What You Need to Know
A service level is a promise about how fast, available or accurate a system is. The envelope is what the real pipeline, on the real corpus, measurably holds. Engineer first, then promise, never the reverse.
Take a car valeting firm whose Claude assistant answers booking and price questions from its price list and bookings. A reply time the owner names in a meeting and hands to engineering is a liability with a date on it. Measure instead: Domain 3's latency decomposition and Domain 4's percentiles, under realistic concurrency, with headroom for load the measurement never saw.
The Anatomy of a Defensible Commitment
A bare figure is a wish. Six elements make it operable:
- The metric: what is measured, and how it is computed.
- The percentile or threshold, because a commitment on the mean is a commitment nobody experiences.
- The measurement point: where the clock starts and stops, which separates a promise about the model from a promise about the user.
- The window: over what period compliance is assessed, so one bad hour cannot breach a year and one good week cannot hide a bad month.
- The population: who and what it covers, defined honestly before the measurement, not after it.
- The breach behaviour: what happens when it is missed.
The valeting firm's clock should start when the customer sends a message. Anthropic's success-criteria guidance applies this rigour to quality: a named held-out set of stated size, a quality floor, a safety rate and a response-time bound.
You Cannot Promise Above Your Dependencies
Before writing an availability or tail-latency clause, check what the platform underneath commits to. Anthropic documents its standard service tier as best-effort availability. A prioritised tier carries an explicit uptime target but needs a committed capacity arrangement, itself no longer generally purchasable. Traffic beyond that commitment falls back to standard. A promise above the envelope underneath it is written against somebody else's best effort. The valeting firm's uptime promise cannot exceed its tier's.
Negotiating the Impossible Ask
When a stakeholder wants a number the pipeline cannot hold, show the decomposition so the figure stops being arbitrary. State the envelope it does hold, then close what gap you honestly can: perceived responsiveness. Anthropic's latency guidance separates baseline latency from time to first token. It recommends streaming because it improves how responsive an application feels. Progressive disclosure, partial results and an asynchronous path with notification apply the same lever. None of them moves total completion time.
Accepting the number is capitulation. Deleting the retrieval or verification steps the answer depended on to hit it is worse: a measured problem becomes an unmeasured one. Dropping the firm's price lookup would speed replies and leave wrong prices unmeasured.
Quality Is a Measured Claim
An accuracy commitment needs the same six elements. "It will be very accurate" fails the same way a bare duration does. Where the shortfall concentrates in an identifiable segment, say so. Commit on the population genuinely measured, and give that segment its own improvement plan, evidence and owner. Fleet-van quotes missing the floor while car quotes meet it would be one such segment.
The Trap That Looks Like Rigour
Narrowing the population until the headline number becomes true leaves the excluded cases in production with no threshold, no review path and nobody accountable. The precise schedule hides the carve-out. Narrow only where the exclusion is visible and nobody is left uncovered by accident.
Feedback Is Instrumentation
A feedback channel must produce evidence and a disposition. It turns a report into a trace, a case identifier or a reproducible input. A valeter's "it gets things wrong sometimes" cannot be actioned; the trace of the wrong quote can. Triage needs a visible outcome, or the channel trains everyone to stop reporting and silence gets mistaken for quality. Confirmed failures route into Domain 4's regression set. Anecdotes routed straight into prompt edits make a system rot. Each edit fixes the reported case and moves one nobody mentioned.
Reports lead with the success criteria agreed at discovery, backed by engineering measurements. Scope change gets the same discipline. Weigh each request, such as adding fleet invoicing, against the agreed criteria and baseline. Put its cost, date and risk to the sponsor before anyone commits. Then reset the plan under their signature.
In the exam
The keyed answer matches a measured envelope, carries all six elements, and streams where a human is waiting. Capitulation and deleting a control are distractors. Hidden population narrowing is the most dangerous option in this objective; visible narrowing can be keyed.
Sourcing Notes
The six-element shape is professional practice, not Anthropic documentation. The tier specifics are a dated snapshot for the quick reference.
Exam Traps
Writing a bare number into a schedule as the service level
A figure with no percentile, measurement point, window, population or breach behaviour cannot be assessed or operated. It reads as a commitment and behaves as an argument waiting to happen.
Meeting a latency target by dropping a retrieval or verification stage
It reads as engineering pragmatism and is evasion. The clock is satisfied while the error rate moves unmeasured, and the promise that was met is no longer the promise that was made.
Offering streaming or partial results where the consumer is a pipeline rather than a person
Those levers move perceived responsiveness, not total completion time. With nobody watching the response arrive, the measured figure is unchanged and the gap is still there.
Answering a disengaged sponsor with more frequent reports and more engineering measurements
Theatre. The sponsor is being reported to against the wrong criteria, and adding volume of the same thing widens the gap it was meant to close.
Practice Scenario
An insurer's service-centre sponsor wants a flat maximum response time written into the schedule for a correspondence-drafting assistant: every draft back inside the same short window, no exceptions. Measurement on the real pipeline holds that window for routine cases but runs well past it on multi-policy customers, a small and permanent share of the queue. Agents wait at the screen while the draft is produced. Any promise must rest on what the pipeline has been shown to do, cover the whole queue and leave draft quality as it is. What should the architect commit to?
Build Exercise
Write the Service-Level Schedule and the Sponsor Page
What you'll learn
- How to derive a commitment from a measured envelope rather than a demonstration
- How to write a latency commitment carrying all six elements
- How to check a promise against the service tier underneath it
- How to close a perceived gap with experience design rather than by deleting a control
- How to lead a sponsor report with the criteria that were agreed at the start
- Measure or estimate the envelope for a Claude system you know, then write one latency commitment with all six elements: metric, percentile, measurement point, window, population, breach behaviour
Why: A commitment derived from an envelope survives the first real week; one derived from a demonstration is a liability with a date on it.
You should see: One sentence a lawyer could put in a schedule and an engineer could alert on.
- Write one quality commitment to the same standard, including who reviews the residual errors, within what time, and at what cost
Why: An accuracy promise fails exactly the way a latency promise fails, and the residual errors are certain rather than hypothetical.
You should see: A threshold on a named population, plus a review path with an owner and a cost.
- Check both commitments against your dependencies — the service tier the system runs on and what it actually promises — and mark anywhere your promise exceeds the envelope beneath it
Why: An availability or tail-latency clause written above the platform underneath it is written against somebody else’s best effort.
You should see: A dependency line per commitment, and at least one honest note where the promise is thinner than it looked.
- Design the gap-closing experience for the ask you could not meet: streaming, progressive disclosure, partial results or an asynchronous path with notification
Why: Perceived responsiveness is a real lever on interactive surfaces and an irrelevance everywhere else, so the choice has to be justified by who is waiting.
You should see: One lever chosen, with the sentence naming whether a human is watching the response arrive.
- Draft the sponsor page led by the success criteria discovery agreed, with the engineering measurements underneath as support, then delete anything that is on the page only because it was easy to measure
Why: A sponsor who agreed to a business outcome and receives infrastructure telemetry has stopped being communicated with, and more of it makes that worse.
You should see: A page whose first section is an agreed criterion with its trend, and a shorter appendix than you started with.
Sources
- Reducing latency — Anthropic
- Service tiers — Anthropic
- Define success criteria and build evaluations — Anthropic
- Claude Certified Architect Professional Exam Guide (v1.0) — Domain 6 — Anthropic