# What changed in review

Fictional GreenLeaf UK, paid-click cohorts September 1–14, 2026; snapshot September 22, 23:59:59 UTC.
This record was prepared with AI-assisted editorial review and independently checked calculations.
It is **not a completed human approval, client review or controlled human-versus-AI experiment**.
The owner’s publication/release decision remains pending. Original model responses are unchanged.

## The decision before the wording

The first draft recommends diagnosis before a creative or budget pivot. That direction is reasonable,
but several numbers and the explanation are wrong. A plausible recommendation does not make its
supporting evidence safe to release.

| Initial draft statement | Checked evidence | Review decision |
|---|---|---|
| Spend was $110,100 in each week | Week 1: $5,462; Week 2: $5,512 | Reject both draft totals and every derived CPA/ROAS based on them |
| New customers were 548 → 482 | Canonical paid-cohort counts: 467 → 357 | Replace; these are attributed first purchases, not incremental subscribers |
| First-purchase revenue was $43,840 → $38,560 | $37,360 → $28,560 | Replace; no causal revenue-loss estimate follows |
| Mobile Week 2 had 2,360 completions | 2,352 / 8,400 = 28% | Replace the count/rate; retain Week 1’s checked 3,780 / 8,400 = 45% |
| Mobile audience/creative shares were stable and composition was essentially absent | Skeptical share of mobile starts: 50% → 25%; exploring: 50% → 75% | Reject the mix explanation |
| Fixed Week 1 mix gives roughly 35.3%–35.5% | Exactly 35⅓% before display rounding | Use 35.33%; state that both mix and within-group changes matter |
| Founder should freeze spend reallocation | The CSV alone does not establish what all commercial/risk decisions should be | Narrow to “do not recommend a budget move from this evidence alone”; retain existing authorized risk processes |
| A diagnostic succeeds if completion returns toward 45% | Returning to a target is not a causal test design | Require verified event definitions, version/cohort evidence and an explicit comparison plan |

All channel, desktop, audience and creative tables in the raw response remain **unapproved**. The
final editorial drafts use the checked metric pack instead. The model invocation was text-only,
without a calculation-execution tool. This is evidence about this run, not every AI analysis product.

## What the focused follow-up did

The second response recognizes the audience-mix shift and gives the correct Week 2 starts per
stratum: skeptical 350, exploring 1,050. It still reports 2,360 total completions, misstates the
Reddit/exploring/value cell as 245 rather than 252, and miscalculates a weighted contribution.
Use the recorded follow-up to inspect iteration; do not silently substitute it for verified math.

## Approved facts for the production draft

| Measure | Week 1 | Week 2 |
|---|---:|---:|
| All-paid media spend | $5,462 | $5,512 |
| All-paid quiz starts | 15,040 | 15,120 |
| All-paid quiz completions | 6,770 | 5,460 |
| Attributed first-purchase customers | 467 | 357 |
| Attributed first-purchase revenue | $37,360 | $28,560 |
| Media CPA | $11.70 | $15.44 |
| Attributed first-order ROAS | 6.84 | 5.18 |
| Mobile quiz starts | 8,400 | 8,400 |
| Mobile quiz completions | 3,780 | 2,352 |
| Mobile completion | 45% | 28% |
| Desktop completion | 45.03% | 46.25% |

Mobile change: −17 percentage points; approximately −37.8% relative. Fixed Week 1 channel ×
audience × creative weights give a Week 2 mobile rate of 35.33%. Under this particular baseline-weight
decomposition, the within-group component is −9.67 points and the remaining composition component
is −7.33 points. Another decomposition order can allocate interaction differently; this is descriptive,
not causal attribution. All 12 comparable mobile strata decline.

## New fictional context and its effect

The context note was prewritten and hashed before the run. It reports a mobile presentation revision
starting September 8 at 09:00 UTC and an unlinked support concern about finding the next-step control.
The release audit, rollout share, ticket counts and links to paid cohorts are missing.

**Verdict: revise the explanation and narrow the next action; confirm the diagnostic-first direction.**
Ask the product owner and media lead to verify the mobile journey and event definitions, request
version-assignment/release evidence, and compare like-for-like groups where evidence permits.
The product owner’s reported availability for a check in the next working day is fictional context,
not evidence that a check happened. A product cause, creative failure or revenue loss remains unknown.

Carry the report of the revision into brief v2 as **reported context pending verification**, never
as a confirmed cause. No automatic pause, spend change or external alert is approved.

## Instructions to the production pass

Preserve the approved table, uncertainty and decision above across three drafts:

- Founder/CEO: trajectory and the diagnostic decision, with at most three actions and role owners.
- Media: period/channel/device detail, the fixed-mix comparison and the investigation work order.
- Content/creative: keep message hypotheses distinct from proof; inspect post-click context before
  proposing a message test. Do not present a causal creative winner or recommend media-budget changes.

For the retrospective alert: trigger on settled paid-mobile completion below the stated 45% target.
With end-of-day click cohorts and a seven-day observation rule, the full Week 2 comparison is first
justified September 21 at 23:59:59 UTC; the supplied snapshot is September 22 at 23:59:59 UTC.
Detect → Analyze → Route → Review is a static demonstration. Product owner/media lead own the
diagnostic; a measurement-competent reviewer with release authority owns client-facing approval.
No human approval or external delivery has occurred. Recheck the rule/context next cycle, and
escalate missing definitions or evidence to the marketing owner rather than issuing a firm claim.

## Provenance

Checked against the supplied reconciled CSV, independently reviewed deterministic calculations and
`data/analysis.json`. The raw initial and follow-up responses remain downloadable evidence. Numeric
claims in final editorial drafts must map to the claim ledger. “Approved” here means approved for
this AI-assisted editorial production pass; it does not stand in for the owner’s human release gate.

## Check of the recorded production pass

The recorded formatting pass preserves the checked period/device counts and financial ratios.
It also introduces “hold the diagnostic decision until role owners complete their checks,” which
could be read as postponing the investigation we intend to start. That wording is not retained in
the separate editorial reports. Its content section groups “targeting/mix changes” under hypotheses:
the mix shift is observed in the synthetic data; what caused it remains a hypothesis.

The supplied founder/media/content files are separately prepared, checked editorial drafts from the
same approved facts. They are not advertised as verbatim copies of the formatting response. The raw
production output remains unchanged so readers can inspect this last review boundary as well.