced on January 26th and not caught until January 28th — two full days of degraded outputs reaching users before Anthropic's team identified and rolled back the change. The regression was confirmed publicly by an Anthropic team member.
Users lost tokens to lower-quality outputs during that window, and the incident generated significant frustration around both the detection lag and the inconsistent handling of refund requests. This was not a hypothetical failure mode. It was a named model, specific dates, and documented user impact — and it happened to one of the most closely watched AI products in the industry.
The developer community has drawn the obvious conclusion: practitioners are now building daily benchmark pipelines specifically to track model degradation, because they cannot rely on providers to surface regressions proactively. The tooling is still largely custom-built. The need is real and immediate.
Monitoring for model drift and prompt sensitivity
What teams should watch for in production falls into two categories. The first is model drift — gradual or sudden changes in output quality that result from provider-side updates, fine-tuning changes, or shifts in the underlying model behavior. The second is prompt sensitivity at scale: prompts that performed well in evaluation may behave differently when exposed to the full variance of real user inputs, edge cases your test set never covered, or subtle rephrasing patterns that shift output quality in ways that aggregate metrics won't catch.
A monitoring stack for this needs end-to-end tracing across prompts, model calls, retrieval steps, and tool use — not just output-level scoring. It also needs prompt versioning and evaluation management so that changes to prompts are treated with the same rigor as code changes. The goal is to make quality degradation detectable before users report it, not after.
Guardrail metrics and rollback capabilities
Detection without the ability to act quickly is incomplete. The operational complement to monitoring is rollback — and the infrastructure layer that makes rollback practical is guardrail metrics.
Within a feature flagging and experimentation platform, guardrail metrics, gradual rollouts, and rollback capabilities operate as integrated parts of the same workflow — not separate tools bolted together. Guardrail metrics can be configured per-metric with customizable thresholds, so a regression in error rate or response quality triggers an alert before it propagates to your full user base.
Suspicious uplift detection flags when a metric moves by an implausible amount in a single experiment — often a signal of a bug rather than a genuine effect. Gradual rollouts limit blast radius by exposing AI features to a subset of users first, using engagement data to validate quality before full deployment. And when something goes wrong, instant feature deactivation means you can pull a feature without a code deploy.
These capabilities work together because they share the same targeting, segmentation, and metrics infrastructure — which is what makes the detection-to-rollback cycle fast enough to be useful.
Building a feedback loop from user behavior to model evaluation
The final piece is closing the loop between what happens in production and what you test against. User behavior — task completion, engagement patterns, abandonment — is a quality signal that no pre-launch eval can fully replicate. That signal should feed back into your evaluation process, turning production failures into stronger test coverage and surfacing the real-world conditions your benchmarks missed.
Character.AI's head of post-training described this directly: using GrowthBook to compare different modeling techniques from the perspective of users, guiding research in the direction that best serves the product. That framing — model evaluation driven by user outcomes rather than internal benchmarks — is what continuous monitoring is actually for.
As GrowthBook's AI testing playbook puts it, evals are just the tip of the iceberg. The ongoing loop between user behavior and model evaluation is where quality compounds over time.
Continuous monitoring is not just a safety net. It is the mechanism by which teams build organizational confidence in AI outputs — and the only way to know whether the quality you measured before launch is the quality your users are actually experiencing.
Building a measurement program that catches what single scores hide
The argument this article has made is simple, even if the execution isn't: AI output quality is multidimensional, the dimensions move independently, and any measurement program that collapses them into a single score is hiding the failure modes that matter most.
A green accuracy score and quietly churning users are not contradictory — they're what happens when your measurement stack is narrower than the problem it's supposed to catch.
Start with the dimensions most likely to mask risk in your specific use case
Not every quality dimension carries equal risk for every product. A customer-facing support bot has a different risk profile than an internal code assistant — safety and hallucination rate may dominate in one context, while task completion and latency dominate in the other.
Before you build anything, identify which two or three dimensions, if they degraded silently, would cause the most damage. Those are your first monitoring targets, not the ones that are easiest to automate.
Layer your evaluation stack: automated scoring, human review, and behavioral signals
The three-layer stack — automated metrics as the monitoring layer, human review as the calibration layer, behavioral signals as the outcome layer — works because each layer compensates for what the others miss. The temptation is to pick one and call it done. Resist it.
Automated scoring without behavioral signals means you're measuring model performance, not user outcomes. Behavioral signals without human review means you can see that something is wrong but not why. The layers are interdependent, and the stack only works when all three are running.
Instrument for continuous monitoring before you scale AI features
The Claude Code regression — two days of degraded outputs reaching users before detection — is a useful forcing function. Ask yourself honestly: if your model provider pushed a silent update tonight, how long would it take you to notice? If the answer is "when users complain," your monitoring is behind your deployment.
Set up guardrail metrics and gradual rollouts before you scale, not after. Feature flagging and guardrail metric infrastructure exists precisely for this: catch regressions when the blast radius is still small enough to recover from.
The tension worth holding onto is this: building a complete measurement program takes real investment, but the cost of not having one compounds quietly. You won't see it in your dashboards. You'll see it in retention.
This article is meant to give you a framework you can actually act on — not a theoretical model, but a practical starting point grounded in how these systems fail in the real world.
What to do next: Map your current measurement stack against the seven dimensions. If you're only tracking one or two — accuracy, latency, or throughput — identify which unmeasured dimensions carry the most risk for your specific use case. Start there. Add one behavioral signal tied to a downstream user outcome (task completion, retention, regeneration rate) and one guardrail metric with a defined threshold. That combination — one outcome signal, one guardrail — is enough to start catching the failure modes your current stack is making invisible.
Related insights
- What I learned from Khan Academy about A/B testing AI
- AI Evals vs. A/B Testing: Why You Need Both to Ship GenAI
- What Is A/B Testing? A Complete Guide
Related Articles
CLI vs MCP: TLDR, it doesn’t matter as much as you think
Your agent should know how to do the job whichever interface it uses.
Your coding agent can already run terminal commands. Your product manager's assistant can connect to remote tools. Both need to inspect the same experiment, understand the same metric, and prepare changes to the same feature flag.
That creates a familiar question: should you give them a CLI or an MCP server?
The usual answers mix several concerns. A CLI is easy to compose with scripts. MCP gives compatible clients a standard way to discover and invoke tools. Skills can supply the instructions that make either route useful. None of those statements tells you whether the agent knows your release process or can distinguish a draft from a live change.
For GrowthBook, the choice has become smaller. We've built the agent layer around a REST API and reusable workflow skills, then made those available through different interfaces. Your choice of CLI or MCP no longer needs to determine which product workflow your agent can understand.
It still determines how the agent connects, executes requests, and handles credentials. This article explains that distinction, shows the same disabled-flag operation through both routes, and gives you a way to evaluate the integration you actually intend to use.
The debate combines three different questions
Start with a specific job: prepare a release control for a new checkout flow. The agent needs to know which project owns checkout, which environments exist, what the default behavior should be, and whether the proposed flag key already exists.
Choosing an interface answers only part of that job.
How can the agent reach the product?
A command-line interface lets an agent invoke a program in an environment where it has shell access. That program can make authenticated API requests and return data. You can also run the same commands yourself, put them in scripts, and preserve output for inspection.
The Model Context Protocol architecture connects a host application to servers through clients. An MCP server exposes capabilities that a compatible host can discover and call. The host might be a coding environment, a desktop assistant, or another agent application.
These categories overlap. A terminal-based agent can also be an MCP client. A local MCP server can be a process launched by that client. A remote MCP connection can work without the host giving the model a general-purpose shell.
So “terminal agent” doesn't automatically mean “use the CLI,” and “MCP” doesn't automatically mean “remote service.” Ask what access the runtime provides before choosing the adapter.
What procedure should the agent follow?
An API can expose a create operation without teaching an agent when to use it. A syntactically valid feature flag might still have the wrong project, an ambiguous key, or a default value that doesn't match the application.
That's the job of a skill: reusable instructions for a recognizable task. The Agent Skills format gives those instructions a portable structure, including a skill entry point and supporting references. A workflow can describe the context to inspect, the sequence to follow, and the evidence to return.
For the checkout example, useful guidance says to resolve environments, check the permanent key, choose a value type, and create the flag disabled. Merely exposing a create command leaves those decisions to the agent's general knowledge.
Skills are also useful when the right answer is to take no action yet. “Create a flag” and “design an experiment” are related requests, but they can require different objects and sequencing. An agent needs enough product guidance to make that distinction.
What context makes the procedure relevant?
Even a good playbook needs the current product state. The agent should inspect the objects in your GrowthBook organization and, when it is changing application code, the repository's SDK initialization and tests.
Those are different sources of context. GrowthBook supplies flags, experiments, metric definitions, and other authorized product resources. Your coding runtime supplies the application files it can access. An MCP connection doesn't by itself grant repository access, and a CLI doesn't by itself explain your metric definitions.
Evaluate the combination: access to the product, a maintained procedure, and enough context to apply that procedure. A comparison that changes all three at once cannot isolate the effect of CLI versus MCP.
GrowthBook puts the durable parts below the interface
We separated the agent layer into responsibilities that can be reused. The product API exposes operations. Skills explain workflows. The interface connects the agent to those two layers.
This is the architectural shift behind GrowthBook's agent experience. Instead of making every client integration carry its own interpretation of experimentation, we can share the interpretation and change how it is delivered.
The API provides the operations
GrowthBook's REST API makes product objects addressable. An agent can inspect a project, retrieve a flag, or perform an authorized operation against those objects.
The rebuilt API described in GrowthBook 4.4 uses shared schemas to support types, validation, and API documentation. The CLI is generated from the API definition. That reduces the amount of separately maintained command logic needed to expose product operations.
The OpenAPI specification explains the underlying principle: a machine-readable description can represent an HTTP API's paths, operations, inputs, and responses. Generation is useful because it derives another interface from that contract.
Generation doesn't mean any installed binary knows every future endpoint. The generated client still has a version, and your GrowthBook server still has a version. A self-hosted instance can lag behind the public documentation. Treat API compatibility as a deployment fact you verify.
Skills provide the workflow guidance
GrowthBook's feature-flags skill routes requests to more specific workflows. Creating a flag, inspecting dependencies, changing targeting, requesting review, and publishing a revision are separate procedures.
That structure lets the agent load the guidance relevant to the current job. The skill authoring guidance on progressive disclosure supports this approach: keep the entry point focused and place detail in references that are loaded as needed.
The benefit is maintainability as well as context management. A correction to a flag workflow can be made in the maintained skill rather than repeated across a series of tool descriptions and client-specific prompts.
There are two shell concepts to keep distinct. The generated growthbook CLI is a terminal interface to the product API. The skills repository also includes a small gb-call helper for making REST requests. Installing a skill doesn't make that helper a globally available command; its location depends on the installation. Both can reach the API, but their configuration and invocation are different.
MCP provides a thin bridge
The current GrowthBook MCP implementation exposes four tools: growthbook_list_skills, growthbook_read_skill, growthbook_api_read, and growthbook_api_write.
An agent discovers a domain, reads the relevant workflow, then makes API calls. It doesn't need a dedicated MCP tool for every product endpoint. Calls described as gb-call GET in a skill map to the read tool; mutating methods map to the write tool with the same path and body.
The bridge does not run the shell helper. It forwards authenticated requests. Product-side permissions and validation remain important because a small tool surface can still reach many operations.
Shared guidance also has a delivery boundary: the MCP server bundles skills when it is built, using a pinned source revision. A newly edited skill in the repository needs to be included in a subsequent server build before that server supplies it. “Shared source” is a maintainability claim, not a promise of instant synchronization.
Explore the shared flag workflows
Open the feature-flags skill to see the procedures that guide creation, targeting, reviews, and publishing. Bring the same workflow knowledge to the agent interface your team uses.
View the Flag Skill
The same disabled flag can use either route
The practical question is whether the interface changes the job. Let's create a Boolean flag called checkout-redesign, with a false default and every environment disabled.
This is an API example for a sandbox project, not an application integration tutorial. It creates a control object; your application still needs a correctly configured GrowthBook SDK to evaluate it. The feature flag documentation covers the product concepts behind that integration.
Connect and verify before preparing the write
For the terminal route, install the current GrowthBook CLI, configure a Personal Access Token through its interactive setup, and inspect the available command syntax:
Check that the active server and identity are the ones you intend to use. Prefer a tested, pinned CLI version in repeatable automation. Don't include a live token in a command you plan to share.
For an MCP route, the GrowthBook Cloud setup uses the hosted endpoint and browser OAuth. In a client using the documented mcpServers configuration shape, such as Cursor, add:
Complete the browser authorization and confirm tool discovery. Other clients can use a different configuration shape. Self-hosted GrowthBook users should follow the local stdio or self-hosted HTTP instructions rather than point a private instance at the Cloud service.
Next, let the connected agent discover the flag-creation workflow. These are tool-call parameters, shown separately in the order to invoke them:
The returned guidance should describe the prerequisites for creating the flag. If the workflow isn't present, resolve the server's bundled skill version before substituting instructions from an older integration.
Resolve real IDs and construct one payload
The flag-creation workflow checks existing keys, lists projects, and resolves environments. In the MCP client, invoke the read tool with each of these argument objects:
For the CLI route, perform that discovery with the corresponding command groups documented in the CLI command catalog:
Inspect each command's help if your installed version differs. The returned keys, project IDs, and environment IDs are the inputs to the payload below. In MCP, the equivalent read-tool arguments are:
Verify that checkout-redesign isn't taken, identify the sandbox project's ID, and enumerate every actual environment. These read responses determine the payload; don't treat example environment names as a complete inventory of your organization.
Save the following as flag.json, replacing <PROJECT_ID> and the environment keys with the discovered values. This example assumes the only environments are staging and production:
The API represents defaultValue as a string, even for a Boolean flag. Explicitly disabling each environment avoids relying on the organization's default state for newly created flags. Use a PAT for this example so the caller has a user identity for attribution.
Review the permanent key, project, value, and disabled states before submitting. If the intended job is an experiment, route to experiment design first instead of creating an unrelated flag and hoping to connect it later.
Submit once, then read the object back
With the CLI, the create command accepts a JSON body. In a Bash-compatible shell, inspect the request with dry run before submitting:
The first command previews a request; it doesn't prove the server will accept it. Run the second only after checking the preview and approving this creation. In PowerShell, use that shell's file-reading syntax rather than paste the Bash substitution unchanged.
The PowerShell equivalent, using the same reviewed file, is:
The same distinction applies: preview first, approve the specific creation, then submit. Neither command changes the application's deployed code.
Through MCP, call growthbook_api_write with the same payload serialized into its body string. Substitute the same resolved project and environment names:
Choose one creation route. Don't run both against the same key; they target the same product object, and the second creation should encounter an existing key.
Verify through the CLI get command:
Or invoke growthbook_api_read with:
Check the returned key, project, value type, default value, disabled state of every environment, and absence of unintended rules. Report those facts rather than “success.” No targeting or exposure change is included in this example.
Choose the route that fits the runtime
Once the procedure and API operation are shared, the remaining choice is operational. Where does the agent run? What can it execute? How does it authenticate? What does your team need to inspect afterward?
Shell access favors terminal composition
A shell-capable agent can run the CLI alongside repository tools, save structured output, and combine product inspection with local checks. That is useful when the job includes code changes, generated files, or a reproducible script.
For example, a flag audit can produce an inventory file while an engineer separately checks code references. The agent should preserve the association between the GrowthBook key and each repository reference. The final artifact can include both the product state and the proposed code diff.
The shell makes composition available; it doesn't make a workflow safe by default. A script can select the wrong profile or swallow an error. Decide which failures stop the workflow, where outputs are stored, and which steps need a reviewable artifact.
Remote tool access favors MCP clients
MCP is useful when the host already provides a tool connection but doesn't expose a terminal. A PM can inspect an experiment through a compatible assistant without installing a product CLI inside that assistant's runtime.
The MCP authorization specification defines an authorization framework for HTTP transports. GrowthBook Cloud uses browser OAuth, while local stdio configurations use credentials supplied to the process. Those are distinct deployment paths.
MCP also gives hosts a structured description of tools. The tool specification includes annotations such as read-only and destructive hints. Hosts can use those signals in their interaction design, but the signals are not an access-control boundary. The product must still enforce the caller's actual permissions.
Use a small decision table
| Your immediate requirement | Route to consider | What to verify |
|---|---|---|
| Combine product operations with local files and scripts | CLI plus maintained skills | Shell access, profile, installed version, structured output |
| Connect a compatible assistant without a shell | Hosted MCP | HTTP/OAuth support, identity, tools, bundled workflow |
| Keep requests inside a self-hosted deployment | CLI or self-hosted MCP | Network reachability, server version, credentials, trusted app origin |
| Reuse a procedure across different team clients | Shared skills through either route | Workflow revision, API compatibility, expected review artifact |
You can support more than one row. An engineer's CLI and a PM's MCP connection can operate on the same objects with different identities and permissions. Standardizing the workflow is often more useful than standardizing everyone's client.
Compare actual outcomes, not tool counts
Four MCP tools can address a large API. A large CLI command catalog can expose the same operations. Counting entries tells you little about whether an agent completed the intended job correctly.
Likewise, there is no universal token or speed result implied by the architecture. The client, model, instructions, response size, pagination, and number of requests all affect a real run. Measure the work you need instead of importing a result from an unrelated integration.
Keep the evaluation task constant
Use one sandbox project and one bounded task. For a read test, ask each route to find a named flag and identify its enabled environments, current revision, and relevant dependencies. For a write test, use separate temporary keys with the same reviewed disabled-flag requirements.
Give both runs the same expected output. Record whether each identified the correct objects, loaded the relevant workflow, handled missing data, and preserved the boundary between creation and exposure. Measure elapsed time and context consumption only alongside those correctness checks.
If one run gets rich workflow guidance and the other gets only a command name, you're testing the guidance as well as the interface. That can still be a useful product evaluation, but describe what changed.
Test failure behavior deliberately
Try a nonexistent object, an existing key, and a credential without the required permission. The agent should identify the failure and stop the dependent action. It should not invent an ID, silently pick another environment, or reinterpret a permission error as permission to retry with a broader credential.
The revision workflow adds another important test: multiple open drafts. When a chain of changes belongs in one revision, the agent must preserve its version rather than repeatedly ask for whichever draft is latest.
A creation receipt and a publication receipt are different artifacts. If your evaluation only checks that a request returned without error, it can miss the point where a proposed change became live or where a draft failed to include the intended edits.
Record the versions behind the result
Capture the CLI version, server version when available, client configuration, and the skill text or revision used. With MCP, inspect the guidance actually returned by the connected server. With installed skills, inspect the local instructions actually loaded.
That record makes a later regression diagnosable. A changed result may come from new guidance, a client update, a changed API, or different product state. “MCP became worse” is too broad to help you fix it.
Build workflows that survive the next client
The most useful consequence of GrowthBook's design is organizational. Engineers, PMs, and analysts can choose the agent surface that fits their work while referring to the same flags, experiments, and metrics.
An engineer can prepare a disabled flag from a terminal. A reviewer can inspect its draft in the app. An assistant connected through MCP can retrieve the same object and explain its state. The handoff should carry an ID, a revision where applicable, the proposed behavior, and the decision still required.
This is where the CLI-versus-MCP question loses much of its importance. Your release procedure doesn't need to be rewritten around each client. The product holds state; the API exposes operations; skills explain the work; the chosen interface delivers access.
Start with one inspection or sandbox creation. Require the agent to return a concrete object and its verified state. Then add the next workflow only when the current one is reliable enough for your team to review.
GrowthBook's feature flag platform gives that work a common place to live. The durable choice is how you connect release control and measurement to your team's process. CLI and MCP are ways to carry that process into the tools people already use.
Start with a common foundation
Set up GrowthBook for the flags, experiments, and metrics your agents will share. Then choose the CLI or MCP connection that fits each team member's work.
Get Started With GrowthBook
Standard deviation vs variance: Key differences for A/B testers
A metric can have wide user-to-user variation and still yield a precise experiment estimate.
That is why confusing standard deviation with standard error creates bad A/B testing decisions. A dashboard may show that users differ substantially while the average effect is estimated closely enough to guide a release. Those statements can both be true.
Variance and standard deviation describe spread. Variance works on a squared scale; standard deviation takes its square root to return to the metric's original units. Neither is automatically the uncertainty around an experiment result.
For A/B testers, the useful distinction is practical: use standard deviation to communicate outcome variability, use variance in the calculations that combine uncertainty, and inspect the standard error or interval for the treatment effect. Start by making sure all three describe the correct unit of analysis.
Understand the difference in units
Variance uses squared distances
Outcomes above and below a mean have deviations with opposite signs. Simply averaging those signed deviations gives zero. Variance avoids that cancellation by squaring each deviation before combining them.
The sample variance formula is s² = Σ(xᵢ − x̄)² / (n − 1), where x̄ is the sample mean and n is the number of observations. The symbol Σ means to add the terms across observations.
If the outcome is measured in seconds, variance is measured in seconds squared. That is mathematically useful but less intuitive for explaining a product experience. “Variance is 4 seconds squared” does not mean users differ by 4 seconds.
Standard deviation returns to the original scale
Standard deviation is s = √s². A variance of 4 seconds squared corresponds to a standard deviation of 2 seconds. Both describe the same dispersion, expressed on different scales.
Do not define standard deviation as the average absolute distance from the mean. That is a different measure. Squaring deviations gives large deviations more influence, and taking the square root afterward does not undo that weighting.
| Quantity | What it describes | Units for task time |
|---|---|---|
| Variance | Squared dispersion in outcomes | Seconds squared |
| Standard deviation | Outcome dispersion on the original scale | Seconds |
| Standard error of the mean | Sampling variability of an estimated mean | Seconds |
Neither variance nor standard deviation requires a normally distributed population merely to be calculated. Normality matters for particular interpretations and inferential procedures, not for the existence of these summaries when the required moments are finite.
Calculate both from the same observations
Work through a small dataset
Use four synthetic task times: 2, 4, 4, and 6 seconds. This tiny dataset is for arithmetic, not evidence about a real product or a basis for confident inference.
The mean is (2 + 4 + 4 + 6) / 4 = 4 seconds. Deviations from the mean are −2, 0, 0, and 2 seconds. Their squares are 4, 0, 0, and 4 seconds squared, totaling 8.
Treating these observations as a sample, the usual sample variance is 8 / (4 − 1) = 8/3, or about 2.667 seconds squared. The sample standard deviation is √(8/3), or about 1.633 seconds.
This is a useful check on a report. If someone describes the same sample's variance as 1.633, they have probably substituted the standard deviation. If the units are missing, the mistake is much harder to spot.
Choose sample or population formulas deliberately
If those four observations constitute the entire finite population you want to describe, population variance divides by 4: 8/4 = 2 seconds squared. Population standard deviation is √2, or about 1.414 seconds.
The population and sample distinction depends on what you are estimating. Having every recorded event from an experiment does not necessarily make those events the full population of future users or future conditions you care about.
The n − 1 denominator makes the usual sample variance an unbiased estimator of population variance under independent, identically distributed sampling with finite variance. It does not make the square root exactly unbiased for population standard deviation. Nor does it correct biased sampling or dependence between observations.
Verify the calculation with explicit function names
Python's statistics functions distinguish sample calculations from population calculations. With Python 3 installed, this standalone example requires no external packages or data:
The first line of output is 2.667 1.633; the second is 2 1.414. The different denominators explain the difference. Before comparing statistics from two tools, check their sample/population defaults and handling of missing values.
Find the noise you can reduce
Explore practical variance reduction methods, including metric design and CUPED, and the conditions that make each approach useful.
Read the Variance Reduction Guide
Use standard errors for experiment uncertainty
Distinguish users from estimated means
Standard deviation describes how users' outcomes vary. Standard error describes how an estimator would vary across repeated samples. For an independent sample, the estimated standard error of its mean is s / √n.
Suppose a user-level task-time metric has a sample standard deviation of 20 seconds among 10,000 independent users. The estimated standard error of the mean is 20/√10,000 = 0.2 seconds. Large differences between users coexist with a much smaller uncertainty around their average.
Collecting four times as many independent users would halve this standard error if the outcome variance stayed similar. It would not mechanically halve user-level standard deviation. The individual experience remains variable even as the estimated mean becomes more precise.
Combine arm variances before taking the square root
For independent treatment and control groups, a simple unadjusted difference in means has estimated variance s²ₜ/nₜ + s²꜀/n꜀. Its standard error is the square root of that sum. You do not add the two arm standard deviations and call the result the experiment's uncertainty.
Take a separate synthetic summary with 10,000 independent users in each arm and a 20-second standard deviation in both. The variance of the estimated difference is 400/10,000 + 400/10,000 = 0.08 seconds squared. Its standard error is √0.08, about 0.283 seconds.
If the observed treatment-minus-control difference is −1 second, an illustrative large-sample, unadjusted 95% normal interval is −1 ± 1.96 × 0.283, approximately −1.554 to −0.446 seconds. Negative values mean faster completion under this metric definition.
This example assumes independent groups and an appropriate large-sample approximation. It is not the exact output of GrowthBook's relative-lift, sequential, covariate-adjusted, or ratio-metric methods. GrowthBook's statistical details describe how the platform estimates effects and their uncertainty.
Avoid interpreting intervals as outcome ranges
That interval concerns the mean treatment effect. It does not say that 95% of individual users improved by between 0.446 and 1.554 seconds. An individual user's causal effect is a different quantity, and the experiment does not observe both potential outcomes for the same person.
Likewise, mean ± one standard deviation is not a universal statement about where 68% of users fall. That familiar interpretation depends on an approximately normal distribution. Skewed task times or revenue can look very different.
The frequentist confidence level describes the repeated-sampling coverage of the interval procedure under its assumptions. It is not a posterior probability assigned to a fixed realized interval.
Check what your metric actually measures
Match the variance to the randomization unit
If users are randomized and each contributes many requests, request rows are not automatically independent experimental units. Treating 100 requests from one person as 100 unrelated users can make uncertainty look smaller than it should.
Define the outcome at the intended unit or use a method that accounts for clustering. GrowthBook's aggregate data requirements are relevant when supplying precomputed summaries: the sample size, sum, and dispersion must correspond to the statistic and analysis method.
Two teams can report different standard deviations without either making an arithmetic mistake. One may summarize per-request latency while the other summarizes each user's average latency. They are describing different outcome distributions. Clarify the grain before comparing their numbers.
Paired measurements need another distinction. If the same users provide before-and-after outcomes, the measurements within each pair can be correlated. Analyze each user's difference or use an appropriate paired method; do not apply the independent-arm formula automatically. The paired-means framework treats those differences as the observations of interest. Pairing also does not remove time trends or establish causality by itself; the design still determines which comparison is defensible.
Inspect tails instead of deleting them casually
Large observations influence both variance and standard deviation. A few unusually long tasks can dominate a time metric; a few large purchasers can dominate revenue. Determine whether those values reflect valid behavior, instrumentation defects, retries, or unit errors.
A valid expensive user is not automatically an outlier to discard. Removing high values after seeing an experiment result can change the estimand and introduce selective analysis. Decide justified exclusions or transformations before interpreting the test.
Report complementary summaries when they answer useful questions: a histogram, median, or a relevant tail percentile can show what a single spread measure hides. Those summaries do not replace the mean effect if the mean remains the decision metric.
Keep units consistent across pipelines
Converting seconds to milliseconds multiplies every value by 1,000. Standard deviation also multiplies by 1,000, but variance multiplies by 1,000,000. A dramatic numerical variance increase can therefore be a unit conversion rather than a product regression.
For the four-value example, sample standard deviation becomes approximately 1,633 milliseconds and variance becomes approximately 2,666,667 milliseconds squared. The underlying task times have not changed. Store units with metric definitions and inspect them when comparing dashboards.
A lower variance caused solely by rescaling a metric does not create more statistical information. The effect size rescales too. Separate a numerical change in units from an actual improvement in the precision of your estimate.
Also avoid averaging standard deviations from daily reports and labeling the result a weekly standard deviation. The complete-period spread depends on both variation within days and differences between daily means. Keep the underlying observations or sufficient summaries that support the intended calculation. A simple average of displayed daily values generally loses information needed to reconstruct the full distribution.
Turn spread into a better testing plan
Define the outcome before planning the sample
Start with the metric definition: population, identity, aggregation, window, units, exclusions, and desired effect. Estimate variability from representative historical data at that same grain.
Historical variance is an input, not a guarantee. New traffic, a different user mix, or a changed measurement window can alter it. Explore plausible variance and effect-size assumptions rather than treating one historical number as exact.
Keep the two arm variances separate when appropriate. A treatment can change variability as well as its mean. The simple equal-standard-deviation illustration above is a teaching assumption, not a requirement that production groups have identical spread.
Reduce estimation noise without hiding behavior
CUPED and post-stratification use suitable pre-experiment information and covariates to improve effect estimation. CUPED stands for Controlled-experiment Using Pre-Experiment Data. Its value depends on useful predictive information and a valid implementation.
Adjustment can reduce uncertainty in the treatment effect without making real users' outcomes less variable. That distinction prevents a reporting error: a narrower adjusted interval is not evidence that customers now experience a more consistent product.
Also separate missing data from genuine zeros. A user with no purchase may correctly contribute zero revenue; a missing instrumentation event might represent an unknown outcome. Treating both identically can distort the mean and its variance before any statistical engine sees the data.
Report the right quantity for the decision
| Reader's question | Useful quantity |
|---|---|
| How much do individual outcomes differ? | Standard deviation, with distribution context |
| What enters this uncertainty calculation? | Variance at the correct analysis unit |
| How precisely is the effect estimated? | Standard error and an appropriate interval |
| Is the measured difference worth acting on? | Effect size, uncertainty, guardrails, and decision context |
For a clear experiment readout, show the metric's unit and analysis grain, the estimated effect, and its appropriate uncertainty interval. Add outcome spread when it helps readers understand the population or diagnose a noisy metric.
GrowthBook's experimentation platform brings defined metrics and statistical analysis together. A trustworthy interpretation still starts with your measurement choices: describe spread with standard deviation, calculate with variance where required, and judge precision using the estimator's uncertainty.
Build a clearer experiment readout
Connect your metrics to GrowthBook and evaluate effects with the analysis method and measurement definitions your team has chosen.
Get Started With Experimentation
Catching 80 defects means little until you know how many defects existed.
Suppose a detector flags 125 suspicious software changes. Engineers inspect them and confirm 80 real defects. That tells you something about alert quality, but it does not tell you how many defects the detector missed. To calculate true positive rate, you need the actual positives outside the alert queue too.
True positive rate, or TPR, measures the share of actual positive cases that a system correctly identifies. It is also called recall or sensitivity. The formula is TP / (TP + FN), where TP means true positives and FN means false negatives.
The arithmetic is simple. Choosing the right population, labels, threshold, and denominator is the difficult part. This guide works through an engineering example, distinguishes neighboring metrics, and explains what changes when the “test” is an A/B experiment rather than an item classifier.
Start with actual positives, not alerts
Define the positive class
“Positive” means the condition you want to detect. It does not necessarily mean a good outcome. In a defect detector, positive means a real defect. In a spam filter, it means spam. Write that definition before calculating anything.
A true positive is an item that is actually positive and predicted positive. A false negative is actually positive but predicted negative. Those two groups together form the denominator: all actual positives in the evaluation sample.
The recall definition is therefore conditional on actual class membership. TPR asks, “Among the real defects, what proportion did we catch?” It does not ask, “Among the alerts, what proportion were real?”
Fix the unit and observation window
Choose a consistent evaluation unit. A code change, an individual finding, and a repository are different units. If one change generates 12 duplicate alerts, counting alerts in the numerator and changes in the denominator produces a meaningless rate.
For the example, each unit is one code change, evaluated once by one fixed detector version at one fixed threshold. An independent review process eventually labels each change as defective or not defective. The observation window is long enough for those labels to mature.
This resembles the discipline needed when defining experiment metrics: the event, identity, time window, and denominator must describe the same question. A correct formula applied to incompatible counts still gives the wrong answer.
Calculate TPR from the confusion matrix
Put every evaluated item in one cell
Consider an original synthetic evaluation of 1,000 code changes. Reviewers establish that 100 contain a defect and 900 do not. The detector catches 80 defects, misses 20, and incorrectly flags 45 nondefective changes.
| Actual label | Predicted positive | Predicted negative | Total |
|---|---|---|---|
| Defective | 80 true positives | 20 false negatives | 100 |
| Not defective | 45 false positives | 855 true negatives | 900 |
| Total | 125 | 875 | 1,000 |
Read the labels rather than relying on cell position. Libraries and reports can arrange axes differently. The scikit-learn confusion matrix convention uses actual classes on rows and predicted classes on columns, with class order also affecting where each value appears.
Check the totals before calculating rates. The four cells sum to 1,000. True positives plus false negatives sum to 100 actual defects. Predicted positives total 125, which is a different denominator.
Divide detected positives by all actual positives
The calculation is:
TPR = TP / (TP + FN) = 80 / (80 + 20) = 80 / 100 = 0.80 = 80%.
Interpret it in words: the detector caught 80% of the actual defects in this evaluation sample. It missed 20%, so the false negative rate is 20%. For a binary task with complete labels, TPR and false negative rate sum to 100%.
Do not divide 80 by 1,000. That would measure true positives as a share of all evaluated changes, not recall. Do not divide 80 by 125 either. That is precision, which answers the alert-quality question.
The result describes this sample and threshold. It does not establish that every future workload will have the same recall, or that all defect categories are equally detectable. Keep the counts beside the percentage so readers can see the support behind it.
Handle zero positive support explicitly
If the sample contains no actual defects, TP + FN equals zero. Sample TPR is undefined. Returning 0% would imply the detector missed known defects; returning 100% would imply it successfully found them. Neither interpretation is supported.
Software libraries need a practical zero-division convention. That convention can return a placeholder or warning, but it does not create evidence. Report “TPR unavailable: zero actual positives” and preserve the support count. A pipeline can then distinguish an unevaluable slice from measured poor performance.
Check the experiment error risks
See how power, repeated checks, and multiple comparisons affect the reliability of experimental decisions beyond a single classifier score.
Read the False Positives GuideCompare TPR with precision and false positives
Keep the denominators visible
For the same 1,000 changes, the other common metrics tell different stories:
| Metric | Calculation | Result |
|---|---|---|
| True positive rate / recall | 80 / 100 actual positives | 80% |
| Precision | 80 / 125 predicted positives | 64% |
| False positive rate | 45 / 900 actual negatives | 5% |
| Specificity | 855 / 900 actual negatives | 95% |
| Accuracy | (80 + 855) / 1,000 changes | 93.5% |
The classification metric definitions differ because they condition on different populations. Here, 36% of alerts are false, but the false positive rate is 5%. Calling both numbers “the false positive rate” hides the operational distinction.
Accuracy can also look reassuring in an imbalanced population. A detector that labels all 1,000 changes nondefective would be 90% accurate while detecting none of the 100 defects. TPR exposes that failure, though it still needs the false-positive side of the picture.
Choose a threshold for the decision
Suppose a stricter threshold, evaluated on the same labeled changes, catches 60 defects and incorrectly flags 10 nondefective changes. Its TPR is 60%, its precision is about 85.7%, and its false positive rate is about 1.11%.
This threshold reduces the review queue from 125 changes to 70. It also misses 40 defects instead of 20. Whether that tradeoff is acceptable depends on the consequences of a missed defect, the review team's capacity, and what action follows an alert.
A ROC curve plots TPR against false positive rate across thresholds. It helps compare operating points. It does not pick the operational threshold for you, and an area-under-curve summary does not specify the queue size at the threshold you intend to deploy.
If you tune thresholds using the same labels you later report as a final evaluation, the score can be optimistic. Use separate validation and held-out evaluation data, or another defensible evaluation design, and freeze the decision rule before the final assessment.
Specify the action at each threshold too. A low-confidence finding might create a suggestion, while a high-confidence finding might block a merge. Those are different interventions with different error costs. One aggregate recall figure should not imply that every detected item received the same treatment.
Make sure the labels can support the claim
Audit predicted negatives too
If engineers review only alerts, you can estimate precision within the reviewed alert population. You cannot directly count missed defects among unreviewed negatives. Assuming those changes are clean makes FN artificially small and TPR artificially high.
An evaluation may need independent review of a representative sample of predicted negatives, delayed incident labels, or a curated benchmark with known defects. Each approach has limits. Incident labels miss undetected failures; curated examples can differ from ordinary production work.
If the evaluation oversamples suspicious changes or certain repositories, report that design. Use appropriate weighting when estimating performance for the real workload. A simple unweighted confusion matrix describes the sampled dataset, which may not represent the deployed population.
Preserve uncertainty and segment support
An 80% estimate based on 4 detected defects out of 5 is much less precise than one based on 800 out of 1,000. The important sample size for recall is the count of actual positives, not just the total number of changes evaluated.
Report positive support by important slices, such as language, defect category, or repository type. A high pooled TPR can hide weak performance on a rare but expensive failure mode. At the same time, avoid declaring tiny slices definitively better or worse from a few examples.
Counts alone do not make observations independent. Many similar changes from one repository can share a common pattern. Use uncertainty methods that fit the evaluation design rather than assuming every row is an unrelated trial.
Separate label delay from label absence
A recently deployed change might appear nondefective because nobody has had time to find the bug. Treating it as a confirmed negative gives the newest cohort a different labeling standard from older cohorts.
Define when an item becomes eligible for evaluation and track unresolved labels separately. Keep model version, threshold, label version, and observation dates with the result. That makes a later score change interpretable: you can tell whether the detector changed, the workload changed, or the labels matured.
For a warehouse-native measurement workflow, these definitions belong in the underlying data model. Joining model decisions to review outcomes still requires identity and time alignment; a queryable warehouse does not automatically resolve ambiguous ground truth.
Keep comparable populations together
When comparing detector versions, evaluate them on the same held-out cases where possible. Otherwise, an apparent recall improvement could come from an easier mix of defects. Preserve a breakdown of new, removed, and unchanged cases so reviewers can see what changed between evaluations.
When combining reports, add the underlying TP and FN counts for nonoverlapping, consistently defined populations, then calculate the pooled rate. An unweighted average of repository percentages answers a different question because it gives a tiny repository the same influence as a large one. Either summary can be useful, but label the weighting explicitly.
If the same change appears in several reports, deduplicate or model that repeated observation before pooling. Repeated scans do not create additional independent defects. Keep an audit trail from the final counts back to the evaluated item identifiers so a suspicious rate can be investigated.
Connect detection quality to product decisions
Distinguish classifier recall from statistical power
In hypothesis testing, statistical power is the probability that a testing procedure rejects the null when a specified alternative is true. It is a repeated-test detection probability, depending on assumptions such as effect size, variance, sample size, and decision rule.
That is related to the idea of detecting a real signal, but it is not the same dataset as a labeled-item confusion matrix. In an ordinary A/B program, you do not observe the true causal effect of every experiment with certainty. Dividing declared winners by all experiments does not reveal a ground-truth TPR.
Power planning asks how often your procedure would detect a particular real effect under its assumptions. A classifier evaluation asks how many labeled positives were detected in the evaluated sample. Keep those questions separate when discussing GrowthBook's statistical methods.
Evaluate the intervention as well as the detector
An improved offline detector can still create a worse product workflow. More alerts might overwhelm engineers, delay releases, or catch defects too late to matter. Conversely, a modest recall change might be valuable if it catches the failures with the highest operational cost.
When appropriate, use randomized experiments to evaluate the downstream intervention: which review experience helps teams resolve defects, with what delay and cost? Keep detector recall, review workload, and product outcomes as separate measures.
GrowthBook's experimentation platform can analyze defined outcomes from such tests. Your team supplies reliable assignment, exposure, labels, and metric definitions. Neither a high TPR nor a significant product result eliminates the need to inspect the other evidence.
Use a repeatable reporting checklist
Before sharing a TPR number, state the positive class, evaluation unit, detector version, threshold, and label process. Show TP and FN beside the rate. Include precision or false positive rate so readers can assess the workload created by detection.
Then ask whether predicted negatives were evaluated, whether labels had time to mature, and whether the sample reflects the population where you will act. Mark zero-support slices unavailable and explain material gaps in coverage.
The useful result is a statement someone can audit: “At this threshold, the detector found 80 of 100 labeled defects in these 1,000 evaluated changes, while producing 45 false alerts.” That is a clearer basis for a decision than “our model is 80% accurate.”
Measure what the change delivers
Connect your data and defined metrics to GrowthBook to evaluate how a product or workflow change affects the outcomes your team cares about.
Start Measuring With GrowthBookReady to ship faster?
No credit card required. Start with feature flags, experimentation, and product analytics—free.
