The diagram below details the full process from a one-sentence topic to a published report page. Its design draws on best practices from comparative journalism, statistics, data visualization, and AI-agent orchestration.
One-sentence task:
topic + A/B grouping + time window
1. scope
flesh out the detailed collection plan
Confirm the scope + approve reference questions
(touchpoint one)
⟲ revise → 1. scope2. collect
independent per-side collection + balance checks
sentence-segmented corpus +
independent reporting clusters · data store
3. annotate
build the question set (templates/reference/custom)
annotate: answered/answers/sources
question × answer annotation set
4. analyze
stats, answer-rate comparison,
answer comparison, confidence intervals, ranking
statistically tested findings by angle:
consensus / divergence / silence
⟲ new questions surface → 3. annotate⟲ a side's sample too thin → 2. collect5. write
data-journalism writing:
summary → angle selection →
explained conclusions + sentence-level tracing
page copy
(English master)
6. render + localize
automatic page build/checks
+ independent judge reviews
+ localized into the reviewer's language
finished preview page
(reviewer's language)
7. review
review + approve
(touchpoint two)
⟲ sent back: review comments pinpoint the exact step → 5. writeapproved
8. publish
freeze the version, verify, go live
(plus localized into each reading language)