Experiments

Canary releases for AI models: what changes vs traditional software

May 29, 2026

min read

A graphic of a bar chart with an arrow pointing upward.

A hallucinating language model returns HTTP 200.

A model that's drifted off-distribution responds to every request — it just responds worse. If your canary release monitoring stops at error rates and latency, you have no instrumentation for either of those failures. That's the core problem this article addresses.

Traditional canary releases were built to catch crashes, 5xx errors, and latency spikes. Those are binary failures — they either happen or they don't. AI model degradation doesn't work that way.

Output quality can quietly regress while every infrastructure metric stays green. Fixing that requires a different monitoring layer, different traffic routing logic, and different rollback triggers — not just more dashboards.

This article is for engineers, PMs, and data teams who are shipping AI models to production and want their canary process to actually catch the failures that matter. Here's what you'll learn:

Why standard canary monitoring creates a false sense of safety for AI models — and what the signal gap actually looks like
Which output quality metrics to track during a canary, and how to make rollback decisions defensible
How to structure traffic routing and hold periods around statistical requirements, not just baking time
How to design rollback triggers that fire on quality degradation, not just error thresholds
How to treat the canary release as a controlled experiment so every rollout generates evidence, not just a pass/fail result

Each section builds on the last, moving from the conceptual gap to the specific instrumentation, routing, and rollback decisions you'll need to make before your next model release goes live.

Why standard canary release logic breaks down for AI models

If you've shipped software with canary deployments before, the pattern feels reliable: send a small slice of traffic to the new version, watch the dashboards, and either promote or roll back based on what you see. For deterministic software, this works well.

For AI models, it creates a dangerous illusion of safety — you can watch every metric stay green while your model silently degrades in ways that matter enormously to users.

How traditional canary releases work — and what they're designed to catch

The mechanics of a traditional canary release are straightforward. Route somewhere between 1% and 5% of production traffic to the new version, monitor infrastructure signals over a validation window, and either expand the rollout or revert.

Google's SRE practice captures the intent precisely: canary servers exist to "detect dangerous effects from the behavior of the new software under real user traffic."

That definition sounds comprehensive. The problem is what it implicitly assumes about what "dangerous effects" look like. Traditional canaries are optimized for binary failure detection — crashes, HTTP 5xx errors, latency spikes, memory exhaustion, throughput collapse.

These are events that either happen or don't. A bug either causes a measurable failure or it doesn't. The signal is immediate and unambiguous.

This works because deterministic software has a predictable failure signature. When something goes wrong, the infrastructure tells you.

The signal layer traditional canaries rely on

The specific metrics that feed traditional canary dashboards reflect this binary assumption: HTTP error rates, p95 and p99 latency, CPU and memory utilization, crash rates, and request throughput. Each of these is a proxy for a single underlying question: is the system still running correctly?

For infrastructure regressions, these signals are well-calibrated. A bad deployment that introduces a memory leak or a broken API endpoint will show up quickly in these numbers. The monitoring layer and the failure mode are matched.

The structural problem is that these signals say nothing about what the system is producing. They measure the health of the pipe, not the quality of what flows through it.

Where those signals fail for AI models — concrete examples

Consider what a degraded AI model actually looks like at the infrastructure layer. A language model that begins hallucinating returns HTTP 200. Systematic bias in completions for certain demographic inputs runs at normal latency with normal throughput — nothing in the infrastructure layer registers it.

A code generation model whose suggestions compile but contain logical errors passes every uptime check. Drift off-distribution produces the same response pattern as a healthy model; the outputs are just worse.

None of these failure modes produce an error event. None of them spike a latency metric. From the perspective of your canary monitoring dashboard, the deployment looks healthy.

NVIDIA has acknowledged this gap directly in the context of AI model monitoring: "Currently, these detection capabilities for AI models are identical to those used for monitoring any other sensitive data — no detection capability focuses on the unique nature of AI/ML." The default tooling treats an AI model like any other software artifact, which means it's instrumented to catch the wrong class of failures entirely.

It's worth noting that even well-engineered traditional canary systems can miss non-binary failures in deterministic software — the Google 2019 outage is a documented case where canaries didn't catch a configuration side effect before it cascaded.

If that's possible with conventional software, the problem is compounded significantly when the failure mode is probabilistic and gradual rather than discrete.

The conceptual shift required — from infrastructure health to output quality

The gap isn't a tooling deficiency you can patch by adding more infrastructure metrics. It's a category error. Traditional canaries answer the question "is the system healthy?" AI canaries have to answer a different question: "are the outputs good?"

These require different instrumentation entirely. Output quality metrics, user engagement signals, task completion rates, drift metrics, and guardrail violation rates are not derivable from CPU utilization and error logs.

They require a second monitoring layer, sitting above infrastructure observability, purpose-built for evaluating what the model is actually producing and how users are responding to it.

Landon Smith, Head of Post-Training at Character.AI, describes this operational reality 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 phrase — from the perspective of our users — identifies exactly the signal layer that traditional canary monitoring cannot provide.

Platforms like GrowthBook are built around this gap, connecting model performance directly to user outcomes rather than infrastructure metrics, and enabling rollback decisions based on quality signals rather than error thresholds. The subsequent sections of this article address how to instrument that signal layer, structure the exposure window, and design rollback triggers that can actually detect the failures that matter.

The metrics that actually matter: monitoring AI output quality during a canary release

When Anthropic's Claude Code team identified a quality regression in early 2025, they confirmed it publicly: "We fixed a Claude Code harness issue that was introduced on 1/26. This was rolled back on 1/28 as soon as we found it." Two days between introduction and detection — at a company with substantial ML infrastructure investment.

The issue wasn't caught by error rates or uptime monitors. It was caught by daily benchmarks specifically instrumented to track output quality.

That gap between "infrastructure healthy" and "model actually degrading" is exactly the problem that output quality metrics exist to close. A model can return HTTP 200 on every request while producing responses that are subtly less accurate, more verbose, increasingly off-distribution, or simply less useful to users.

Infrastructure telemetry has no surface area for any of that. If your canary release monitoring stops at error rates and p99 latency, you are operationally blind to the failure modes that matter most for AI systems. What follows is the metric framework that closes that gap.

Inference latency and throughput — the baseline you still need

Latency and throughput remain necessary, but they're the floor of your monitoring stack, not the ceiling. For generative models specifically, median latency is a misleading signal — output length varies dramatically by request, which means tail latency (p95, p99) is far more meaningful than p50.

A model that handles most requests quickly but occasionally hangs on long-form generation will look fine at the median while degrading user experience at the tail. Track both, treat latency as a guardrail metric, and don't mistake it for a quality signal.

Accuracy and drift signals — detecting what error rates miss

This is the core instrumentation gap. Task-specific accuracy scores, output distribution drift, and hallucination rate proxies all require deliberate instrumentation that sits entirely outside standard observability tooling.

Output distribution drift is particularly easy to overlook: are responses shifting in average length, topic coverage, or tone in ways that correlate with the model version change?

A new model checkpoint might produce responses that are technically coherent but systematically shorter, more hedged, or less complete than the previous version — none of which triggers an alert in your infrastructure layer. The Claude Code incident is instructive here precisely because Anthropic is not a resource-constrained team.

Even with significant investment in ML infrastructure, detecting a quality regression required dedicated daily benchmarks. Teams that haven't built equivalent instrumentation are running canary releases for AI models without the sensors needed to detect the most common failure mode.

User-outcome metrics — where model benchmarks and business reality diverge

Model-level metrics tell you how the model is behaving. User-outcome metrics tell you whether that behavior is actually serving users better. Both layers are required, and neither substitutes for the other — a model can score well on internal accuracy benchmarks while producing outputs users find less useful in practice.

The metrics that operationalize this user-outcome perspective include task completion rate, engagement depth (session length, follow-up query rate, retry rate), and downstream conversion or action rates. Warehouse-native experiment platforms address this layer directly — connecting experiment assignments to behavioral data stored in the data warehouse to surface whether a model change is actually moving user outcomes, not just model benchmarks.

Token efficiency and cost signals

Token cost is both a quality proxy and a direct business metric, and it belongs in your canary monitoring stack for both reasons. Verbose or repetitive outputs inflate token usage without improving user outcomes — so a rising cost-per-successful-task-completion signal often indicates output quality degradation before accuracy metrics catch it.

Raw token count is the wrong unit; cost per successful task completion is the right one.

The right framing treats cost efficiency and quality as jointly optimized, not in tension — a rising cost-per-task signal is evidence of quality degradation, not just a budget concern.

Why rollback decisions require reproducible analysis, not black-box dashboards

The instrumentation architecture matters as much as the metric selection. During a canary, rollback decisions need to be defensible — which means the analysis behind them needs to be reproducible and queryable, not locked in a black-box analytics layer.

Warehouse-native architectures address this directly. GrowthBook runs experiment analysis directly in the customer's data warehouse — Snowflake, BigQuery, Redshift, Databricks, and others — with only aggregates leaving the warehouse. No separate event pipeline, no data duplication, and full reproducibility of every calculation.

That last point matters specifically for canary rollback decisions: when you're deciding whether to halt a model rollout based on a quality metric trend, you need to be able to confirm the numbers independently.

Equally important is the ability to add metrics after a canary begins. Teams frequently discover mid-rollout that they need a signal they didn't anticipate. Merritt Aho, Digital Analytics Lead at Breeze Airways, put it plainly: "Being able to spin up new metrics mid-experiment is a game changer. This was simply never possible before."

For AI canaries specifically, where the failure modes are less predictable than traditional software releases, that flexibility isn't a convenience — it's a requirement.

Incremental traffic routing: structuring the canary exposure window for AI models

The mechanical pattern of traffic splitting in an AI canary release looks identical to what you'd do for any other deployment: route a fraction of requests to the new version, observe it, then decide whether to promote or roll back. The infrastructure primitives are the same.

What's categorically different is how long you need to hold each stage, how you determine when it's safe to advance, and why the sample size requirements for generative AI outputs make traditional "baking period" logic dangerously insufficient.

Initial traffic split mechanics

AWS SageMaker's canary traffic shifting documentation offers a reasonable infrastructure baseline: provision a separate green fleet, route up to 25% of traffic to it initially, and monitor CloudWatch alarms during a baking period. For traditional ML inference endpoints serving deterministic outputs, this is a workable starting point.

For generative AI models, the initial split should typically be considerably lower — the 1–5% range is common practitioner convention — and the reasoning is different from what you'd expect. The concern isn't primarily infrastructure risk.

It's that output quality signals need time to accumulate, and you want to limit user exposure to potentially degraded generative outputs before you have enough signal to evaluate whether degradation is actually occurring. Starting at 25% with a generative model means you've already exposed a quarter of your user base before your quality metrics have reached any meaningful statistical resolution.

The other critical mechanic at this stage is consistent user assignment. A user who hits the canary model on their first request needs to hit it on their second and third requests as well — otherwise your output quality measurements reflect a mix of model versions per user, which corrupts your signal.

Deterministic hashing, where each user is assigned a stable value used to route them to a specific serving path, is the correct implementation pattern here. GrowthBook's SDK uses MurmurHash3 for deterministic hashing, assigning each user a value between 0 and 1 and routing based on configurable ranges, which ensures clean, stable assignment throughout the canary window.

Scaling increments and hold periods

Traditional canary logic is primarily time-gated: hold for N minutes, check whether any alarms fired, promote if clean. This works when the failure mode you're detecting is a crash or a 5xx error — those signals are strong, fast, and unambiguous.

Detecting a 3% regression in output quality scores for a generative model is a fundamentally different statistical problem. The signal is weak, noisy, and slow to accumulate. Advancing the canary on a time-based schedule alone — without validating that your quality metrics have reached statistical significance and show no degradation trend — means you're making a promotion decision on insufficient evidence.

The correct pattern is validation-gated scaling: each increment (say, 1% → 5% → 10% → 25% → 50% → 100%) requires the previous stage's quality metrics to clear defined thresholds before the next increment begins. Elapsed time is a floor, not a ceiling. You don't advance because the clock ran out; you advance because the data cleared the bar.

Compute isolation requirements

AI model serving infrastructure introduces cross-contamination risks that don't exist for stateless application code. If your canary model and your baseline model share the same GPU pool, the same memory cache, or the same request batching queue, they can interfere with each other — a heavily loaded canary endpoint can slow down baseline responses, or a shared memory cache can serve outputs from the wrong model version.

Either scenario corrupts your quality measurements in ways that are hard to diagnose after the fact.

The AWS SageMaker blue/green architecture — where canary and baseline are provisioned as entirely separate fleets — provides the right infrastructure template. Both fleets receive traffic simultaneously during the baking period, but they are isolated at the resource level.

The assignment layer needs to match: canary and baseline users should be routed to separate inference endpoints, not handled as a weighted routing rule on a shared endpoint.

Sample size considerations for generative AI outputs

The asymmetry here is stark. Detecting a server error requires a sample size of one. Detecting a meaningful quality regression in generative outputs may require thousands of interactions, depending on the baseline variance of your quality metrics and the effect size you're trying to detect.

The minimum number of interactions you need to observe should be calculated before the canary starts — based on how large a quality difference you're trying to detect and how noisy your quality metrics are — not set as a fixed time window. Running a canary for 48 hours regardless of how much data you've collected is not a statistical method; it's a schedule.

Running the canary for 48 hours at 1% traffic on a low-volume product might yield 200 interactions, which is almost certainly insufficient to distinguish a real quality regression from noise in a generative model's output distribution.

One practical tool for managing this is sequential testing, which allows valid interim looks at accumulating data without inflating false positive rates. GrowthBook's API exposes a sequentialTestingEnabled parameter that supports exactly this pattern — continuous monitoring during the hold period without the statistical penalty of repeated significance testing.

Combined with warehouse-native analysis that can aggregate quality signals directly from Snowflake, BigQuery, Databricks, and others without requiring a separate event pipeline, this architecture makes it operationally feasible to hold canary stages long enough to accumulate the sample sizes generative AI evaluation actually requires.

Character.AI's use of GrowthBook for incremental model exposure — described by Head of Post-Training Landon Smith as enabling comparison of "different modeling techniques from the perspective of our users" — illustrates what this looks like at production scale: not just a deployment safety mechanism, but a structured process for accumulating user-outcome evidence before committing to a new model version.

The through-line across all four of these considerations is the same: the traffic routing mechanics are familiar, but the exposure window must be designed around the statistical requirements of output quality detection, not the infrastructure requirements of crash detection. Those are different problems with different answers.

Designing rollback triggers beyond error thresholds for AI canary deployments

The instinct when inheriting a canary deployment system is to reach for the same rollback logic you'd use for a microservice: if error rate crosses 1%, roll back; if p99 latency exceeds 500ms, roll back. That pattern works because microservice failures tend to be loud — they surface in HTTP status codes, exception logs, and infrastructure metrics.

AI model degradation is structurally different. A model can produce subtly biased completions, quietly regress on accuracy for a specific user segment, or begin hallucinating in ways that never register as a 5xx error. The rollback trigger layer has to be redesigned from the ground up to account for this.

Why binary thresholds fail for AI model rollbacks

The core problem is that AI models don't fail the way software fails. As models encapsulate learned weights, training data lineage, feature preprocessing pipelines, and implicit behavioral assumptions — failure is distributed across all of these dimensions simultaneously, not concentrated in a single error signal.

A model that's "healthy" by every infrastructure metric can still be silently degrading a niche slice of production traffic. Aggregate error rates look fine; a user segment is being harmed. A hard percentage threshold will never catch this, because the signal you need to act on isn't in the error log at all.

Output quality score triggers and guardrail metric design

The practical alternative is to define guardrail metrics that capture output behavior directly — task completion rates, engagement depth, output quality scores — and wire rollback logic to statistical degradation in those metrics rather than to threshold breaches.

The key design insight from GrowthBook's Safe Rollouts with guardrail metrics is that the trigger threshold should be zero: as soon as there's statistical certainty that a metric is being harmed at all, even by very small amounts, the rollout is marked as failing. This is a fundamentally different posture than "roll back if error rate exceeds 5%." It treats any statistically confirmed harm as sufficient cause to stop, regardless of magnitude.

The mechanism that makes this workable without constant false positives is one-sided sequential testing, which allows rollback decisions to be made as soon as statistical significance is reached without the false positive accumulation that comes from repeatedly peeking at results.

This matters because AI output metrics are inherently noisy — response quality scores fluctuate, engagement signals vary by time of day, and a naive threshold approach will generate rollback events that don't reflect real model regression. One practical constraint worth taking seriously: GrowthBook's documentation explicitly warns that choosing too many guardrail metrics increases false positive risk.

A focused set of critical metrics — the ones that most directly reflect model output quality for your specific use case — will produce more reliable rollback signals than an exhaustive monitoring dashboard.

Anomaly detection and model drift as rollback signals

Beyond quality score triggers, anomaly detection deserves treatment as a first-class rollback signal, not a secondary alert. item.com's canary deployment framework explicitly includes anomaly detection signals alongside latency and accuracy in its rollback trigger specification — the same tier, not a supplementary layer.

The reason is that model drift can originate from data distribution shifts rather than model weight changes, meaning a model that performed well at deployment can degrade as the distribution of incoming requests shifts away from the training distribution. This kind of drift doesn't produce errors; it produces outputs that are subtly off-baseline in ways that only anomaly detection will surface.

Automated vs. human-in-the-loop rollback decisions

Not every rollback signal warrants the same response. When guardrail metrics cross statistical significance thresholds — clear, statistically certain regression — automated rollback is appropriate and fast. GrowthBook's Safe Rollouts support an Auto Rollback toggle for exactly this case, and the system exposes webhook events (feature.saferollout.rollback, experiment.decision.rollback) that can wire rollback decisions directly into CI/CD pipelines.

But ambiguous degradation signals — cases where results are inconclusive or where traffic distribution looks unhealthy rather than regressed — warrant human review rather than automated action.

GrowthBook distinguishes between a "Guardrails Failing" state (regression detected, consider reverting) and an "Unhealthy" state (traffic imbalanced, check implementation), treating these as distinct failure modes that require different responses. The default for genuinely inconclusive results is to ship, not block — a deliberate bias toward action over paralysis that teams should adopt as an explicit policy rather than leaving undefined.

The engineering decision of where to draw the automated/manual boundary is not a tooling question — it's an organizational one. Teams that haven't defined it before the canary starts will make it under pressure, which is the worst time to reason clearly about statistical ambiguity.

Treating the AI canary release as a controlled experiment, not just a safe deployment

There's a meaningful difference between asking "did anything break?" and asking "did this version produce better outcomes than the one before it?" Most canary deployments are designed to answer the first question. The most mature AI teams have learned to answer both simultaneously — and that shift in framing has concrete consequences for how the rollout is designed before a single request is routed.

The defensive posture — monitor for errors, watch latency, roll back if something explodes — is necessary but not sufficient for AI. A model can pass every failure threshold while quietly producing responses that are less helpful, less accurate, or less aligned with what users actually need.

If you're not measuring that, you're not learning anything from the canary beyond "it didn't crash." That's a low bar for a production model release.

Defining 'better' before traffic routes, not after results arrive

The practical implication of treating a canary as an experiment is that you must define what "better" means before you route traffic, not after you've seen the results. This sounds obvious, but it's routinely violated in practice.

Teams deploy a new model version, watch the dashboards for a few days, and then decide whether the numbers look good enough to promote. That's post-hoc rationalization, not evidence.

The experimental framing imposes a pre-condition: write the hypothesis first. What specifically do you expect this model version to improve, for which users, and by how much? GrowthBook's experiment data model enforces this structurally — the schema requires a hypothesis field, making the discipline of defining expected outcomes a required step in the deployment workflow rather than an optional one.

That's not just a UX nicety; it's the platform encoding experimental discipline into the deployment workflow.

Character.AI is a concrete example of this approach in production. Landon Smith, Head of Post-Training at Character.AI, describes it directly: "GrowthBook has been an invaluable tool for Character.AI, helping us develop our models into a great consumer experience. 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 — comparing modeling techniques from the perspective of user outcomes — is precisely the shift this section is arguing for.

Selecting primary vs. guardrail metrics before rollout

Pre-registering metrics before a canary begins is what separates a controlled experiment from a monitoring exercise. GrowthBook's experiment schema separates three distinct tiers: primary goals, secondary metrics, and guardrails — each with configurable risk thresholds.

Primary metrics answer the hypothesis directly: did task completion improve, did engagement depth increase, did user satisfaction scores move? Guardrail metrics define the floor — the conditions under which the experiment stops regardless of how the primary metrics look. Latency SLA violations, safety guardrail breaches, and cost-per-query spikes are all candidates for guardrail status.

This separation matters because it prevents the peeking problem: if you're free to adjust which metric you're optimizing for after seeing early results, your statistical conclusions are invalid. The same discipline that rigorous A/B testing programs and clinical trials require now applies to model deployment.

Feature flags as the traffic control layer

Feature flags are the mechanical layer that makes canary-as-experiment operationally feasible. They control which users see which model version, generate the assignment data needed for statistical analysis, and enable instant rollback without redeployment.

The key distinction from a pure canary deployment is that the flag assignment is logged and tied to outcome metrics — creating the experimental record that drives the promotion decision.

GrowthBook's phases structure allows configurable coverage percentages and traffic split weights per variation, mapping directly to canary-style incremental exposure within an experiment framework. Reducing the friction of instrumenting every model release as a proper experiment — rather than an ad-hoc rollout — is what makes this discipline sustainable at shipping cadence.

Statistical rigor for promotion decisions

Promoting a model to 100% traffic because "it looked fine" is not the same as having statistical evidence it outperforms the previous version. The promotion decision should be driven by primary metric improvement above a pre-defined threshold, guardrail metrics within bounds, and statistical significance — not just the absence of incidents.

GrowthBook supports both Bayesian and frequentist statistical approaches, so teams can use whichever methodology their data science function prefers. Sequential testing allows the canary to be monitored continuously — checking results as data accumulates — without increasing the risk of false alarms that comes from repeatedly peeking at incomplete data.

Regression adjustment reduces the noise in your metrics, which means you need less data to reach a confident conclusion. Sequential testing is particularly relevant for AI canaries: it allows teams to make promotion decisions as soon as statistical confidence is reached rather than waiting for a fixed sample size, which matters when new model versions are shipping frequently.

Each canary that runs this way becomes a data point in an institutional learning library — evidence about what actually moves user outcomes, not just a deployment event that happened without incident.

Closing the gap between canary safety theater and actual AI deployment confidence

As established at the outset: the through-line of this article is a single category error — treating AI model degradation as if it were software failure. Software fails loudly, models fail quietly. A hallucinating model returns HTTP 200. A drifting model serves every request — just worse.

If your canary process is instrumented only to catch the loud failures, you have a false sense of safety, not actual safety.

The single question that reveals whether your canary can detect AI failure

The most useful thing you can do right now is look at your existing canary setup and ask one question: if your new model version started producing subtly worse outputs tomorrow, would any of your current monitors catch it?

If the honest answer is "probably not," that's your gap. You don't need to rebuild everything at once — but you do need to know where the blind spots are before the next release goes out.

One output quality signal beats zero infrastructure metrics for AI rollback

You don't need a complete observability platform before your next canary. You need one output quality signal that's closer to user outcomes than error rates — task completion, engagement depth, cost per successful interaction — and a rollback trigger wired to statistical degradation in that signal rather than a hard threshold.

That's a meaningful improvement over where most teams start, and it's achievable before your next model release.

Promotion decisions that require evidence, not just the absence of incidents

The shift from "did anything break?" to "did this version produce better outcomes?" is not a tooling upgrade — it's a discipline change. It means writing the hypothesis before routing traffic, pre-registering your primary and guardrail metrics, and treating the promotion decision as something that requires evidence, not just the absence of incidents.

Warehouse-native experiment analysis and statistical guardrails are built specifically to make this workflow operationally feasible — connecting model assignment data to user outcome metrics without a separate event pipeline, and surfacing rollback signals the moment statistical certainty is reached.

This article is meant to be genuinely useful to teams who are shipping AI models and want their deployment process to match the actual risk profile of what they're releasing — not a theoretical framework, but a practical reorientation.

What to do next: Pull up your current canary monitoring setup and identify whether you have any metric that reflects what your model is producing — not just whether it's running. If you have zero output quality signals today, start there: instrument one user-outcome metric (task completion rate is a good default) and define what statistical degradation in that metric would cause you to pause a rollout. That single addition closes the most dangerous gap this article describes, and it's the right first step regardless of what tooling you're using.

Sign up for free

Take Growthbook for a spin, no credit card required.

Create my account

Example H2

See All Articles

Experiments

Feature Flags

CLI vs MCP: TLDR, it doesn’t matter as much as you think

Oct 21, 2026

min read

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:

npm install -g growthbook
growthbook configure
growthbook whoami
growthbook features create --help
growthbook features get --help

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:

{
  "mcpServers": {
    "growthbook": {
      "url": "https://mcp.growthbook.io/mcp"
    }
  }
}

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:

{"name":"growthbook_list_skills","arguments":{}}

{"name":"growthbook_read_skill","arguments":{"name":"feature-flags"}}

{"name":"growthbook_read_skill","arguments":{"name":"feature-flags/references/flag-create"}}

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:

growthbook features list-keys --output-format json
growthbook projects list --output-format json
growthbook environments list --output-format json

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:

{"path":"/api/v2/feature-keys"}

{"path":"/api/v1/projects"}

{"path":"/api/v1/environments"}

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:

{
  "id": "checkout-redesign",
  "valueType": "boolean",
  "defaultValue": "false",
  "description": "Control exposure to the redesigned checkout",
  "project": "<PROJECT_ID>",
  "environments": {
    "staging": { "enabled": false },
    "production": { "enabled": false }
  }
}

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:

growthbook features create --body "$(cat flag.json)" --dry-run
growthbook features create --body "$(cat flag.json)" --output-format json

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:

$flagPayload = Get-Content -LiteralPath .\flag.json -Raw
growthbook features create --body $flagPayload --dry-run
growthbook features create --body $flagPayload --output-format json

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:

{
  "method": "POST",
  "path": "/api/v2/features",
  "body": "{\"id\":\"checkout-redesign\",\"valueType\":\"boolean\",\"defaultValue\":\"false\",\"description\":\"Control exposure to the redesigned checkout\",\"project\":\"<PROJECT_ID>\",\"environments\":{\"staging\":{\"enabled\":false},\"production\":{\"enabled\":false}}}"
}

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:

growthbook features get --id checkout-redesign --output-format json

Or invoke growthbook_api_read with:

{"path":"/api/v2/features/checkout-redesign"}

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

Experiments

Analytics

Standard deviation vs variance: Key differences for A/B testers

Oct 21, 2026

min read

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:

from statistics import variance, stdev, pvariance, pstdev

times = [2, 4, 4, 6]
print(round(variance(times), 3), round(stdev(times), 3))
print(round(pvariance(times), 3), round(pstdev(times), 3))

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

Experiments

Analytics

How to calculate true positive rate (TPR)

Oct 21, 2026

min read

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 Guide

Compare 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 GrowthBook

Ready to ship faster?

No credit card required. Start with feature flags, experimentation, and product analytics—free.

Get Started

Book a Demo

Simplified white illustration of a right angle ruler or carpenter's square tool.

White checkmark symbol with a scattered pixelated effect around its edges on a transparent background.

A hallucinating language model returns HTTP 200.

Why standard canary release logic breaks down for AI models

How traditional canary releases work — and what they're designed to catch

The signal layer traditional canaries rely on

Where those signals fail for AI models — concrete examples

The conceptual shift required — from infrastructure health to output quality

The metrics that actually matter: monitoring AI output quality during a canary release

Inference latency and throughput — the baseline you still need

Accuracy and drift signals — detecting what error rates miss

User-outcome metrics — where model benchmarks and business reality diverge

Token efficiency and cost signals

Why rollback decisions require reproducible analysis, not black-box dashboards

Incremental traffic routing: structuring the canary exposure window for AI models

Initial traffic split mechanics

Scaling increments and hold periods

Compute isolation requirements

Sample size considerations for generative AI outputs

Designing rollback triggers beyond error thresholds for AI canary deployments

Why binary thresholds fail for AI model rollbacks

Output quality score triggers and guardrail metric design

Anomaly detection and model drift as rollback signals

Automated vs. human-in-the-loop rollback decisions

Treating the AI canary release as a controlled experiment, not just a safe deployment

Defining 'better' before traffic routes, not after results arrive

Selecting primary vs. guardrail metrics before rollout

Feature flags as the traffic control layer

Statistical rigor for promotion decisions

Closing the gap between canary safety theater and actual AI deployment confidence

The single question that reveals whether your canary can detect AI failure

One output quality signal beats zero infrastructure metrics for AI rollback

Promotion decisions that require evidence, not just the absence of incidents

Sign up for free

Table of Contents

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.

The debate combines three different questions

How can the agent reach the product?

What procedure should the agent follow?

What context makes the procedure relevant?

GrowthBook puts the durable parts below the interface

The API provides the operations

Skills provide the workflow guidance

MCP provides a thin bridge

Explore the shared flag workflows

The same disabled flag can use either route

Connect and verify before preparing the write

Resolve real IDs and construct one payload

Submit once, then read the object back

Choose the route that fits the runtime

Shell access favors terminal composition

Remote tool access favors MCP clients

Use a small decision table

Compare actual outcomes, not tool counts

Keep the evaluation task constant

Test failure behavior deliberately

Record the versions behind the result

Build workflows that survive the next client

Start with a common foundation

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.

Understand the difference in units

Variance uses squared distances

Standard deviation returns to the original scale

Calculate both from the same observations

Work through a small dataset

Choose sample or population formulas deliberately

Verify the calculation with explicit function names

Find the noise you can reduce

Use standard errors for experiment uncertainty

Distinguish users from estimated means

Combine arm variances before taking the square root

Avoid interpreting intervals as outcome ranges

Check what your metric actually measures

Match the variance to the randomization unit

Inspect tails instead of deleting them casually

Keep units consistent across pipelines

Turn spread into a better testing plan

Define the outcome before planning the sample

Reduce estimation noise without hiding behavior