/segmently-unit-economics

UC-12 — Growth cycle: from the economics to a measured change, and back

Route: "growth cycle", "where can we grow", "what next after this test", "is this test real?" → step 8 (SKILL.md § Router) · Eval: P12 · MCP: uc-12, cases-uc-12 Contents: Turns · Calls · Answer skeleton · Boundary · Stops · Worked example

Turns

On .ue/<slug>.json (none → cases/uc-00-launch-card.md's interview and init first). One question per step, only what the message does not already carry:

• Step 1: "Where do your numbers come from — analytics, segmently.ai, RevenueCat, the Facebook Ads MCP, or by hand?"
• Step 2: "In how many days do you want to read the result — 7, 14 or 30?" — none named → 30, labelled "default horizon", every later run at it with --horizon-default.
• Step 4: "What would you change (X), and by how much do you expect it to move <lever> (Y %)?" — never proposing either; no change in mind → the same question for recommendation.runnerUp next turn; still none → stop, naming both levers.
• Step 6: "What would it cost to build (dollars), and what are the odds it holds (0 to 1)?" — "I don't know" leaves that flag out, never a default.
• Before step 8: "What are the counts once each arm has reached <nPerArm> — per arm, how many saw it and how many converted?" Steps 3, 7 and 9 ask nothing: each runs in the answer of the step before it.

Calls

• Step 1 by source: RevenueCat → cases/uc-04-revenuecat.md; an ad account → cases/uc-09-ad-account.md; segmently.ai reports or a Facebook Ads MCP this host has → read them, write them into the base as measured (re-basing: the old base into variants[] first), measuredOn naming the source and window; by hand → cases/uc-03-your-numbers.md (a fee the reader states goes in with no preset).
• Step 2: ue rank .ue/<slug>.json --horizon 7,14,30 (realistic judged at 7 days); step 3: ue rank .ue/<slug>.json --horizon <N>.
• Step 6, each of --cost and --odds only with the reader's number: ue rank .ue/<slug>.json --horizon <N> --override <lever>=<Y> --card .ue/<slug>/card-<lever>.json --cost <lever>=<usd> --odds <lever>=<p>
• Step 8, field 10's read (any lever the card tested but renewals); step 9, the same read, every flag kept (--belief, --horizon-default), plus --rebase (the CLI writes the re-base itself): ue stat read --a <n>/<x> --b <n>/<x> --project-file .ue/<slug>.json --lever <lever> --horizon <N> --belief <the card's typed lift>

Answer skeleton

Each step's part of an answer opens with that step's bold name exactly as written below ("Step N — <name>.", its parenthesis included) — never a shortened name — and prints its row whole. A step's question closes the part before it — the step before's row, or for step 1 the onboarding turn — and the reader's reply opens a turn that starts with that step's heading, even when the question already named the step; the exception is step 2's horizon question, which closes step 2's own row: the reply opens with step 3. A reader who names a step directly — "is this test real?" — starts at that step; the steps before it are not re-run. Owner's numbering: there is no step 5.

A default run at steps 2 and 3 carries the structural rows too; a structural pick (a change to the offer) runs cases/uc-13-test-an-offer-change.md's Turns 3–7 in place of step 6, then returns to step 8 with its read. Step 6's card is cases/uc-11-hypothesis-card.md's for a metric lever, cases/uc-13-test-an-offer-change.md's own card (--card) for a change to the offer. Step 8's read of a structural card names its lever --lever <spec>, the structural spec, in place of <lever>.

Boundary

Every step that prints a figure closes with its case's rendered boundary — steps 2 and 3 their rank run's boundary, step 6 the card run's (after step 7's part; the cost and the odds are its lines, right after the lever's belief), step 8 the read's context.boundary, which carries no cost or odds: step 6's estimate line goes in right after its belief line. After step 9's re-base the walk is the re-based base's.

Stops

• No learnable row at step 3 → "bigger change or more traffic", and the cycle stops here.
• Fewer counts than the planned sample (the card row's nPerArm per arm) → "keep the test running to <nPerArm> per arm", no read — its daily check is ue stat watch; underpowered → the counts question at read.needPerArm, no read of a smaller sample; not yet → read.lines (its context.more.sentence last) is the whole answer — never a planned sample already reached, or one that is n/a.
• Step 9 only on significant; a mix read or a drop has no realized run (its context.rebase.lines are the effect).

Worked example

The example follows one metric lever through the cycle, so its rank runs carry --metrics-only; a default run adds the structural rows.

The story of cases/uc-11-hypothesis-card.md#Worked example, every input invented for the dry run — copy the shape, never a sentence or a figure of it. A walk-through, not an answer: each step's answer still closes with its own case's boundary, and every link it hands over keeps the privacy sentence above its call. A step's answer prints its case's whole output; this walk-through abbreviates only where it points at that worked card. Each heading below is the owner's step name, whole, as each step's part of an answer opens.

Step 1 — check the economics (analytics, segmently.ai, RevenueCat, the Facebook Ads MCP, or by hand). Printed whole: the answer of cases/uc-03-your-numbers.md — its plausibility line (each of the reader's numbers placed against the book's P10–P90), the KPIs, walk.lines word for word, breakEvens.chain, the privacy sentence, the ue link call and its link, then its boundary — here on the story's base, whose KPIs, walk.lines and link are field 2 of cases/uc-11-hypothesis-card.md#Worked example. Step 2 follows in the same answer.

Step 2 — growth points reachable in 7, 14 and 30 days. Printed whole, as the worked card's Turn-1 choice run: sentences.traffic, every sentences.table line, a blank line, every sentences.belowTable line as a - bullet — no sentences.pick: a choice run only helps choose — then its boundary (its horizon line names the horizons offered), then only the horizon question. The story's horizons are 14, 21 and 30 days, where step 2 of the growth cycle runs 7, 14 and 30.

Step 3 — the metric with the most profit that can be verified in time. A fresh 21-day run, printed as the worked card's table. Its sentences.pick — printed whole:

p1 is the pick (recommendation.first), p2 the runner-up (recommendation.runnerUp).

Then sentences.mostProfit — printed whole (recommendation.mostProfit p1):

p1 gains the most of the levers learnable inside 21 days: $2,806.36 a month (gainPerMonth, at its landed +10.53 %).

Then its boundary, printed whole, and only step 4's question: "What would you change (X), and by how much do you expect it to move p1 (Y %)?"

Step 4 — the hypothesis "if I change X, the chosen metric moves by Y %". The reader's change and belief, restated with their X and Y and the lever's stage label (field 3 of cases/uc-11-hypothesis-card.md):

If the landing's first screen asks one goal question and routes to the matching step 1, then p1 (acquisition) moves by +12 % — your belief.

Then only step 6's question: "What would it cost to build (dollars), and what are the odds it holds (0 to 1)?"

Step 6 — the hypothesis in numbers (profit if it holds, time and traffic, cost to build, odds it holds). The reader's cost and odds, invented for the dry run: $2,400.00 and 0.5. ONE run:

npx -y @segmently/cli ue rank base.json --metrics-only --horizon 21 --override p1=0.12 --card .ue/synthetic-meditation/card-p1.json --cost p1=2400 --odds p1=0.5

The ten-field card of cases/uc-11-hypothesis-card.md#Worked example, fields (1) to (10) in order — field 5 its card block (card.lines when the card comes inline), field 7 its plan, every sentences.warnings line and its while-it-runs line (UC-11 field 7) — is printed whole; this run's card, table and warnings are that worked card's override run's — only field 9 differs.

Field 9 — the p1 row's bet.sentence, then "Beside it: $<spendRouted> routed over <days> days." — printed whole:

9. Cost — If the +11.84 % on p1 holds, it gains $3,157.16 a month (gainPerMonth). At your odds of 50 % that it holds, the expected gain is $1,578.58 a month (50 % × $3,157.16). Your build cost of $2,400.00 is recovered in 0.76 months of that gain ($2,400.00 ÷ $3,157.16). Beside it: $4,100.80 routed over 4.56 days.

Step 7 — the metrics to watch (the steps that must not drop; for the nearest step the drop that cancels the whole gain and whether it is visible within the test). guardrail.watch and guardrail.sentence — printed whole, right after the card's field 10, in the same answer (fields 4 and 6 keep them too), then the handoff on its own line:

Watch the steps after p1 — p2, p3, bought, close, trialConv and renewals must not drop while it is tested.

p2 must not fall: −10.59 % there cancels the +11.84 % on p1 ($3,157.16 a month, gainPerMonth); guardrail «p2» is not readable within the test's 4.56 days (readable from day 7.43); extend the test to 7.43 days or keep it as a kill switch only.

While it runs: ue exp start the day it goes live, then ue stat watch once a day with the day's counters; the one read, at 1864 per arm, is step 8.

The base's walk.lines are the first 16 lines of the boundary below. The card run's boundary closes steps 6 and 7 — printed whole (the estimate line right after the belief), on the project file:

These figures are this scenario's arithmetic on the inputs above.

• cost per start $1.10, p1 38 %, p2 55 % and p3 32 % — measured: your project file — SYNTHETIC: invented for a process dry run, 2026-09-23
• Monthly with trial: $19.99, monthly, a 7-day free trial — measured: your project file (price, cadence, trial)
• Monthly with trial's take 24 % — measured: your project file (the CLI's init split for these products is 30.1 %, not this)
• Monthly with trial's charges 2.9 — measured: your project file (the book's monthly count is 2.86)
• Monthly with trial's completion 95 % — measured: your project file (the book's is 100 %)
• Monthly with trial's trial conversion 45 % — measured: your project file (the book's is 50 %)
• Annual: $99.99, yearly, no trial — measured: your project file (price, cadence, trial)
• Annual's take 12 % — measured: your project file (the CLI's init split for these products is 14 %, not this)
• Annual's charges 1 (its cadence cap) — assumed: the book
• Annual's completion 97 % — measured: your project file (the book's is 100 %)
• $900 a day — measured: your stated budget
• the fee 2.9 % + $0.30 on the charged base — assumed: the book's card_processor preset
• refunds 3 % — measured: your project file
• disputes 0.4 % at $15.00 each — measured: your project file
• activation — 88 % open the app, 1 charge from those who do not — measured: your project file
• the book's first renewal behind payback (monthly 60 %, from words.payback) — assumed: the book
• p1 +11.84 % (typed +12.00 %) — assumed: your belief, at its landed value.
• p1: your cost to build $2,400.00 and your odds of 50 % that it holds — assumed: your estimate
• the +10 % on p2, close, p3, bought, mix, trialConv and renewals — assumed: ue rank's default ("assumed lift"), at each lever's landed lift in the table.
• the 21-day horizon — measured: your answer
• the 50/50 split — assumed: the test design
• p1 is decidable in 4.56 days, p2 in 8.14 days and close in 17.97 days at this budget.
• not testable soon, in this 21-day run: p3 45.44 days, bought 111.66 days, mix 438.74 days and trialConv 319.55 days — past 21 days at this budget, each at its own days; renewals — a calendar read.

Then only the question: "What are the counts once each arm has reached 1864 — per arm, how many saw it and how many converted?"

While the test runs — before step 8. The reader launches the test in their own A/B tool (the owner's list has no step 5). In a session step 6's card run is on the project file and writes the planned experiment e1 (card.experiment.id); the day the test goes live ue exp start sets it running, and each day after, one ue stat watch takes the reader's counters — p1 and p2 so far, and p2 on the last full day, invented for the dry run (--date pins the dry run's days; a session's check is today's):

npx -y @segmently/cli ue exp start .ue/synthetic-meditation.json e1 --date 2026-09-24
npx -y @segmently/cli ue stat watch --project-file .ue/synthetic-meditation.json --lever p1 --a 1227/466 --b 1225/521 --guardrail p2=466/256,521/280 --today p2=156/86,174/95 --date 2026-09-27

watch.lines — printed whole:

Day 3 of 5 — p1 (e1), running since 2026-09-24.

p2: noise — −2.17 % (pValue 0.7072, not below 0.01); a rise of ≥ 16.4% is readable today (day 3); the drop that cancels the gain (−10.59 %) needs 1,157 per arm for its read at the planned end — 466 so far.

Primary: 1,225 of 1,864 per arm — read on 2026-09-29.

Daily looks: 0.05 ÷ 5 = 0.01 per look — Bonferroni over the planned days, a stand-in for a sequential boundary.

A {{…}} line stands for lines already shown above; an answer prints the lines, never the marker (ue check flags one).

watch.boundary — printed whole:

These figures are this scenario's arithmetic on the inputs above.

{{walk: base — the 16 lines above}}

• the variant: e1's card "card p1 +11.84 % (believed)" (.ue/synthetic-meditation/card-p1.json) holds p1 +11.84 % — assumed: the card's beliefs, as the card carries them
• the 21-day horizon — measured: your answer
• the 50/50 split — assumed: the test design
• the counters of day 3 — measured: your counts

Day 3 of 5, status continue: the test keeps running, one watch a day; it never prints p1's rates, and at 1864 per arm it prints the one read to run — step 8.

Step 8 — judge only on the planned sample (real / not / not enough people and how many more). Ending B of cases/uc-11-hypothesis-card.md#Worked example, at the planned 1864 per arm — field 10's read with the reader's counts:

npx -y @segmently/cli ue stat read --a 1864/708 --b 1864/736 --project-file .ue/synthetic-meditation.json --lever p1 --horizon 21 --belief 0.12

read.lines — printed whole (its drift line, then the verdict's sentence):

pA 37.98 %, pB 39.48 %, lift 0.0395, diff 0.015, ci95 -0.0163 … 0.0463, z 0.94, pValue 0.3465, needPerArm 16509, verdict underpowered

p1 scenario 38 % → observed 39.48 % (1.48 points higher)

Not enough people yet: the observed difference needs 16,509 per arm (read.needPerArm) and the smaller arm has 1,864 — 14,645 more per arm, 35.8 more days at 818.18 funnel starts a day, split 50/50.

Not enough people — underpowered: the test keeps running to the new planned sample. The read's context.boundary — printed whole (the read sent with --belief 0.12), step 6's estimate line added right after the belief line — a read has no cost or odds; no realized lift on underpowered:

These figures are this scenario's arithmetic on the inputs above.

{{walk: base — the 16 lines above}}

• p1 observed in the test, read.pB 39.48 % as printed — measured: prepared from your counts
• the counts 708/1864 (control) and 736/1864 (new) — measured: your answer
• p1 +11.84 % (typed +12.00 %) — assumed: your belief, at its landed value.
• p1: your cost to build $2,400.00 and your odds of 50 % that it holds — assumed: your estimate
• the 21-day horizon — measured: your answer
• the 50/50 split — assumed: the test design
• not testable soon, in the card's 21-day table run: p3 45.44 days, bought 111.66 days, mix 438.74 days and trialConv 319.55 days — past 21 days at this budget, each at its own days; renewals — a calendar read.

Then only the question, at read.needPerArm: "What are the counts once each arm has reached 16509 — per arm, how many saw it and how many converted?"

Step 9 — update the economics with what was measured (new base, the old one beside it, KPIs recalculated). Ending A of cases/uc-11-hypothesis-card.md#Worked example, significant → real — step 8's word for the read, said first. Printed whole as that worked card prints its read: read.lines (the realized gain and its warnings among them), the same read plus --rebase (the CLI writes the re-base), the re-base block with context.rebase.lines and the re-based link — then the read's context.boundary, step 6's estimate line added right after the belief line. Then back to step 2 on the re-based file, in the same answer: a new --horizon 7,14,30 run on it, printed whole as step 2.

A read of a lever off the chain. On a fresh copy of the pre-test project, .ue/synthetic-meditation-copy.json, a close test's counts, invented for the dry run:

npx -y @segmently/cli ue stat read --a 1000/957 --b 1000/975 --project-file .ue/synthetic-meditation-copy.json --lever close --horizon 21 --rebase

Its subject: close (payment completion). Its drift line: close (payment completion) scenario 95.67 % → observed 97.50 % (1.83 points higher). Its re-base block's prepared lines — printed whole:

• the realized lift 0.019164 (read.pB 0.975 ÷ today 0.956667) through ue rank's own close move — prepared from your counts
• Monthly with trial's completionRate 0.95 × 1.019164 = 0.968206 — prepared from your counts
• Annual's completionRate 0.97 × 1.019164 = 0.988589 — prepared from your counts

Every context.rebase.lines entry — printed whole:

• profit per start −$0.03 → −$0.01
• ROAS 0.97 → 0.99 (valueBasis net)
• CAC per payer $75.10 → $73.57
• payback not covered within 12 months → not covered within 12 months
• stage live_no_data → live_measured