Colophon

The one rule

Nothing renders as measured unless it is measured. Where a number could not be reconstructed, the site shows nothing, or shows the number explicitly marked as unreconstructed. Empty fields are not an unfinished state to be apologised for — they are the honest output of a reconstruction with real gaps.

This was not the original design — earlier drafts rendered placeholder values as though they were findings. Material changes to how the site reports its data are listed under Revisions.

What the site is built from

City guides

15 machine-generated 10-category guides, one per city, built in 2026 from Atlas Obscura, Eater, Time Out, TasteAtlas and web search. 787 recommendations across 15 of 18 cities.

Location history

A Google Takeout export covering 242 days — Maps activity plus 73 saved lists totalling 2,338 places. This is the only independent record of where the trips actually went.

Everything else

Memory. Treated as such throughout.

The comparison, and how it changed

The site was built around a question: does curation change where you end up? The premise was that the first trip was unguided instinct and the second was guide-led. That premise turned out to be false. The saved-list export shows every first-trip city already had a 32–42 place list, self-made, before departure.

So the comparison is not instinct versus curation. It is self-curation — own judgment, no external sources — against sourced curation, from published guides. Under that framing the finding sharpens rather than dissolves: the degree of list-following barely moved — 47.4% of places navigated to on the first trip were already on the list, against 52.5% on the second — but the content inverted: books ×6.9, museums and history ×2.9, record stores from zero to 24. Outdoors was flat.

The source of curation changed what got curated, not how faithfully it was followed.

Error bars

Visit counts are floors, not rates

The public export truncates each city to its top ~8–14 places: 151 entries against 1,102 distinct destinations navigated to. A match proves a visit; a non-match proves nothing. 50 of 787 recommendations are confirmed visited, and 737 have no record either way. Hence “≥” on every city page, and a distinct “visit unknown” state that is deliberately not the same as “skipped”.

Adherence is directions-only

Map-view events were excluded as circular — browsing a saved list generates a view for every item on it. Only navigation counts.

Saves cannot be dated

The export carries no per-item save timestamps, and the lists were edited during the trips. A place saved before leaving cannot be separated from one saved while standing in it. This is why the overlap between the guides and the self-made lists is computed but not published — it would read as a stronger claim than the data supports.

Interest categories are keyword-tagged

A rough instrument, applied to destination names. Places with uninformative names tag as nothing.

The retro guides are an anachronism

First-trip cities had no guides at the time. Theirs were generated in 2026, from 2026 sources, and scored against a trip taken in 2025 — a control group built after the fact. Their recommendations are provenance markers, not outcome claims: “the guide picked this” never means “this was skipped”. Portland is excluded from the comparison entirely; scoring a guide against your own hometown is a category error.

Not yet reconstructed

Known gaps, stated rather than filled:

• Miles driven, and every per-leg distance. Location-history tracking was off during the trips, so there are no GPS traces. Distances need retroactive routing; drive times are bounded only by search timestamps.
• Spend. Card statements have not been exported. No cost figure on this site is reconstructed spend.
• Ratings. No recommendation carries a would-return rating. The only ratings that ever existed here were invented for layout and have been removed. Superlatives cannot be derived until real ones are recorded.
• City character. Vibe, verdict, favourites, field notes and the six-axis fingerprints are unscored. The fingerprints are deliberately deferred to a single sitting — scoring them piecemeal over weeks would let the scale drift and quietly corrupt every comparison between cities.
• Four cities. Portland, Austin, New Orleans and Philadelphia have no guide data. The first three are origin, anchor and interlude; Philadelphia is home.

Revisions

Material changes to how this site reports its data.

1. The hit-rate metric was reframed as a floor. It had counted visits from status values no data had yet populated, so most city pages reported a rate of zero. It now reads as a minimum, and the curation comparison it had been conflated with moved to its own panel with its own source.
2. Washington DC was rebuilt from its guide. Its page had been hand-built early as a schema test, with statuses and ratings assigned for layout rather than from any record. It is now generated like every other city. Fourteen venues that only the original survey covered — the Library of Congress, NMAAHC, Air and Space, the Folger, the Phillips Collection among them — were restored with their citations intact.
3. DC's elevation is blank by override. A city spanning roughly 1–410 ft is misrepresented by any single figure, so the number was removed. Reviewed decisions like this one are now recorded in a file the rebuild reads, so regenerating a page cannot quietly undo them.
4. The matcher joining guides to location history was rewritten. An early version discarded words like “coffee” and “books” as noise, which reduced Coffee Garden to garden and matched it to an unrelated sculpture park. It now requires one name to be a prefix of the other, weighted so that generic words — and each city's own name — cannot carry a match alone.

Built with

SvelteKit, prerendered to static files. No database, no analytics, no third-party embeds, nothing that phones home. Every recommendation carries its source; where a citation could not be recovered it says so rather than inventing one. The data is JSON in a public Git repository, and the transforms that produce it are committed alongside — so every number here can be traced to the file and the line of code that produced it.