Growth experiments

Growth-enabled worlds accept growth_experiment and growth_rollout through the normal commitment tool. A player composes an audience, one or more channels, quoted offer, delivery dates, allocation and budget. A hypothesis explains the decision but does not change physics or earn points.

What executes

An experiment registers two to four arms and an explicit baselineArmId. Allocations total 10,000 basis points. Identical A/A policies are legal; a baseline need not be untreated. The addressable market produces ordinary organic signups even without a study; registering an experiment does not create that population. During enrollment, every prospect is assigned before audience filtering. Ineligible, unexposed and nonconverting people stay in the intention-to-treat denominator. Prospect identities depend on company, seed, calendar day, segment and ordinal; changing an episode, experiment or arm label cannot buy another population.

Each channel charges its approved daily budget only when an eligible assigned population is available after setup. Its effective reach probability is effectiveSpend / (effectiveSpend + eligibleAssigned * marginalReachCost), where effective spend includes the published ramp. This finite saturating model makes doubling spend less than double reach. Channels use separate event-keyed draws; one person can see several channels but create only one economic account. Exposure reports identify the first reached channel, not causal attribution credit.

A signup enters the ordinary account kernel. Procurement, trials, usage, support, capacity contention, conversion, payment failures, retries, refunds, cancellation and renewal remain active. Introductory discounts apply to a locked recurring-price basis for a declared number of billing periods. For per-unit contracts, unit prices scale while fixed fees remain fixed; usage overages remain separately chargeable. Annual contracts collect twelve monthly units and carry calendar-year service obligations. Introductory offers therefore affect actual cash, MRR, earned revenue and later renewal decisions; they do not create a separate synthetic reward account.

Rollout requires a registered arm and the experiment’s completed outcome lag. It assigns that arm to at most 90% of future prospects and delivers the registered baseline to the remainder, including baseline advertising costs. Prior accounts and offers remain unchanged. Only one experiment can enroll a shared market at a time, and new enrollment cannot overlap an unsettled rollout. Cancellation ends future enrollment and expenditure while preserving past charges, accounts, invoices, and paid-through service. The ordinary cancellation fee still applies.

Public reports and qualification

snapshot.growth publishes known assignment bounds, daily addressable market size, marginal reach costs, setup and ramp dates, allowed discount bounds, current quoted list-price basis, minimum follow-up guidance, and task requirements. Private segment response probabilities are omitted. Results follow the scenario’s reporting delay and separate experiment and rollout populations. They report assigned, eligible, exposed, signup, mature, paid and retained counts, cash collected, earned core revenue, acquisition cost, allocated variable service cost and contribution.

Contribution means earned core subscription revenue minus campaign/setup/cancellation and allocated variable service costs. Variable operating costs are allocated across actual active/trial users with deterministic integer-cent remainder allocation. Common payroll, overhead and add-on economics are excluded from this cohort figure and remain in ordinary company financial objectives. Annual cash is never substituted for earned contribution.

Task requirements can additionally impose retained-customer counts, rollout sample size, an acquisition spending ceiling or a dated paid-account checkpoint. A task cannot qualify by writing a persuasive hypothesis. Primitive account and billing verifiers reconcile economic receipts; the growth audit checks registration, real channel expenditure, quoted offers, exposure eligibility and a reconstructed daily population census, including missing nonconverters. Private random draws additionally require the existing source-bound command replay. A standalone seal does not reveal the private RNG seed or prove a hidden propensity draw by itself.

The public reference policy uses only published terms and reports, reserves real project costs and maintains the business. Audience-specific arms use the full disclosed prospect census because ineligible assignments remain in their denominators. Service-readiness tasks enroll a monthly cohort through the lowest quoted marginal-reach-cost channel, with a bounded introductory offer; the preparation control receives the same commercial treatment. Other families compare multiple approaches. Rollout uses a disclosed cash-payback estimate, normalizing annual receipts and already possible monthly invoices, and allows at most a month of future delivery with the required follow-up. That estimate never substitutes for the grader’s earned-contribution accounting.

A separate preregistered control changes the relevant audience, budget, offer, timing, preparation or rollout decision. catalog admit retains the existing replay and feasibility evidence. New task metadata remains structurally_validated; it is not a claim of executed admission. Keep failed feasibility and audience-control attempts with their original source and inputs. A later successful path does not erase an earlier counterexample.

Evidence and uncertainty

The numeric model in data/growth-model.json explicitly labels its population, units, nominal parameters and sensitivity bounds as authored. Current task cases embed the nominal values; the bounds describe proposed sensitivity checks and are not an already-executed robustness study.

• Microsoft Research’s Diagnosing sample ratio mismatch in A/B testing informs preserving assignment denominators and distinguishing assignment from exposure.
• Dropbox Engineering’s subscription experimentation write-up informs explicit outcome lag and the distinction between early proxies and realized subscription outcomes.
• DoorDash Engineering’s switchback experimentation discussion informs the shared-resource interference limitation.

The catalog source entries link the public articles and retain their reuse constraints. These references support design requirements, not the fictional companies’ response, saturation, discount or retention coefficients. Articles remain publisher-copyrighted; the repository retains links and paraphrases, not customer data or redistributed article bodies.

Arms share actual service and support capacity. Reported effects describe the tested mixture, not an isolated all-market intervention. There is no natural-language creative-quality judge, ad auction, post-campaign adstock, referral network, automatic statistical-significance claim, or empirical industry calibration. Repeated policies and seeds are not counted as new mechanisms. The catalog composes growth decisions across distinct company structures. Admission and sensitivity evidence must identify the exact cases and authored assumptions they tested.

Human experiment editor

In a running growth task, choose Add work to plan → Test a growth idea. Give the hypothesis, enrollment window and follow-up, then use the approach tabs to choose audiences, channels, daily budgets and offers. The explicit baseline can carry the same policy as another approach for an A/A check. Shares must total 100%; Split equally handles the integer remainder. The spending limit includes setup and the full declared delivery budget.

After the published follow-up has matured, Expand a tested approach shows assigned populations, actual payments and earned contribution. Select the approach, delivery period and treatment share; at least 10% continue receiving the registered baseline. The ordinary plan cancellation control stops future delivery while preserving earlier assignments, expenses and customer promises.

The public reference can use a second allowed study when a mature first study has no paid treatment accounts. Audience tasks then compare channels on the same broad audience; activation tasks can compare introductory offers and billing cadence. An already paying cohort is allowed to earn revenue rather than automatically funding another study because early earned contribution is negative. Enrollment and rollout never overlap. These are disclosed planning heuristics, not a feasibility guarantee.

Paid core infrastructure

Operating companies may offer service_capacity_expansion { blocks } through company.operating.capacityExpansion; growth tasks may override that quote with task.growth.capacityExpansion. Built-in operating companies disclose an authored offer. Legacy/custom companies without either quote cannot purchase capacity. Changing that menu changes company/case identities and requires validation of the new bindings. It expands the core daily work-unit capacity used by actual trial and paid customer service, separately from employee hours and add-on capacity. The order pays its upfront purchase before engineering setup; both complete setup and the quoted lead time must elapse before capacity is usable. Hosting then accrues in exact integer cents over each calendar month, including idle days. It affects cash, spending, operating margin and the published monthly operating cost. Cancelling unfinished work preserves sunk spending and grants no capacity. Completed infrastructure cannot currently be retired or scaled down during an episode.

The quote in data/growth-model.json uses normalized work units: 100 units/day per block, 5,000 cents upfront, 500 cents/month, eight engineering hours, three lead days and at most 100 installed or pending blocks. These are authored contract terms, not cloud-provider prices, measured machine throughput or estimated causal effects. The business snapshot publishes the effective quote, installed units, hosting cost and latest observed demand, including for tasks without growth experiments. Demand follows the ordinary report delay; it is not privileged current state. The existing purchase form uses those public terms, and the reference planner reserves remaining hosting cost separately from immediate setup spending. That reserve uses a conservative 28-day calendar bound and retains completed purchases until the delayed business report includes their hosting cost; publishing that report replaces, rather than duplicates, the reservation. The reference’s initial 50% headroom and subsequent 70% utilization target are disclosed planning heuristics, not calibrated optima.

AWS Well-Architected’s REL07-BP03 Obtain resources upon detection that more resources are needed for a workload supports the requirements to provision against demand, preserve headroom during delayed capacity delivery, and account for resource costs. Its primary documentation is linked in the growth-aws-capacity source record. The daily, manually approved simulator operation is coarser than production minute-scale autoscaling; it does not simulate instance types, regional cloud supply, billing reservations or an autoscaler.

Optional funded-cohort accounting requirement

growth.requirements.fundedOutcome adds a requirement for one declared phase (experiment or rollout), minimumMatureAssigned (at least one person), and signed integer minimumNetContributionCents. An optional nonnegative minimumAcquisitionCostCents requires actual attributed spending. These checks supplement the existing aggregate requirements. Omitting the object preserves the previous contract and behavior.

Funded arms are identified from positive daily budgets in their immutable registered channel terms, regardless of the chosen baseline label. Every assigned person in those arms and that phase counts, including ineligible people, people never exposed, and non-signups. Every assignment must complete the registered follow-up before the pooled result can pass. An organic rollout holdout contributes neither population nor money to the funded pool; a baseline with paid terms still counts. Experiment results cannot substitute for rollout outcomes.

Net contribution and acquisition costs use the existing public arm results reconstructed from the ledger. Acquisition cost includes attributed setup, actual delivery and cancellation costs; a cost floor is not a channel-spend-only floor. Integer cents are pooled exactly, without averaging rounded arm rates. Costs remain charged when delivery is stopped, and no positive result is inferred from selecting a paid channel without sufficient matured assignments and, when required, attributed expenditure. This is observed funded-cohort accounting, not an estimate of advertising causal superiority. Shared-capacity spillovers and mixture effects remain.