Every documented change, by year and area.
Each mark is one entry in the record. Select a mark for its source and comparison. Several entries can describe the same system.
Dates keep the precision in the research. On a small screen, scroll inside each picture or table. All entries are also in the table at the end.
How big were the measured effects?
These are different metrics on different populations, and they were published selectively. A lower cost is a fall, even when that is useful.
Each dot is an entry. All groups share one scale, compressed so that 1% and 1,000% both fit: each labelled step is ten times the last. The dark tick is the median of the dots in that group that are not marked unverified, not a typical feature lift. Filters change the dots and medians. Repeated reports are kept.
The plot keeps only one clear percentage change per entry. It leaves out ranges, bounds, shares, multiple effects and basis points.
How it landed
What the launch catalogue retains from press and users, alongside the measured record. Reception is qualitative evidence, not a measured feature effect or a representative survey.
Loading reception from the launch catalogue.
Reception counts as a table
Each feature counts once, in its primary area. The ink bars share one count scale across all areas. Sentiment is the catalogue's research classification. “Little” means little recorded reaction; “unknown” means sentiment unknown, even if a reception item was retained. “Most discussed” below means the most retained press/user items, including summaries and snippets, with ties ordered by name. It does not measure audience size, popularity or independent voices. These summaries show the whole catalogue regardless of the record filters. Read the full launch catalogue →
What did not work.
Failures and quiet removals from the research, plus removals Pinterest said helped. A missing usage figure means it was not found, not that usage was zero.
Lessons are the research's reasoning unless marked as a source's statement. Long text is cut short here: open an entry for all of it, with its caveat. This list shows the full record regardless of filters.
Patterns.
The research draws these from selected examples. They are interpretations, not rules that predict the next launch. Open each supporting entry to check its outcome and evidence.
Claims about the reason for The Yes reversal are inference.
Precedents for three open bets
Compare the mechanism, metric and study design before borrowing a conclusion. Close, partial and loose describe comparability, not evidence strength. Even a close match may cover only one step of a proposed feature.
Loading public-source precedents.
What this means for the simulator.
These cards have a plain precedent in this record. A precedent supports a mechanism or a question. It does not validate the simulator's effect size.
Simulator outputs are ranges under stated assumptions, not forecasts or probabilities of real outcomes. Country and state figures are allocations.
The full record.
Sort by a column heading. The filters above apply to this table and to both pictures.
Long cells are cut short: open an entry to read all of it, with its source links. READ means the note's author fetched the source and saw the text. SNIPPET means only a search snippet was seen. MIXED combines the two. SECONDARY means the original was not read. GIVEN was supplied in the brief. UNSTATED means the record has no reading label.